WO2019208158A1 - Image identification device, image identification method, and image identification program - Google Patents

Image identification device, image identification method, and image identification program Download PDF

Info

Publication number
WO2019208158A1
WO2019208158A1 PCT/JP2019/015064 JP2019015064W WO2019208158A1 WO 2019208158 A1 WO2019208158 A1 WO 2019208158A1 JP 2019015064 W JP2019015064 W JP 2019015064W WO 2019208158 A1 WO2019208158 A1 WO 2019208158A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
public
target
target image
similarity
Prior art date
Application number
PCT/JP2019/015064
Other languages
French (fr)
Japanese (ja)
Inventor
浩司 藤本
孝尚 宮武
Original Assignee
テンソル・コンサルティング株式会社
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by テンソル・コンサルティング株式会社 filed Critical テンソル・コンサルティング株式会社
Publication of WO2019208158A1 publication Critical patent/WO2019208158A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/53—Querying
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/53—Querying
    • G06F16/532—Query formulation, e.g. graphical querying
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis

Definitions

  • the present invention relates to a technique for detecting image theft.
  • Patent Documents 1 and 2 disclose similar image determination apparatuses that determine the degree of similarity between an original work image and another image.
  • Patent Document 1 an electronic original work image is prepared, and content presumed to include an image that is the same as or similar to the original work image is acquired from the content holding device, and the original work image is included in the content.
  • a mode is described in which a unit histogram group for each unit width is generated by changing the unit width, and the similarity between both images is determined by comparing the generated unit histogram group and the feature amount of the original work image.
  • Patent Document 2 a predetermined digitized original work image and content presumed to include the same or similar image as the original work image are acquired, and the acquired content is compared with the original work image.
  • a target image that is a target is specified, information indicating similarity between the images is generated by comparing feature quantities of the specified target image and the original work image, and information indicating the similarity is obtained from the acquisition destination of the target image.
  • a mode of outputting together with information is described.
  • Patent Document 2 image conversion processing such as enlargement, reduction, rotation, and tone conversion is applied to at least one of the specified target image and original work image to more closely approximate the feature amounts of both images, It also describes that feature values of both images after image conversion processing are compared.
  • Patent Documents 1 and 2 If the methods described in Patent Documents 1 and 2 are used, it is possible to calculate how similar two images are, or to accurately detect that the same object appears in the two images. is there. However, even with these techniques, it is difficult to achieve both accuracy and speed.
  • An object of the present invention is to provide a technique for searching for a public image that may have stolen a target image in a short time and with high accuracy.
  • the first similarity calculated by comparing each of the plurality of public images with the target image based on the statistic of the pixel value and / or the statistic of the feature point is predetermined.
  • a first determination unit that selects a combination of a target image and a public image that exceeds a threshold value of the target image, and, for the selected combination of the target image and the public image, a feature point of the public image is set as a predetermined feature point and target image
  • a second determination unit that selects a combination of a target image and a public image whose second similarity calculated by comparing with a feature point of the processed image subjected to the above processing exceeds a predetermined threshold.
  • FIG. 1 is a block diagram showing a processing configuration of the image identification apparatus according to the present embodiment.
  • FIG. 2 is a block diagram illustrating a hardware configuration of the image identification device.
  • the image identification device 10 includes a group creation unit 11, a public image collection unit 12, a first determination unit 13, a second determination unit 14, and a display unit 15.
  • the image identification device 10 is a device that performs processing for finding an image that depends on the target image from among a plurality of public images.
  • a public image is an image published on the Internet by a company or an individual through a web page provided by a web server.
  • the target image is a copyrighted image to be searched for whether it has been stolen on the Internet. It is assumed that the public image includes an image that depends on the target image. An image created based on the target image may depend on the target image. An image based on the target image can be stolen if it is created without the copyright holder of the target image.
  • the image identification device 10 searches for a public image presumed to be the same as the target image or a public image presumed to be based on the target image in order to find a public image in which the target image has been stolen.
  • the image identification device 10 searches for an image presumed to have stolen the target image from the public images based on the two-stage determination by the first determination unit 13 and the second determination unit 14.
  • the group creation unit 11 prepares the target image for comparison with the public image. Detailed processing of the group creation unit 11 will be described later.
  • the public image collection unit 12 collects public images. Detailed processing of the public image collection unit 12 will be described later.
  • the first determination unit 13 includes a target image in which a first similarity calculated by comparing a plurality of public images and a target image with pixel value statistics and / or feature point statistics exceeds a predetermined threshold value. Select a combination of public images.
  • the statistic of the pixel value and / or the statistic of the feature point is either or both of the statistic of the pixel value and the statistic of the feature point.
  • the first similarity may be a value calculated by comparing the statistic of the pixel value of the public image with the statistic of the pixel value of the target image.
  • the first similarity may be a value calculated by comparing the statistic of the feature point of the public image with the statistic of the feature point of the target image.
  • the first similarity is calculated by comparing the statistic of the pixel value of the public image and the statistic of the pixel value of the target image, the statistic of the feature point of the public image, and the target image. It may be a value calculated from the similarity calculated by comparing the statistics of feature points.
  • the second determination unit 14 performs a predetermined process on the feature points of the public image and the feature points of the target image and the target image.
  • a combination of the target image and the public image whose second similarity calculated by comparing with the feature point of the processed image exceeds a predetermined threshold is selected.
  • the combination of the target image and the public image selected here is a combination of the target image and the public image estimated to be created based on the target image.
  • primary selection is performed by comparing pixel value statistics and / or feature point statistics, and only the images remaining in the primary selection are compared with the target image in detail.
  • the feature points of the remaining image are compared not only with the target image itself but also with a processed image obtained by processing the target image that can be a form of theft. As a result, theft of the target image can be investigated at high speed and with high accuracy.
  • the public image and the feature point are compared for both the target image and the processed image that relies on the target image, so that an image showing the same person or the same object is simply shown.
  • the second similarity of the processed image depending on the target image can be calculated higher than that of an image belonging to a general similar image concept such as an image.
  • the calculation method is particularly effective.
  • the statistic of the pixel value is, for example, a pixel value histogram
  • the first determination unit 13 compares the pixel value histogram included in the public image Pi with the pixel value histogram of the target image Oi.
  • the first similarity is calculated.
  • a primary selection is performed based on a histogram of pixel values, and only an image remaining in the primary selection is compared with a target image in detail.
  • the feature points of the remaining image are compared not only with the target image itself but also with a processed image obtained by processing the target image that can be a form of theft. As a result, theft of the target image can be investigated at high speed and with high accuracy.
  • the first determination unit 13 determines similarity based on the first similarity based on the pixel value histogram mainly focusing on the hue
  • the second determination unit 14 mainly determines the second similarity based on the feature point focusing on the shape. Therefore, it is possible to determine the similarity without being biased to either the hue or the shape.
  • the group creation unit 11 creates a predetermined number of processed images for each of the plurality of target images by a predetermined processing method common to the plurality of target images, and the same image and the processed image of the target images. Are grouped as comparison images and associated with each of a plurality of target images.
  • the second determination unit 14 calculates a second similarity between the public image and the target image based on a comparison between the feature point of the public image and the feature point of the comparison image included in the group associated with the target image. You may decide to do it.
  • the second determination unit 14 includes a plurality of comparison images included in the public image and the group of the target images in the combination of the target image and the public image selected by the first determination unit 13. It is also possible to calculate the second similarity to the target image of the public image by calculating the similarity with respect to each of the feature points and adding the similarities.
  • the public image is compared with a plurality of comparison images, and the second similarity is calculated based on a value obtained by adding the similarities, for example, a total value. Therefore, a comprehensive determination is made based on the plurality of comparison images. Even if the plagiarism is performed by a processing method different from any of the above processing methods, the plagiarism can be estimated by comprehensive judgment.
  • the addition here includes addition of a simple total value, and addition of weights to the processing method or the comparison image and weighting the similarities for integration.
  • the weight of similarity with an image that has not been processed at all in the target image may be set larger than the weight of another comparative image that has been processed in some way.
  • the second determination unit 14 includes, in the combination of the target image and the public image selected by the first determination unit 13, the public image and a plurality of comparison images included in the target image group.
  • the second similarity of the public image with the target image may be calculated based on the maximum value of the similarities with respect to each of the feature points. If the public image is obtained by processing the target image by any of the predetermined processing methods, it is considered that the open image shows a high degree of similarity with any comparison image (processed image). Therefore, if the second similarity is calculated on the basis of the similarity between the plurality of comparison images and the maximum value, it is possible to estimate theft by the same processing method as the assumed processing method. For example, if the maximum value of the similarities with a plurality of comparative images is the second similarity, it is effective for theft estimation by the same processing method as the assumed processing method.
  • the group creation unit 11 creates a comparison image for each target image group in advance and creates a database
  • the public image collection unit 12 collects the public images in advance and creates a database.
  • the second determination unit 14 may access the database and compare the feature points of the public image collected in advance with the feature points of the comparison image created in advance.
  • the comparison image and the public image are stored in a database in advance, and their feature points are compared while accessing the database. Therefore, when both the target image and the public image exist, the comparison process is executed efficiently. Calculation time can be shortened.
  • the first determination unit 13 compares the pixel value histogram of the target image with the pixel value histogram of the public image.
  • a statistic calculated for each type from the feature points of the target image may be compared with a statistic calculated for each type from the feature points of the public image.
  • the first determination unit 13 may compare a feature point index indicating the number of feature points for each type of target image with a feature point index indicating the number of feature points for each type of public image. .
  • the first determination unit 13 uses a binary feature point index indicating whether or not there is a feature point for each type of target image, and whether or not the feature point exists for each type of public image. The feature point index shown may be compared.
  • the first determination unit 13 may calculate the first similarity using both comparison of pixel value histograms and feature point index comparison.
  • the first determination unit 13 may match the sizes of the target image and the public image before comparing the pixel value histogram or the feature point index. This facilitates the comparison operation.
  • the image identification apparatus 10 can be realized by causing a computer to execute a software program that defines the processing procedure of each unit shown in FIG.
  • FIG. 2 shows an example of a hardware configuration of a computer that implements the image identification device 10.
  • the image identification device 10 can be connected to the web server 91 via a communication network 90 such as the Internet.
  • the image identification device 10 includes a processing device 21, a main memory 22, a storage device 23, a communication device 24, an input device 25, and a display device 26, which are connected to a bus 27.
  • the storage device 23 stores data that can be written and read.
  • the storage device 23 stores image data used by the image identification device 10 for processing. For example, public image data collected from a plurality of web servers 91 is stored in the storage device 23. Further, data of a target image to be searched for whether or not the image is stolen is also recorded in the storage device 23.
  • the processing device 21 is a processor that reads data stored in the storage device 23 to the main memory 22 and executes processing of the software program using the main memory 22.
  • the processing device 21 implements the group creation unit 11, the public image collection unit 12, the first determination unit 13, the second determination unit 14, and the display unit 15 illustrated in FIG.
  • the communication device 24 transmits information processed by the processing device 21 via a communication network 90 including wired or wireless or both, and transmits information received via the communication network 90 to the processing device 21.
  • the received information is used by the processing device 21 for software processing.
  • the input device 25 is a device that accepts information by an operation input by an operator such as a keyboard and a mouse, and the input information is used by the processing device 21 for software processing.
  • the target image is input to the storage device 23 via the communication device 24 or the input device 25.
  • the display device 26 is a device that displays image and text information on the display screen in accordance with software processing by the processing device 21.
  • FIG. 3 is a flowchart of the entire process.
  • FIG. 4 is an image diagram for explaining the overall processing.
  • the storage device 23 holds data of the target image for searching for theft.
  • the image identification device 10 first performs group creation processing by the group creation unit 11 (step S101).
  • the group creation unit 11 creates a processed image of the target image stored in the storage device 23, creates a pixel value histogram of the target image, and features of the target image and the processed image (comparison image). Perform extraction.
  • the processed image is an image obtained by rotating the target image, a divided image, a corrected image, or the like. It is preferable to prepare a processed image obtained by processing the target image by a processing method assumed when the target image is processed and stolen.
  • a comparison image obtained by combining the target image and the processed image created by processing the target image is a group associated with the target image.
  • the pixel value histogram is a graph showing the distribution of pixel values of pixels in an image, with the horizontal axis representing pixel values and the vertical axis representing the number of pixels.
  • the image identification device 10 executes a public image collection process by the public image collection unit 12 (step S102).
  • the public image collection unit 12 circulates each website of the web server 91 via the communication network 90, collects data of the public image published on the web page included in the website, A pixel value histogram of the public image is created, and the public image data and the pixel value histogram data are stored in the storage device 23.
  • the public image collection process is performed after the group creation process is performed is shown as an example, but the process order is not limited to this.
  • the public image collection process may be executed before the group creation process.
  • the image identification device 10 performs a first determination process by the first determination unit 13 (step S103).
  • the first determination unit 13 first compares the pixel value histogram of the public image with the pixel value histogram of the target image, and determines the similarity (first similarity) of the respective combinations. calculate.
  • the first determination unit 13 compares the first similarity with a predetermined threshold, associates the public image with the first similarity exceeding the threshold with the target image, and records the association information in the storage device 23. To do.
  • the similarity between pixel value histograms may be calculated based on the similarity of the graph shape of the pixel value histogram. For example, for each bin of the pixel value histogram of the target image and the bin of the pixel value histogram of the public image corresponding to that bin, the deviations of the number of pixels of the bin from the average value of the number of pixels of all bins are compared. If the total value of the differences between these deviations is small, the first similarity between the target image and the public image is high.
  • the color of the target image and the entire image are compared, the color of the target image and the entire image (the tendency of the color of the entire image) regardless of the shape of the target image or the public image.
  • a public image with similar is selected primarily.
  • the pixel value histograms of the images may be normalized and then compared. By doing so, it is possible to determine whether the color of the entire image is similar even if the image size is different.
  • the first determination unit 13 compares the target image and the public image using a histogram of pixel values.
  • the first determination unit 13 may compare the target image and the public image using a feature value statistic.
  • the statistic of the feature point can be represented by a vector having a predetermined number of dimensions, and the first determination unit 13 includes a vector representing the statistic of the feature point of the target image and a vector representing the statistic of the feature point of the public image. What is necessary is just to calculate a similarity.
  • the method for calculating the similarity between vectors is not particularly limited.
  • the cosine similarity may be used as the similarity between vectors, or the dissimilarity between character strings may be regarded as vector elements.
  • the degree of similarity between vectors may be obtained from the Hamming distance representing
  • the image identification device 10 executes the second determination process by the second determination unit 14 (step S104).
  • the second determination unit 14 determines the feature points of the public image for the combination of the public image and the target image associated with the association information recorded in the storage device 23. The comparison is made with the feature points of each comparison image belonging to the group associated with the target image.
  • the comparison of the feature points of the images may be performed based on, for example, feature amounts calculated from the feature points.
  • a feature point is a characteristic point in an image. For example, there are feature points at the corners and boundaries of an object shown in the image.
  • the feature amount is a value calculated from the feature point. For example, a score weighted according to the type of feature point may be assigned to each feature point, and an integrated value obtained by integrating the scores of feature points whose positions and types match in the image may be used as the second similarity between the images.
  • the distance between the feature points of the images of the public image and each of the comparison images belonging to the target image group is calculated, the similarity is calculated so that the similarity is higher as the distance between the feature points is closer, and the group
  • the calculation of the second similarity may be defined so as to calculate the second similarity by combining the similarities.
  • the feature point comparison method used in the present embodiment is not particularly limited. Various existing feature point matching methods can also be used.
  • the feature point matching for calculating the similarity based on the distance between the feature points is weak against image rotation, and the similarity decreases if the image is rotated even in the same image. Therefore, when such a method is used, it is also effective in this respect to prepare an image obtained by rotating the target image as the processed image.
  • this comparative image includes a target image and a processed image obtained by processing the target image. For example, the similarity between the feature point of the public image and each comparison image belonging to the group associated with the target image is calculated, and based on the similarity, the similarity between the public image and the target image ( (Second similarity) is calculated. Further, the second determination unit 14 determines whether or not the public image is estimated to have stolen the target image based on the comparison result. For example, if the second similarity exceeds a predetermined threshold, the public image may be estimated to have been created by plagiarizing the target image.
  • the image identification device 10 causes the display unit 15 to display the determination result by the second determination unit 14 on the screen.
  • a score representing the second similarity used for the determination may be displayed for a public image estimated to have been created by stealing the target image.
  • the target image and a public image estimated to have been created by plagiarizing the target image may be displayed side by side.
  • FIG. 4 is an image diagram for explaining the overall processing.
  • step S101 a pixel value histogram Oh of the target image Oi is created. Further, the target image Oi and the feature point Of of the processed image are extracted. Although only one target image Oi is shown in FIG. 4, there are actually a large number of target images Oi1, Oi2,.
  • step S101 processing is performed on the large number of target images Oi1, Oi2,.
  • step S102 more public images Pi1 to Pin than the target image Oi are collected from the web server 91 via the communication network 90, and pixel value histograms Ph1 to Phn of the respective public images Pi1 to Pin are created. Feature points Pf1 to Pfn of the public images Pi1 to Pin are extracted.
  • the first similarity between the pixel value histogram Oh of the target image Oi and the pixel value histograms Ph1 to Phn of the public images Pi1 to Pin is calculated, and the first similarity exceeds the threshold value.
  • a combination of the image Oi and the public image Pi is extracted.
  • the combination of the target image Oi and the public image Pi extracted here is that it is necessary to determine whether or not the public image Pi was created by stealing the target image Oi in the second determination process. is there.
  • step S104 the combination of the target image Oi and the public image Pi extracted in step S103 is based on the target image Oi, the feature point Of of the processed image, and the feature point Pf of the public image Pi.
  • the second similarity is calculated, and a combination of the target image Oi and the public image Pi whose second similarity exceeds the threshold is extracted.
  • the combination of the target image Oi and the public image Pi extracted here is presumed that the public image Pi was created by stealing the target image Oi.
  • FIG. 5 is a flowchart of the group creation process.
  • FIG. 5 shows details of the group creation processing in step S101.
  • the group creation unit 11 creates a processed image of the target image (step S201).
  • a processed image processed by various processing methods assumed when the target image Oi is stolen and processed to be the public image Pi is preferably prepared as a comparative image.
  • a processed image an enlarged image obtained by enlarging the image size of the target image Oi, a reduced image obtained by reducing the image size of the target image Oi, a rotated image obtained by rotating the target image Oi by a predetermined angle, and a part of the target image Oi It is preferable that at least one of the divided images obtained by cutting out is included in the comparison image. Since the processed image obtained by processing the target image Oi by an easy processing method that may be used for theft can be compared with the public image Pi, the detection accuracy for the theft can be ensured with a small amount of processing.
  • the divided image is obtained by dividing the target image Oi from the target image Oi by a fixed division method, for example, dividing the target image Oi into two equal parts, two equal parts horizontally, or four equal parts vertically and horizontally.
  • a human face may be detected from the target image Oi, and a processed image may be created so as to cut out the periphery of the human face.
  • the group creation unit 11 creates a pixel value histogram of the target image Oi (step S202).
  • the pixel value histogram can be created by counting the number of pixels corresponding to the pixel value for each pixel value.
  • the group creation unit 11 extracts each feature point of the target image Oi and its processed image, that is, the comparison image (step S203).
  • the feature point extraction method is not particularly limited, and various existing methods can be applied.
  • the group creating unit 11 groups the target image Oi and the processed image obtained by processing the target image Oi as a comparison image corresponding to the target image Oi, and the pixel value histogram of the target image Oi and the target image Oi. Data of Oh, the target image Oi, and the feature point Of of the processed image is recorded in the storage device 23 (step S204).
  • FIG. 6 is a diagram showing data recorded in the recording device by the group creation processing.
  • group creation processing is performed for each of a large number of target images Oi.
  • FIG. 6 shows group data corresponding to each of the target images Oi1, Oi2, Oi3.
  • Data of the target image Oi1, the pixel value histogram Oh1, and the feature point Of1 is registered in the group Og1 corresponding to the target image Oi1, and the target image Oi2, the pixel value histogram Oh2, and the feature are registered in the group Og2 corresponding to the target image Oi2.
  • the data of the point Of2 is registered, and the data of the target image Oi3, the pixel value histogram Oh3, and the feature point Of3 are registered in the group Og3 corresponding to the target image Oi3.
  • FIG. 7 is a flowchart of the first determination process.
  • the first determination unit 13 acquires the pixel value histogram data of all target images Oi and the pixel value histogram data of all public images Pi from the storage device 23 (step S301).
  • the pixel value histogram of the target image is created in advance in the pixel value histogram creation processing in step 202, but the present invention is not limited to this.
  • a histogram of pixel values may be created in step S301.
  • the first determination unit 13 compares the pixel value histogram of the target image Oi with the pixel value histogram of the public image Pi, and calculates the similarity (first similarity) of these pixel value histograms (step S302). ).
  • the first determination unit 13 compares the pixel value histograms of all the target images Oi with the pixel value histograms of all the public images Pi.
  • the first determination unit 13 compares the calculated first similarity with a predetermined threshold, and selects a combination of the target image Oi and the public image Pi whose first similarity exceeds the threshold (step S303). .
  • the first determination unit 13 records, in the storage device 23, first selection information in which the target image Oi whose first similarity between the pixel value histograms exceeds the threshold and the public image Pi are associated with each other.
  • the threshold value to be compared with the first similarity may be a predetermined fixed value, or may be a value that can be set or changed by the user. In the first determination process, it is only necessary to narrow down the combination of the target image and the public image to be further compared and reduce the time required for the processing time based on the similarity of the pixel value histograms. Accordingly, a series of processes for extracting the similar image Pi that is estimated to have stolen the target image Oi can be completed within one cycle of repeating the process of extracting the similar image Pi that is estimated to have stolen the target image Oi.
  • the threshold value may be determined so that the number of combinations of the target image Oi and the public image Pi can be narrowed down and the public image Pi estimated to have stolen the target image Oi is not missed as much as possible at the stage of the first determination process. .
  • FIG. 8 is an image diagram for explaining the first determination process.
  • the pixel value histograms Oh1, Oh2, Oh3,... Of the target images Oi1, Oi2, Oi3,... And the pixel value histograms Ph1, Ph2 of the public images Pi1, Pi2, Pi3,. , Ph3,...
  • FIG. 8 shows a state in which the pixel value histogram Phk of the public image Pik is compared with the pixel value histograms Oh1, Oh2, and Oh3 of the target images Oi1, Oi2, and Oi3.
  • a comparison between the target image Oi and the public image Pi is indicated by a bidirectional arrow.
  • whether or not the first similarity calculated by the comparison exceeds the threshold is indicated by a circle and a cross.
  • a circle indicates that the first similarity exceeds the threshold
  • a cross indicates that the first similarity is equal to or less than the threshold.
  • the first similarity between the pixel value histogram Phk and the pixel value histogram Oh1 and the first similarity between the pixel value histogram Phk and the pixel value histogram Oh2 exceed the threshold value, and the pixel value histogram Phk and the pixel value The first similarity with the histogram Oh3 is equal to or less than the threshold value. Therefore, information on the combination of the target image Oi1 and the public image Pik and information on the combination of the target image Oi2 and the public image Pik are recorded in the storage device 23 as the first selection information. On the other hand, information on the combination of the target image Oi3 and the public image Pik is not recorded in the storage device 23 as the first selection information.
  • FIG. 9 is a flowchart of the second determination process.
  • the second determination unit 14 extracts each feature point of the public image Pi in which information on the combination of the target image Oi and the public image Pi is recorded in the first selection information (step S401).
  • the second determination unit 14 records the extracted feature point information in the storage device 23.
  • the feature points of the target image and the processed image are extracted in advance, but the present invention is not limited to this.
  • feature points of the target image and the processed image may be extracted.
  • the second determination unit 14 includes the feature points of the comparison image belonging to the group corresponding to the target image Oi, The feature point of the public image Pi is compared, and the second similarity between the target image Oi and the public image Pi is calculated (step S402).
  • the second determination unit 14 gives a predetermined score to each feature point of the public image, gives a score to the comparison image having a feature point that matches the feature point, and the score of the given score
  • the total may be used as the score of the comparison image
  • the second similarity between the target image associated with the group and the public image may be calculated based on the scores of the plurality of comparison images belonging to the group.
  • the score given to each feature point is not particularly limited. For example, the score may be equally distributed to each feature point so that the total value of the scores in the image is constant. Or you may weight the score given to the feature point by the kind of feature point.
  • the total value of scores of a plurality of comparative images belonging to the group may be set as the second similarity. Or it is good also considering the highest value of the score of the image for a comparison which belongs to a group as the 2nd similarity.
  • the second determination unit 14 compares the calculated second similarity with a predetermined threshold, and selects a combination of the target image Oi and the public image Pi whose second similarity exceeds the threshold (step S403). .
  • the second determination unit 14 records, in the storage device 23, second selection information in which the target image Oi whose second similarity exceeds the threshold and the public image Pi are associated with each other.
  • the threshold value to be compared with the second similarity may be a predetermined fixed value, or may be a value that can be set or changed by the user.
  • a threshold value suitable for extracting the public image Pi estimated to have stolen the target image Oi may be set. If the policy is to extract only images with a possibility of theft being higher than a certain level, the threshold for comparing the second similarity may be set higher. If it is assumed that an image extracted by the image identification device 10 is visually confirmed by a human to be stolen, a threshold value to be compared with the second similarity is set low, and a suspicious image is identified by the image identification device. 10 may be as little as possible.
  • FIG. 10 is an image diagram for explaining the second determination process.
  • the feature points of all the comparative images belonging to the group of the target image Oi and the characteristic points of the public image Pi And brute force.
  • FIG. 10 shows that the feature points Of1-1, Of1-2,... Belonging to the group of the target image Oi1 are compared with the feature points Pf1 of the public image Pi1 in a brute force manner. ing.
  • the group creation unit 11 extracts feature points of the target image Oi in advance.
  • the group creation unit 11 does not extract feature points of the target image Oi in advance, and the second determination unit 14 determines only the combination of the target image and the public image selected by the first determination unit 13.
  • the feature points of the comparison image belonging to the target image group may be extracted, and the feature points of the comparison image may be compared with the feature points of the public image. Since the feature points are extracted only for the target image listed in the combination of the target image selected by the first determination unit 13 and the public image, the processing can be reduced.
  • the comparison image obtained by processing the target image with the processing method assumed to be used at the time of theft is compared with the public image.
  • the comparison image processing method is different from the public image. You may decide.
  • the second similarity is calculated by a calculation method in which the similarity is higher as the distance between the feature point of the comparison image corresponding to the target image and the feature point of the public image is smaller. 90 degree rotated image, 180 degree rotated image, and 270 degree rotated image may be included.
  • the second similarity calculated from the distance between the feature points can detect the similarity of the image by a relatively simple process. However, if the image is rotated and stolen, the similarity decreases and it is difficult to detect. Therefore, by preparing a rotation image with a rotation angle that is typically assumed and using it for comparison with a public image, it is possible to ensure the detection accuracy for theft with a small amount of processing.
  • the example in which the image identification device 10 is configured by a single device has been described, but the embodiment is not limited thereto.
  • a configuration in which processing of each unit of the image identification device 10 illustrated in FIG. 1 is shared by a plurality of devices is also possible.
  • the first determination unit 13 and the second determination unit 14 may be realized by separate computers.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Databases & Information Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Mathematical Physics (AREA)
  • Image Analysis (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

An objective of the present invention is to precisely and quickly search for a published image which is based on an image of interest. Provided is an image identification device comprising: a first determination part for selecting combinations of an image of interest and a published image for which a first degree of similarity exceeds a prescribed threshold, said first degree of similarity being obtained by comparing each of a plurality of published images with the image of interest using a pixel value statistic and/or feature points; and a second determination part for, with regard to the selected combinations of the image of interest and a published image, selecting a combination of the image of interest and a published image for which a second degree of similarity exceeds a prescribed threshold, said second degree of similarity being obtained by comparing the feature points of the published image with the feature points of the image of interest and feature points of a processed image obtained by subjecting the image of interest to a prescribed process.

Description

画像識別装置、画像識別方法、および画像識別プログラムImage identification device, image identification method, and image identification program
　本発明は画像の盗用を検出するための技術に関する。 The present invention relates to a technique for detecting image theft.
　特定のテーマに関する情報をインターネット上から収集し、収集した情報を整理し、１つのページにまとめて掲載するウェブサービスがある。インターネット上では他人のコンテンツを容易に盗用できるため、この種のウェブサービスで著作権侵害が生じやすいという問題が指摘されている。特に画像はそのままあるいは簡単な加工をするだけで容易に盗用できてしまう。 There is a web service that collects information on a specific theme from the Internet, organizes the collected information, and publishes it on a single page. It has been pointed out that copyright infringement is likely to occur in this type of web service because the content of others can be easily stolen on the Internet. In particular, images can be easily stolen as they are or simply by simple processing.
　原著作物画像と他の画像との類似度を判定する類似画像判定装置が特許文献１、２に開示されている。 Patent Documents 1 and 2 disclose similar image determination apparatuses that determine the degree of similarity between an original work image and another image.
　特許文献１には、電子化された原著作物画像を用意し、この原著作物画像と同一または類似の画像を含むことが推定されるコンテンツをコンテンツ保有装置から取得するとともに、当該コンテンツにおいて原著作物画像との比較対象となる対象画像を特定し、特定された対象画像と原著作物画像との特徴量を抽出し、抽出された対象画像の特徴量に基づいたヒストグラムを生成し、当該ヒストグラムを所定の単位幅で変化させて単位幅毎の単位ヒストグラム群を生成し、生成された単位ヒストグラム群と原著作物画像の特徴量とを比較して両画像間の類似度を判定する態様が記載されている。 In Patent Document 1, an electronic original work image is prepared, and content presumed to include an image that is the same as or similar to the original work image is acquired from the content holding device, and the original work image is included in the content. A target image to be compared with the target image, extracting feature amounts of the specified target image and the original work image, generating a histogram based on the extracted feature amount of the target image, A mode is described in which a unit histogram group for each unit width is generated by changing the unit width, and the similarity between both images is determined by comparing the generated unit histogram group and the feature amount of the original work image. .
　特許文献２には、電子化された所定の原著作物画像と、該原著作物画像と同一または類似の画像を含むことが推定されるコンテンツとを取得し、取得したコンテンツにおいて原著作物画像との比較対象となる対象画像を特定し、特定された対象画像と前記原著作物画像の特徴量を比較して画像間の類似度を表す情報を生成し、類似度を表す情報を当該対象画像の取得先情報と共に出力する態様が記載されている。 In Patent Document 2, a predetermined digitized original work image and content presumed to include the same or similar image as the original work image are acquired, and the acquired content is compared with the original work image. A target image that is a target is specified, information indicating similarity between the images is generated by comparing feature quantities of the specified target image and the original work image, and information indicating the similarity is obtained from the acquisition destination of the target image. A mode of outputting together with information is described.
　また、特許文献２には、特定された対象画像及び原著作物画像の少なくとも一方に対して、拡大、縮小、回転、色調変換等の画像変換処理を施して両画像の特徴量をより近似させ、画像変換処理後の両画像の特徴量を比較する旨も記載されている。 Further, in Patent Document 2, image conversion processing such as enlargement, reduction, rotation, and tone conversion is applied to at least one of the specified target image and original work image to more closely approximate the feature amounts of both images, It also describes that feature values of both images after image conversion processing are compared.
特開平１１－１１０５５６号公報JP-A-11-110556 特開平１１－５３５４１号公報Japanese Patent Laid-Open No. 11-53541
　インターネット上に膨大な数の画像が存在するため、その中から対象画像を盗用した画像を見つける処理の演算量は膨大となる。例えば、インターネット上の数千万個の公開画像の中から数百万個の対象画像の盗用を探索する場合、数百兆回の画像マッチングを行うことになる。さらに、インターネット上の画像は日々、変化したり、増加されたりしているので、インターネット上の画像から対象画像を盗用した画像を見つける処理は、ある程度の頻度で繰り返すことが求められる。そのため、インターネット上の画像を短い時間に高い精度で調査し、対象画像を盗用した画像を探索する処理が求められる。例えば、数日あるいは数週間という周期で、高い精度の画像マッチング処理を繰り返すことが求められる。 Since there are an enormous number of images on the Internet, the amount of processing for finding an image in which the target image is stolen from among them is enormous. For example, when searching for the misuse of millions of target images from tens of millions of public images on the Internet, image matching is performed several hundred trillion times. Furthermore, since images on the Internet are changing or increasing every day, the process of finding an image in which the target image has been stolen from the images on the Internet is required to be repeated at a certain frequency. Therefore, there is a demand for processing for searching images on the Internet with high accuracy in a short time and searching for images that have stolen target images. For example, it is required to repeat image matching processing with high accuracy in a cycle of several days or weeks.
　特許文献１、２に記載の方法を用いれば、２つの画像がどの程度類似しているかを算出したり、２つの画像に同じ物体が写っていることを精度よく検出したりすることは可能である。しかしながら、それらの技術をもってしても精度と速度を両立することは困難である。 If the methods described in Patent Documents 1 and 2 are used, it is possible to calculate how similar two images are, or to accurately detect that the same object appears in the two images. is there. However, even with these techniques, it is difficult to achieve both accuracy and speed.
　本発明の目的は、対象画像を盗用した可能性のある公開画像を短時間かつ高精度で探索する技術を提供することである。 An object of the present invention is to provide a technique for searching for a public image that may have stolen a target image in a short time and with high accuracy.
　本発明のひとつの態様に従う画像識別装置は、複数の公開画像の各々と対象画像とを画素値の統計量および／または特徴点の統計量により比較することにより算出される第１類似度が所定の閾値を超える対象画像と公開画像の組合せを選択する第１判定部と、その選択された対象画像と公開画像の組合せについて、公開画像の特徴点を、対象画像の特徴点および対象画像に所定の加工を施した加工画像の特徴点と比較することにより算出される第２類似度が所定の閾値を超える対象画像と公開画像の組合せを選択する第２判定部と、を有している。 In the image identification device according to one aspect of the present invention, the first similarity calculated by comparing each of the plurality of public images with the target image based on the statistic of the pixel value and / or the statistic of the feature point is predetermined. A first determination unit that selects a combination of a target image and a public image that exceeds a threshold value of the target image, and, for the selected combination of the target image and the public image, a feature point of the public image is set as a predetermined feature point and target image A second determination unit that selects a combination of a target image and a public image whose second similarity calculated by comparing with a feature point of the processed image subjected to the above processing exceeds a predetermined threshold.
　本発明によれば、対象画像を盗用した可能性のある公開画像を短時間かつ高精度で探索することができる。 According to the present invention, it is possible to search for a public image that may have stolen the target image in a short time and with high accuracy.
画像識別装置の機能構成を示すブロック図である。It is a block diagram which shows the function structure of an image identification device. 画像識別装置のハードウェア構成を示すブロック図である。It is a block diagram which shows the hardware constitutions of an image identification device. 全体処理のフローチャートである。It is a flowchart of the whole process. 全体処理について説明するためのイメージ図である。It is an image figure for demonstrating the whole process. グループ作成処理のフローチャートである。It is a flowchart of a group creation process. グループ作成処理について説明するためのイメージ図である。It is an image figure for demonstrating a group creation process. 第１判定処理のフローチャートである。It is a flowchart of a 1st determination process. 第１判定処理について説明するためのイメージ図である。It is an image figure for demonstrating a 1st determination process. 第２判定処理のフローチャートである。It is a flowchart of a 2nd determination process. 第２判定処理について説明するためのイメージ図である。It is an image figure for demonstrating a 2nd determination process.
　以下、本発明の実施形態について図面を参照して説明する。 Hereinafter, embodiments of the present invention will be described with reference to the drawings.
　図１は、本実施形態による画像識別装置の処理構成を示すブロック図である。図２は、画像識別装置のハードウェア構成を示すブロック図である。 FIG. 1 is a block diagram showing a processing configuration of the image identification apparatus according to the present embodiment. FIG. 2 is a block diagram illustrating a hardware configuration of the image identification device.
　図１を参照すると、画像識別装置１０は、グループ作成部１１、公開画像収集部１２、第１判定部１３、第２判定部１４、および表示部１５を有している。 Referring to FIG. 1, the image identification device 10 includes a group creation unit 11, a public image collection unit 12, a first determination unit 13, a second determination unit 14, and a display unit 15.
　画像識別装置１０は、複数の公開画像の中から対象画像に依拠する画像を発見するための処理を行う装置である。公開画像は、ウェブサーバが提供するウェブページなどによって企業や個人などがインターネット上に公開している画像である。対象画像は、インターネット上で盗用されていないか探索する対象となる著作権のある画像である。公開画像の中には対象画像に依拠する画像が含まれていることが想定される。対象画像を基にして作成されている画像は対象画像に依拠するものとなりうる。対象画像に依拠する画像を、対象画像の著作権者に無断で作成すると盗用になりうる。画像識別装置１０は、対象画像を盗用した公開画像を発見するために、対象画像と同一のものと推定される公開画像や対象画像に依拠していると推定される公開画像を探索する。 The image identification device 10 is a device that performs processing for finding an image that depends on the target image from among a plurality of public images. A public image is an image published on the Internet by a company or an individual through a web page provided by a web server. The target image is a copyrighted image to be searched for whether it has been stolen on the Internet. It is assumed that the public image includes an image that depends on the target image. An image created based on the target image may depend on the target image. An image based on the target image can be stolen if it is created without the copyright holder of the target image. The image identification device 10 searches for a public image presumed to be the same as the target image or a public image presumed to be based on the target image in order to find a public image in which the target image has been stolen.
　本実施形態による画像識別装置１０は、第１判定部１３と第２判定部１４による２段階の判定により、公開画像の中から対象画像を盗用したものと推定される画像を探索する。 The image identification device 10 according to the present embodiment searches for an image presumed to have stolen the target image from the public images based on the two-stage determination by the first determination unit 13 and the second determination unit 14.
　グループ作成部１１は、対象画像に対して公開画像との比較の準備を行う。グループ作成部１１の詳細な処理は後述する。 The group creation unit 11 prepares the target image for comparison with the public image. Detailed processing of the group creation unit 11 will be described later.
　公開画像収集部１２は、公開画像を収集する。公開画像収集部１２の詳細な処理は後述する。 The public image collection unit 12 collects public images. Detailed processing of the public image collection unit 12 will be described later.
　第１判定部１３は、複数の公開画像と対象画像とを画素値の統計量および／または特徴点の統計量により比較することにより算出される第１類似度が所定の閾値を超える対象画像と公開画像の組合せを選択する。画素値の統計量および／または特徴点の統計量というのは、画素値の統計量と特徴点の統計量とのいずれか一方または両方のことである。第１類似度は公開画像の画素値の統計量と対象画像の画素値の統計量とを比較することにより算出される値であってもよい。あるいは、第１類似度は、公開画像の特徴点の統計量と対象画像の特徴点の統計量とを比較することにより算出される値であってもよい。あるいは、第１類似度は、公開画像の画素値の統計量と対象画像の画素値の統計量とを比較することにより算出される類似度と、公開画像の特徴点の統計量と対象画像の特徴点の統計量とを比較することにより算出される類似度とから算出される値であってもよい。 The first determination unit 13 includes a target image in which a first similarity calculated by comparing a plurality of public images and a target image with pixel value statistics and / or feature point statistics exceeds a predetermined threshold value. Select a combination of public images. The statistic of the pixel value and / or the statistic of the feature point is either or both of the statistic of the pixel value and the statistic of the feature point. The first similarity may be a value calculated by comparing the statistic of the pixel value of the public image with the statistic of the pixel value of the target image. Alternatively, the first similarity may be a value calculated by comparing the statistic of the feature point of the public image with the statistic of the feature point of the target image. Alternatively, the first similarity is calculated by comparing the statistic of the pixel value of the public image and the statistic of the pixel value of the target image, the statistic of the feature point of the public image, and the target image. It may be a value calculated from the similarity calculated by comparing the statistics of feature points.
　第２判定部１４は、第１判定部１３で選択された対象画像と公開画像の組合せについて、その公開画像の特徴点を、その対象画像の特徴点およびその対象画像に所定の加工を施した加工画像の特徴点と比較することにより算出される第２類似度が所定の閾値を超える対象画像と公開画像の組合せを選択する。ここで選択される対象画像と公開画像の組合せは、対象画像と、その対象画像に依拠して作成されたと推定される公開画像との組み合わせである。 For the combination of the target image and the public image selected by the first determination unit 13, the second determination unit 14 performs a predetermined process on the feature points of the public image and the feature points of the target image and the target image. A combination of the target image and the public image whose second similarity calculated by comparing with the feature point of the processed image exceeds a predetermined threshold is selected. The combination of the target image and the public image selected here is a combination of the target image and the public image estimated to be created based on the target image.
　本実施形態によれば、画素値の統計量および／または特徴点の統計量の比較により一次選択を行い、一次選択で残った画像のみ対象画像との詳細な比較を行う。詳細な比較においては、残った画像を、対象画像そのものだけでなく、盗用の態様となりうる対象画像を加工した加工画像とも、特徴点の比較を行う。その結果、高速かつ高精度で対象画像の盗用を調査することができる。 According to the present embodiment, primary selection is performed by comparing pixel value statistics and / or feature point statistics, and only the images remaining in the primary selection are compared with the target image in detail. In the detailed comparison, the feature points of the remaining image are compared not only with the target image itself but also with a processed image obtained by processing the target image that can be a form of theft. As a result, theft of the target image can be investigated at high speed and with high accuracy.
　特に、第２類似度の算出において、対象画像とその対象画像に依拠した加工画像の両方について公開画像と特徴点の比較を行うので、単に同一人物が写っている画像や同一物が写っている画像といった一般的な類似画像概念に属するような画像よりも、対象画像に依拠する加工画像の第２類似度が高く算出されうる。インターネット上には例えば同一の有名人や同一の有名観光スポットを撮影した公開画像が多数存在するが、それらの多くは類似画像であっても対象画像に依拠したものではなく、全く別個に撮影されたものである。そのような類似画像は、画像の盗用の推定を目的とする場合には検出しなくてよい。このような状況を考慮すると、単なる画像の類似判定ではなく、画像の盗用の推定を目的とする画像識別装置１０においては本実施形態のように対象画像と加工画像の両方を用いる第２類似度の算出方法は特に有効である。 In particular, in the calculation of the second similarity, the public image and the feature point are compared for both the target image and the processed image that relies on the target image, so that an image showing the same person or the same object is simply shown. The second similarity of the processed image depending on the target image can be calculated higher than that of an image belonging to a general similar image concept such as an image. There are many public images of the same celebrity or the same famous tourist spot on the Internet, for example, but many of them are similar images but not based on the target image and were taken completely separately. Is. Such a similar image may not be detected when the purpose is to estimate theft of the image. In consideration of such a situation, the second similarity using both the target image and the processed image in the image identification apparatus 10 for the purpose of estimating the theft of the image, not just the similarity determination of the image, as in this embodiment. The calculation method is particularly effective.
　また、本実施形態では、画素値の統計量は一例として画素値のヒストグラムであり、第１判定部１３は、公開画像Ｐｉに含まれる画素値のヒストグラムを対象画像Ｏｉの画素値のヒストグラムと比較することにより第１類似度を算出する。画素値のヒストグラムにより一次選択を行い、一次選択で残った画像のみ対象画像との詳細な比較を行う。詳細な比較においては、残った画像を、対象画像そのものだけでなく、盗用の態様となりうる対象画像を加工した加工画像とも、特徴点の比較を行う。その結果、高速かつ高精度で対象画像の盗用を調査することができる。 In the present embodiment, the statistic of the pixel value is, for example, a pixel value histogram, and the first determination unit 13 compares the pixel value histogram included in the public image Pi with the pixel value histogram of the target image Oi. Thus, the first similarity is calculated. A primary selection is performed based on a histogram of pixel values, and only an image remaining in the primary selection is compared with a target image in detail. In the detailed comparison, the feature points of the remaining image are compared not only with the target image itself but also with a processed image obtained by processing the target image that can be a form of theft. As a result, theft of the target image can be investigated at high speed and with high accuracy.
　また、第１判定部１３が主に色合いに着目した画素値ヒストグラムに基づく第１類似度で類否を判定し、第２判定部１４が主に形状に着目した特徴点に基づく第２類似度で類否を判定するので、色合いと形状のいずれか一方に偏ることのない類否の判定が可能である。 The first determination unit 13 determines similarity based on the first similarity based on the pixel value histogram mainly focusing on the hue, and the second determination unit 14 mainly determines the second similarity based on the feature point focusing on the shape. Therefore, it is possible to determine the similarity without being biased to either the hue or the shape.
　また、本実施形態では、グループ作成部１１は、複数の対象画像のそれぞれについて、複数の対象画像に共通する所定の加工方法で所定個の加工画像を作成し、対象画像の同一画像および加工画像を比較用画像としてグループ化して、複数の対象画像のそれぞれと紐づける。第２判定部１４は、公開画像の特徴点と、対象画像に紐づくグループに含まれる比較用画像の特徴点との比較に基づき、その公開画像とその対象画像との第２類似度を算出することにしてもよい。画像の加工方法は無限にあるが、一般に画像の盗用時にはよく用いられる加工方法というのはある程度想定が可能である。その加工方法を複数ある対象画像に共通に適用することで、共通処理により処理を単純化することができる。 In the present embodiment, the group creation unit 11 creates a predetermined number of processed images for each of the plurality of target images by a predetermined processing method common to the plurality of target images, and the same image and the processed image of the target images. Are grouped as comparison images and associated with each of a plurality of target images. The second determination unit 14 calculates a second similarity between the public image and the target image based on a comparison between the feature point of the public image and the feature point of the comparison image included in the group associated with the target image. You may decide to do it. There are an infinite number of image processing methods, but it is generally possible to assume to some extent a processing method that is generally used during image theft. By commonly applying the processing method to a plurality of target images, the processing can be simplified by common processing.
　また、本実施形態では、第２判定部１４は、第１判定部１３で選択された対象画像と公開画像の組合せにおける、その公開画像と、その対象画像のグループに含まれる複数の比較用画像のそれぞれとの特徴点に関する類似度を算出し、それらの類似度を加算することにより、公開画像の対象画像との第２類似度を算出することにしてもよい。公開画像を複数の比較用画像と比較し、それらの類似度を加算した値、例えば合計値に基づいて、第２類似度を算出するので、複数の比較用画像による総合判断となり、比較用画像の加工方法のいずれとも異なる加工方法で盗用がされた場合であっても総合判断により盗用の推定が可能となる。ここでいう加算には、単純な合計値を求める積算の他、加工方法あるいは比較用用画像にそれぞれ重みを付与し、類似度に重みづけして積算するものを含むものとする。例えば、対象画像に何も加工していない画像との類似度の重みを、何らかの加工がされた他の比較用画像の重みよりも大きく設定してもよい。 In the present embodiment, the second determination unit 14 includes a plurality of comparison images included in the public image and the group of the target images in the combination of the target image and the public image selected by the first determination unit 13. It is also possible to calculate the second similarity to the target image of the public image by calculating the similarity with respect to each of the feature points and adding the similarities. The public image is compared with a plurality of comparison images, and the second similarity is calculated based on a value obtained by adding the similarities, for example, a total value. Therefore, a comprehensive determination is made based on the plurality of comparison images. Even if the plagiarism is performed by a processing method different from any of the above processing methods, the plagiarism can be estimated by comprehensive judgment. The addition here includes addition of a simple total value, and addition of weights to the processing method or the comparison image and weighting the similarities for integration. For example, the weight of similarity with an image that has not been processed at all in the target image may be set larger than the weight of another comparative image that has been processed in some way.
　あるいは、本実施形態では、第２判定部１４は、第１判定部１３で選択された対象画像と公開画像の組合せにおける、その公開画像と、その対象画像のグループに含まれる複数の比較用画像のそれぞれとの特徴点に関する類似度のうち最大値に基づいて、公開画像の前記対象画像との第２類似度を算出することにしてもよい。公開画像が、所定の加工方法のいずれかにより対象画像を加工したものであれば、いずれかの比較用画像（加工画像）と高い類似度を示すと考えられる。そのため、複数の比較用画像との類似度のうち最大値を示したものを基に第２類似度を算出すれば、想定される加工方法と同一の加工方法による盗用の推定が可能となる。例えば、複数の比較用画像との類似度のうち最大値を第２類似度とすれば、想定される加工方法と同一の加工方法による盗用の推定に有効である。 Alternatively, in the present embodiment, the second determination unit 14 includes, in the combination of the target image and the public image selected by the first determination unit 13, the public image and a plurality of comparison images included in the target image group. The second similarity of the public image with the target image may be calculated based on the maximum value of the similarities with respect to each of the feature points. If the public image is obtained by processing the target image by any of the predetermined processing methods, it is considered that the open image shows a high degree of similarity with any comparison image (processed image). Therefore, if the second similarity is calculated on the basis of the similarity between the plurality of comparison images and the maximum value, it is possible to estimate theft by the same processing method as the assumed processing method. For example, if the maximum value of the similarities with a plurality of comparative images is the second similarity, it is effective for theft estimation by the same processing method as the assumed processing method.
　また、本実施形態では、グループ作成部１１は、各対象画像のグループの比較用画像を予め作成し、データベース化しておき、公開画像収集部１２は、公開画像を予め収集してデータベース化しておき、第２判定部１４は、データベースにアクセスし、予め収集された公開画像の特徴点と予め作成された比較用画像の特徴点とを比較することにしてもよい。比較用画像と公開画像とを予めデータベース化しておき、データベースにアクセスしながらそれらの特徴点を比較していくので、対象画像も公開画像もどちらも多数存在する場合に比較処理を効率よく実行して演算時間を短縮することができる。 In this embodiment, the group creation unit 11 creates a comparison image for each target image group in advance and creates a database, and the public image collection unit 12 collects the public images in advance and creates a database. The second determination unit 14 may access the database and compare the feature points of the public image collected in advance with the feature points of the comparison image created in advance. The comparison image and the public image are stored in a database in advance, and their feature points are compared while accessing the database. Therefore, when both the target image and the public image exist, the comparison process is executed efficiently. Calculation time can be shortened.
　なお、本実施形態では、第１判定部１３が対象画像の画素値ヒストグラムと公開画像の画素値ヒストグラムとを比較する例を示したが、これに限定されることはない。他の例として、対象画像の特徴点から種類毎に算出される統計量と、公開画像の特徴点から種類毎に算出される統計量とを比較してもよい。例えば、第１判定部１３が、対象画像の特徴点の種類毎の個数を示す特徴点インデックスと、公開画像の特徴点の種類毎の個数を示す特徴点インデックスとを比較することにしてもよい。あるいは、第１判定部１３は、対象画像の種類毎に特徴点があるか否かを２値で示す特徴点インデックスと、公開画像の種類毎にその特徴点があるか否かを２値で示す特徴点インデックスとを比較することにしてもよい。 In the present embodiment, the first determination unit 13 compares the pixel value histogram of the target image with the pixel value histogram of the public image. However, the present invention is not limited to this. As another example, a statistic calculated for each type from the feature points of the target image may be compared with a statistic calculated for each type from the feature points of the public image. For example, the first determination unit 13 may compare a feature point index indicating the number of feature points for each type of target image with a feature point index indicating the number of feature points for each type of public image. . Alternatively, the first determination unit 13 uses a binary feature point index indicating whether or not there is a feature point for each type of target image, and whether or not the feature point exists for each type of public image. The feature point index shown may be compared.
　更に、他の例として、第１判定部１３は、画素値ヒストグラムの比較と、特徴点インデックス比較の両方を用いて第１類似度を算出することにしてもよい。 Furthermore, as another example, the first determination unit 13 may calculate the first similarity using both comparison of pixel value histograms and feature point index comparison.
　また、本実施形態では、第１判定部１３は、画素値ヒストグラムあるいは特徴点インデックスを比較する前に対象画像と公開画像のサイズを一致させることにしてもよい。これにより、比較の演算が容易となる。 In the present embodiment, the first determination unit 13 may match the sizes of the target image and the public image before comparing the pixel value histogram or the feature point index. This facilitates the comparison operation.
　第１判定部１３および第２判定部１４の詳細な処理と、第１判定部１３あるいは第２判定部１４の処理に関連する他の部分の詳細な処理とについては後述する。 Detailed processing of the first determination unit 13 and the second determination unit 14 and detailed processing of other parts related to the processing of the first determination unit 13 or the second determination unit 14 will be described later.
　本実施形態による画像識別装置１０は図１に示した各部の処理手順を規定したソフトウェアプログラムをコンピュータに実行させることにより実現することも可能である。図２には、画像識別装置１０を実現するコンピュータのハードウェア構成の一例が示されている。 The image identification apparatus 10 according to the present embodiment can be realized by causing a computer to execute a software program that defines the processing procedure of each unit shown in FIG. FIG. 2 shows an example of a hardware configuration of a computer that implements the image identification device 10.
　図２を参照すると、画像識別装置１０はインターネット等の通信ネットワーク９０経由でウェブサーバ９１と接続可能である。ハードウェア構成として、画像識別装置１０は、処理装置２１、メインメモリ２２、記憶装置２３、通信装置２４、入力装置２５、および表示装置２６を有し、それらがバス２７に接続されている。 Referring to FIG. 2, the image identification device 10 can be connected to the web server 91 via a communication network 90 such as the Internet. As a hardware configuration, the image identification device 10 includes a processing device 21, a main memory 22, a storage device 23, a communication device 24, an input device 25, and a display device 26, which are connected to a bus 27.
　記憶装置２３は、書込みおよび読み出しが可能にデータを記憶するものであって、この記憶装置２３には画像識別装置１０が処理に用いる画像のデータが記録される。例えば、複数のウェブサーバ９１から収集した公開画像のデータは記憶装置２３に蓄積される。また、その画像の盗用がされているか否か探索する対象となる対象画像のデータも記憶装置２３に記録される。 The storage device 23 stores data that can be written and read. The storage device 23 stores image data used by the image identification device 10 for processing. For example, public image data collected from a plurality of web servers 91 is stored in the storage device 23. Further, data of a target image to be searched for whether or not the image is stolen is also recorded in the storage device 23.
　処理装置２１は、記憶装置２３に記憶されたデータをメインメモリ２２に読み出し、メインメモリ２２を利用してソフトウェアプログラムの処理を実行するプロセッサである。処理装置２１によって、図１に示したグループ作成部１１、公開画像収集部１２、第１判定部１３、第２判定部１４、および表示部１５が実現される。 The processing device 21 is a processor that reads data stored in the storage device 23 to the main memory 22 and executes processing of the software program using the main memory 22. The processing device 21 implements the group creation unit 11, the public image collection unit 12, the first determination unit 13, the second determination unit 14, and the display unit 15 illustrated in FIG.
　通信装置２４は、処理装置２１にて処理された情報を有線または無線あるいはそれら両方を含む通信ネットワーク９０を介して送信し、また通信ネットワーク９０を介して受信した情報を処理装置２１に伝達する。受信した情報は処理装置２１にてソフトウェアの処理に利用される。 The communication device 24 transmits information processed by the processing device 21 via a communication network 90 including wired or wireless or both, and transmits information received via the communication network 90 to the processing device 21. The received information is used by the processing device 21 for software processing.
　入力装置２５は、キーボードやマウスなどオペレータによる操作入力による情報を受け付ける装置であり、入力された情報は処理装置２１にてソフトウェア処理に利用される。例えば、対象画像は通信装置２４や入力装置２５を介して記憶装置２３に入力される。 The input device 25 is a device that accepts information by an operation input by an operator such as a keyboard and a mouse, and the input information is used by the processing device 21 for software processing. For example, the target image is input to the storage device 23 via the communication device 24 or the input device 25.
　表示装置２６は、処理装置２１によるソフトウェア処理に伴って画像やテキストの情報をディスプレイ画面に表示する装置である。 The display device 26 is a device that displays image and text information on the display screen in accordance with software processing by the processing device 21.
　図３は、全体処理のフローチャートである。図４は、全体処理について説明するためのイメージ図である。図３の全体処理が開始される前に記憶装置２３には盗用を探索する対象画像のデータが保持されている。 FIG. 3 is a flowchart of the entire process. FIG. 4 is an image diagram for explaining the overall processing. Before the entire process of FIG. 3 is started, the storage device 23 holds data of the target image for searching for theft.
　図３を参照すると、画像識別装置１０は、まず、グループ作成部１１によりグループ作成処理を行う（ステップＳ１０１）。グループ作成処理において、グループ作成部１１は、記憶装置２３に格納されている対象画像の加工画像の作成と、対象画像の画素値ヒストグラム作成と、対象画像および加工画像（比較用画像）の特徴点抽出と、を行っておく。 Referring to FIG. 3, the image identification device 10 first performs group creation processing by the group creation unit 11 (step S101). In the group creation process, the group creation unit 11 creates a processed image of the target image stored in the storage device 23, creates a pixel value histogram of the target image, and features of the target image and the processed image (comparison image). Perform extraction.
　加工画像は、対象画像を回転した画像、分割した画像、補正した画像、などである。対象画像を加工して盗用する場合に想定される加工方法で対象画像を加工した加工画像を準備しておくことが好ましい。対象画像とその対象画像を加工して作成した加工画像とを合わせた比較用画像がその対象画像に紐づくグループとなる。 The processed image is an image obtained by rotating the target image, a divided image, a corrected image, or the like. It is preferable to prepare a processed image obtained by processing the target image by a processing method assumed when the target image is processed and stolen. A comparison image obtained by combining the target image and the processed image created by processing the target image is a group associated with the target image.
　画素値ヒストグラムは、画像中の画素の画素値の分布を、横軸に画素値をとり縦軸に画素数をとってグラフにしたものである。ここでは画素値ヒストグラムを描画可能にする画素値の度数分布データを作成すればよい。 The pixel value histogram is a graph showing the distribution of pixel values of pixels in an image, with the horizontal axis representing pixel values and the vertical axis representing the number of pixels. Here, it is only necessary to create frequency distribution data of pixel values that can render a pixel value histogram.
　次に、画像識別装置１０は、公開画像収集部１２により、公開画像収集処理を実行する（ステップＳ１０２）。公開画像収集処理にて、公開画像収集部１２は、通信ネットワーク９０経由でウェブサーバ９１の各ウェブサイトを巡回し、ウェブサイトに含まれるウェブページで公開されている公開画像のデータを収集し、公開画像の画素値ヒストグラムを作成し、公開画像のデータおよび画素値ヒストグラムのデータを記憶装置２３に蓄積する。 Next, the image identification device 10 executes a public image collection process by the public image collection unit 12 (step S102). In the public image collection process, the public image collection unit 12 circulates each website of the web server 91 via the communication network 90, collects data of the public image published on the web page included in the website, A pixel value histogram of the public image is created, and the public image data and the pixel value histogram data are stored in the storage device 23.
　なお、本実施形態では、一例としてグループ作成処理を行った後に公開画像収集処理を行う例を示すが、この処理順序に限定されることはない。公開画像収集処理をグループ作成処理の前に実行してもよい。 In this embodiment, an example in which the public image collection process is performed after the group creation process is performed is shown as an example, but the process order is not limited to this. The public image collection process may be executed before the group creation process.
　次に、画像識別装置１０は、第１判定部１３により、第１判定処理を実行する（ステップＳ１０３）。第１判定処理にて、第１判定部１３は、まず、公開画像の画素値ヒストグラムと対象画像の画素値ヒストグラムとを総当たりで比較し、それぞれの組合せの類似度（第１類似度）を算出する。次に、第１判定部１３は、第１類似度を所定の閾値と比較し、第１類似度が閾値を越えている公開画像を対象画像に対応付け、対応づけ情報を記憶装置２３に記録する。 Next, the image identification device 10 performs a first determination process by the first determination unit 13 (step S103). In the first determination process, the first determination unit 13 first compares the pixel value histogram of the public image with the pixel value histogram of the target image, and determines the similarity (first similarity) of the respective combinations. calculate. Next, the first determination unit 13 compares the first similarity with a predetermined threshold, associates the public image with the first similarity exceeding the threshold with the target image, and records the association information in the storage device 23. To do.
　画素値ヒストグラム同士の類似度は、画素値ヒストグラムのグラフ形状の類似性に基づいて算出することにしてもよい。例えば、対象画像の画素値ヒストグラムの各ビンとそのビンに対応する公開画像の画素値ヒストグラムのビンとについて、当該ビンの画素数の全ビンの画素数の平均値からの偏差同士を比較する。それら偏差同士の差分の合計値が小さければ、対象画像と公開画像の第１類似度は高い。 The similarity between pixel value histograms may be calculated based on the similarity of the graph shape of the pixel value histogram. For example, for each bin of the pixel value histogram of the target image and the bin of the pixel value histogram of the public image corresponding to that bin, the deviations of the number of pixels of the bin from the average value of the number of pixels of all bins are compared. If the total value of the differences between these deviations is small, the first similarity between the target image and the public image is high.
　ここで、画素値のヒストグラムを比較しているので、対象画像や公開画像にどんな形状の物が写っているかとは関係なく、対象画像と、画像全体の色合い（画像全体としての色の傾向）が似ている公開画像が一次選択されることになる。 Here, since the histograms of the pixel values are compared, the color of the target image and the entire image (the tendency of the color of the entire image) regardless of the shape of the target image or the public image. A public image with similar is selected primarily.
　なお、画像の画素値ヒストグラムを正規化してから、それらを比較することにしてもよい。そうすることで画像サイズが違っていても画像全体の色合いが似ているかどうかを判断することが可能となる。 Note that the pixel value histograms of the images may be normalized and then compared. By doing so, it is possible to determine whether the color of the entire image is similar even if the image size is different.
　本実施形態では、第１判定部１３は、対象画像と公開画像とを画素値のヒストグラムにより比較するものとしたが、上述のように、他の例として特徴量の統計量により比較してもよい。特徴点の統計量は所定次元数のベクトルにより表すことができ、第１判定部１３は、対象画像の特徴点の統計量を表すベクトルと、公開画像の特徴点の統計量を表すベクトルとの類似度を算出すればよい。ベクトル間の類似度を算出する手法は特に限定されないが、例えばコサイン類似度をベクトル間の類似度として用いても良いし、ベクトルの要素を文字列と見たてて文字列間の非類似度を表すハミング距離からベクトル間の類似度を求めてもよい。 In the present embodiment, the first determination unit 13 compares the target image and the public image using a histogram of pixel values. However, as described above, the first determination unit 13 may compare the target image and the public image using a feature value statistic. Good. The statistic of the feature point can be represented by a vector having a predetermined number of dimensions, and the first determination unit 13 includes a vector representing the statistic of the feature point of the target image and a vector representing the statistic of the feature point of the public image. What is necessary is just to calculate a similarity. The method for calculating the similarity between vectors is not particularly limited. For example, the cosine similarity may be used as the similarity between vectors, or the dissimilarity between character strings may be regarded as vector elements. The degree of similarity between vectors may be obtained from the Hamming distance representing
　次に、画像識別装置１０は、第２判定部１４により、第２判定処理を実行する（ステップＳ１０４）。第２判定処理にて、第２判定部１４は、記憶装置２３に記録されている対応づけ情報にて対応づけられている公開画像と対象画像の組合せについて、その公開画像の特徴点を、その対象画像に紐づけられたグループに属する各比較用画像の特徴点と比較する。 Next, the image identification device 10 executes the second determination process by the second determination unit 14 (step S104). In the second determination process, the second determination unit 14 determines the feature points of the public image for the combination of the public image and the target image associated with the association information recorded in the storage device 23. The comparison is made with the feature points of each comparison image belonging to the group associated with the target image.
　画像の特徴点の比較は、例えば特徴点から算出される特徴量に基づいて行ってもよい。特徴点とは、画像内の特徴的な点のことである。例えば、画像に写っている物体の角や境界などに特徴点が存在する。特徴量は、特徴点から算出される値である。例えば、特徴点の種類により重み付けしたスコアを各特徴点に付与し、画像の中で位置と種類が一致した特徴点のスコアを積算した積算値を画像同士の第２類似度としてもよい。 The comparison of the feature points of the images may be performed based on, for example, feature amounts calculated from the feature points. A feature point is a characteristic point in an image. For example, there are feature points at the corners and boundaries of an object shown in the image. The feature amount is a value calculated from the feature point. For example, a score weighted according to the type of feature point may be assigned to each feature point, and an integrated value obtained by integrating the scores of feature points whose positions and types match in the image may be used as the second similarity between the images.
　また、公開画像と対象画像のグループに属する比較用画像それぞれとの画像同士の特徴点の距離を算出し、特徴点同士の距離が近いほど類似度が高くなるように類似度を算出し、グループの類似度を総合して第２類似度を算出するように第２類似度の演算を定義してもよい。 Further, the distance between the feature points of the images of the public image and each of the comparison images belonging to the target image group is calculated, the similarity is calculated so that the similarity is higher as the distance between the feature points is closer, and the group The calculation of the second similarity may be defined so as to calculate the second similarity by combining the similarities.
　本実施形態で用いる特徴点の比較手法は特に限定されない。既存の様々な特徴点マッチングの手法を用いることもできる。 The feature point comparison method used in the present embodiment is not particularly limited. Various existing feature point matching methods can also be used.
　なお、特徴点同士の距離で類似度を算出する特徴点マッチングは画像の回転に弱く、同じ画像でも画像が回転していると類似度が低下してしまう。そのため、そのような手法を用いる場合には、加工画像として、対象画像を回転させた画像を準備しておくことは、その点でも効果的である。 Note that the feature point matching for calculating the similarity based on the distance between the feature points is weak against image rotation, and the similarity decreases if the image is rotated even in the same image. Therefore, when such a method is used, it is also effective in this respect to prepare an image obtained by rotating the target image as the processed image.
　上述の通り、この比較用画像には対象画像およびその対象画像を加工した加工画像が含まれている。例えば、公開画像の特徴点と、対象画像に紐づけられたグループに属する各比較用画像との類似度をそれぞれ算出し、それらの類似度に基づいて、公開画像と対象画像との類似度（第２類似度）を算出する。更に、第２判定部１４は、比較結果に基づいて、その公開画像がその対象画像を盗用したものと推定されるか否か判定する。例えば、第２類似度が所定の閾値を越えていたら、その公開画像はその対象画像を盗用して作成されたと推定するものとしてもよい。 As described above, this comparative image includes a target image and a processed image obtained by processing the target image. For example, the similarity between the feature point of the public image and each comparison image belonging to the group associated with the target image is calculated, and based on the similarity, the similarity between the public image and the target image ( (Second similarity) is calculated. Further, the second determination unit 14 determines whether or not the public image is estimated to have stolen the target image based on the comparison result. For example, if the second similarity exceeds a predetermined threshold, the public image may be estimated to have been created by plagiarizing the target image.
　次に、画像識別装置１０は、表示部１５により、第２判定部１４による判定結果を画面に表示する。例えば、対象画像を盗用して作成したと推定される公開画像について、判定に用いた第２類似度を意味するスコアを表示してもよい。また、例えば、対象画像と、その対象画像を盗用して作成したと推定される公開画像とを並べて表示することにしてもよい。 Next, the image identification device 10 causes the display unit 15 to display the determination result by the second determination unit 14 on the screen. For example, a score representing the second similarity used for the determination may be displayed for a public image estimated to have been created by stealing the target image. Further, for example, the target image and a public image estimated to have been created by plagiarizing the target image may be displayed side by side.
　図４は、全体処理について説明するためのイメージ図である。ステップＳ１０１では、対象画像Ｏｉの画素値ヒストグラムＯｈが作成される。また、対象画像Ｏｉおよびその加工画像の特徴点Ｏｆが抽出される。図４には、対象画像Ｏｉが１つだけしか示されていないが、実際には多数の対象画像Ｏｉ１、Ｏｉ２、・・・が存在する。ステップＳ１０１では、それら多数の対象画像Ｏｉ１、Ｏｉ２、・・・に対して処理が行われる。 FIG. 4 is an image diagram for explaining the overall processing. In step S101, a pixel value histogram Oh of the target image Oi is created. Further, the target image Oi and the feature point Of of the processed image are extracted. Although only one target image Oi is shown in FIG. 4, there are actually a large number of target images Oi1, Oi2,. In step S101, processing is performed on the large number of target images Oi1, Oi2,.
　ステップＳ１０２では、通信ネットワーク９０経由でウェブサーバ９１から、対象画像Ｏｉよりも更に多数の公開画像Ｐｉ１～Ｐｉｎが収集され、各公開画像Ｐｉ１～Ｐｉｎの画素値ヒストグラムＰｈ１～Ｐｈｎが作成され、更に各公開画像Ｐｉ１～Ｐｉｎの特徴点Ｐｆ１～Ｐｆｎが抽出される。 In step S102, more public images Pi1 to Pin than the target image Oi are collected from the web server 91 via the communication network 90, and pixel value histograms Ph1 to Phn of the respective public images Pi1 to Pin are created. Feature points Pf1 to Pfn of the public images Pi1 to Pin are extracted.
　ステップＳ１０３の第１判定処理では、対象画像Ｏｉの画素値ヒストグラムＯｈと公開画像Ｐｉ１～Ｐｉｎの画素値ヒストグラムＰｈ１～Ｐｈｎとの第１類似度が算出され、第１類似度が閾値を越えた対象画像Ｏｉと公開画像Ｐｉの組合せが抽出される。ここで抽出される対象画像Ｏｉと公開画像Ｐｉの組合せは、公開画像Ｐｉが対象画像Ｏｉを盗用して作成されたものか否か更に第２判定処理で判定する必要があるとされたものである。 In the first determination process in step S103, the first similarity between the pixel value histogram Oh of the target image Oi and the pixel value histograms Ph1 to Phn of the public images Pi1 to Pin is calculated, and the first similarity exceeds the threshold value. A combination of the image Oi and the public image Pi is extracted. The combination of the target image Oi and the public image Pi extracted here is that it is necessary to determine whether or not the public image Pi was created by stealing the target image Oi in the second determination process. is there.
　ステップＳ１０４の第２判定処理では、ステップＳ１０３で抽出された対象画像Ｏｉと公開画像Ｐｉの組合せについて、対象画像Ｏｉおよびその加工画像の特徴点Ｏｆと、公開画像Ｐｉの特徴点Ｐｆとに基づいて第２類似度が算出され、第２類似度が閾値を越えた対象画像Ｏｉと公開画像Ｐｉの組合せが抽出される。ここで抽出される対象画像Ｏｉと公開画像Ｐｉの組合せは、その公開画像Ｐｉがその対象画像Ｏｉを盗用して作成されたものと推定されたものである。 In the second determination process of step S104, the combination of the target image Oi and the public image Pi extracted in step S103 is based on the target image Oi, the feature point Of of the processed image, and the feature point Pf of the public image Pi. The second similarity is calculated, and a combination of the target image Oi and the public image Pi whose second similarity exceeds the threshold is extracted. The combination of the target image Oi and the public image Pi extracted here is presumed that the public image Pi was created by stealing the target image Oi.
　図５は、グループ作成処理のフローチャートである。図５には、ステップＳ１０１によるグループ作成処理の詳細が示されている。 FIG. 5 is a flowchart of the group creation process. FIG. 5 shows details of the group creation processing in step S101.
　図５を参照すると、グループ作成部１１は、対象画像の加工画像を作成する（ステップＳ２０１）。ここでは、対象画像Ｏｉを盗用し加工して公開画像Ｐｉとする場合に想定される各種加工方法で加工した加工画像を比較用画像として用意しておくのが良い。例えば、加工画像として、対象画像Ｏｉの画像サイズを拡大した拡大画像、対象画像Ｏｉの画像サイズを縮小した縮小画像、対象画像Ｏｉを所定角度だけ回転させた回転画像、および対象画像Ｏｉの一部を切り出した分割画像、の少なくとも１つを比較用画像に含むのが好ましい。盗用に利用される可能性がある容易な加工方法で対象画像Ｏｉを加工した加工画像を公開画像Ｐｉと比較することができるので、少ない処理量で盗用の検出精度を確保することができる。 Referring to FIG. 5, the group creation unit 11 creates a processed image of the target image (step S201). Here, a processed image processed by various processing methods assumed when the target image Oi is stolen and processed to be the public image Pi is preferably prepared as a comparative image. For example, as a processed image, an enlarged image obtained by enlarging the image size of the target image Oi, a reduced image obtained by reducing the image size of the target image Oi, a rotated image obtained by rotating the target image Oi by a predetermined angle, and a part of the target image Oi It is preferable that at least one of the divided images obtained by cutting out is included in the comparison image. Since the processed image obtained by processing the target image Oi by an easy processing method that may be used for theft can be compared with the public image Pi, the detection accuracy for the theft can be ensured with a small amount of processing.
　なお、上記分割画像は、対象画像Ｏｉを例えば、縦に２等分したり、横に2等分したり、縦横に４等分したりなど固定的な分割方法で対象画像Ｏｉから加工画像を作成してもよいし、対象画像Ｏｉから人の顔を検出し、人の顔の周辺を切り取るように加工画像を作成してもよい。 Note that the divided image is obtained by dividing the target image Oi from the target image Oi by a fixed division method, for example, dividing the target image Oi into two equal parts, two equal parts horizontally, or four equal parts vertically and horizontally. Alternatively, a human face may be detected from the target image Oi, and a processed image may be created so as to cut out the periphery of the human face.
　次に、グループ作成部１１は、対象画像Ｏｉの画素値ヒストグラムを作成する（ステップＳ２０２）。画素値ヒストグラムの作成は、画素値毎に、その画素値に該当する画素の個数をカウントすれよばい。 Next, the group creation unit 11 creates a pixel value histogram of the target image Oi (step S202). The pixel value histogram can be created by counting the number of pixels corresponding to the pixel value for each pixel value.
　次に、グループ作成部１１は、対象画像Ｏｉおよびその加工画像、すなわち比較用画像のそれぞれの特徴点を抽出する（ステップＳ２０３）。特徴点の抽出方法は特に限定せず、既存の様々な方法を適用することができる。 Next, the group creation unit 11 extracts each feature point of the target image Oi and its processed image, that is, the comparison image (step S203). The feature point extraction method is not particularly limited, and various existing methods can be applied.
　次に、グループ作成部１１は、対象画像Ｏｉとその対象画像Ｏｉを加工した加工画像とを、その対象画像Ｏｉに対応する比較用画像としてグループ化し、対象画像Ｏｉ、対象画像Ｏｉの画素値ヒストグラムＯｈ、対象画像Ｏｉおよび加工画像の特徴点Ｏｆのデータを記憶装置２３に記録する（ステップＳ２０４）。 Next, the group creating unit 11 groups the target image Oi and the processed image obtained by processing the target image Oi as a comparison image corresponding to the target image Oi, and the pixel value histogram of the target image Oi and the target image Oi. Data of Oh, the target image Oi, and the feature point Of of the processed image is recorded in the storage device 23 (step S204).
　図６は、グループ作成処理により記録装置に記録されるデータを示す図である。本実施形態の画像識別装置１０は多数の対象画像Ｏｉのそれぞれについてグループ作成処理が実施される。図６には、対象画像Ｏｉ１、Ｏｉ２、Ｏｉ３・・・のそれぞれに対応するグループのデータが示されている。対象画像Ｏｉ１に対応するグループＯｇ１には、対象画像Ｏｉ１、画素値ヒストグラムＯｈ１、特徴点Ｏｆ１のデータが登録され、対象画像Ｏｉ２に対応するグループＯｇ２には、対象画像Ｏｉ２、画素値ヒストグラムＯｈ２、特徴点Ｏｆ２のデータが登録され、対象画像Ｏｉ３に対応するグループＯｇ３には、対象画像Ｏｉ３、画素値ヒストグラムＯｈ３、特徴点Ｏｆ３のデータが登録されている。 FIG. 6 is a diagram showing data recorded in the recording device by the group creation processing. In the image identification device 10 of the present embodiment, group creation processing is performed for each of a large number of target images Oi. FIG. 6 shows group data corresponding to each of the target images Oi1, Oi2, Oi3. Data of the target image Oi1, the pixel value histogram Oh1, and the feature point Of1 is registered in the group Og1 corresponding to the target image Oi1, and the target image Oi2, the pixel value histogram Oh2, and the feature are registered in the group Og2 corresponding to the target image Oi2. The data of the point Of2 is registered, and the data of the target image Oi3, the pixel value histogram Oh3, and the feature point Of3 are registered in the group Og3 corresponding to the target image Oi3.
　図７は、第１判定処理のフローチャートである。図７を参照すると、第１判定部１３は、記憶装置２３から全対象画像Ｏｉの画素値ヒストグラムのデータと、全公開画像Ｐｉの画素値ヒストグラムのデータとを取得する（ステップＳ３０１）。なお、本実施形態では、ステップ２０２の画素値ヒストグラム作成処理にて、予め対象画像の画素値ヒストグラムを作成しておくことにしているが、これに限定されることはない。他の例として、ステップＳ３０１にて画素値のヒストグラムを作成することにしてもよい。 FIG. 7 is a flowchart of the first determination process. Referring to FIG. 7, the first determination unit 13 acquires the pixel value histogram data of all target images Oi and the pixel value histogram data of all public images Pi from the storage device 23 (step S301). In the present embodiment, the pixel value histogram of the target image is created in advance in the pixel value histogram creation processing in step 202, but the present invention is not limited to this. As another example, a histogram of pixel values may be created in step S301.
　次に、第１判定部１３は、対象画像Ｏｉの画素値ヒストグラムと公開画像Ｐｉの画素値ヒストグラムとを比較し、それらの画素値ヒストグラムの類似度（第１類似度）を算出する（ステップＳ３０２）。第１判定部１３は、全ての対象画像Ｏｉの画素値ヒストグラムと、全ての公開画像Ｐｉの画素値ヒストグラムとを総当たりで比較する。 Next, the first determination unit 13 compares the pixel value histogram of the target image Oi with the pixel value histogram of the public image Pi, and calculates the similarity (first similarity) of these pixel value histograms (step S302). ). The first determination unit 13 compares the pixel value histograms of all the target images Oi with the pixel value histograms of all the public images Pi.
　次に、第１判定部１３は、算出した第１類似度を所定の閾値と比較し、第１類似度が閾値を越えている対象画像Ｏｉと公開画像Ｐｉの組合せを選択する（ステップＳ３０３）。第１判定部１３は、画素値ヒストグラム同士の第１類似度が閾値を越えた対象画像Ｏｉと公開画像Ｐｉとを対応づけた第１選択情報を記憶装置２３に記録する。 Next, the first determination unit 13 compares the calculated first similarity with a predetermined threshold, and selects a combination of the target image Oi and the public image Pi whose first similarity exceeds the threshold (step S303). . The first determination unit 13 records, in the storage device 23, first selection information in which the target image Oi whose first similarity between the pixel value histograms exceeds the threshold and the public image Pi are associated with each other.
　第１類似度と比較する閾値は、予め定めた固定値を用いても良いし、ユーザが設定したり変更したりできる値であってもよい。第１判定処理では、画素値ヒストグラムの類似性により、その後で更に比較処理を行う対象画像と公開画像の組合せを絞り込み、処理時間に要する時間を短縮すればよい。したがって、対象画像Ｏｉを盗用したと推定される類似画像Ｐｉを抽出する処理を繰り返す１周期の間に、対象画像Ｏｉを盗用したと推定される類似画像Ｐｉを抽出する一連の処理を完了できるだけの対象画像Ｏｉと公開画像Ｐｉの組合せの個数に絞り込むことができ、かつ、対象画像Ｏｉを盗用したと推定される公開画像Ｐｉを第１判定処理の段階でできるだけ見逃さないように閾値を決めるとよい。 The threshold value to be compared with the first similarity may be a predetermined fixed value, or may be a value that can be set or changed by the user. In the first determination process, it is only necessary to narrow down the combination of the target image and the public image to be further compared and reduce the time required for the processing time based on the similarity of the pixel value histograms. Accordingly, a series of processes for extracting the similar image Pi that is estimated to have stolen the target image Oi can be completed within one cycle of repeating the process of extracting the similar image Pi that is estimated to have stolen the target image Oi. The threshold value may be determined so that the number of combinations of the target image Oi and the public image Pi can be narrowed down and the public image Pi estimated to have stolen the target image Oi is not missed as much as possible at the stage of the first determination process. .
　図８は、第１判定処理について説明するためのイメージ図である。第１判定処理では、対象画像Ｏｉ１、Ｏｉ２、Ｏｉ３、・・・の画素値ヒストグラムＯｈ１、Ｏｈ２、Ｏｈ３、・・・と、公開画像Ｐｉ１、Ｐｉ２、Ｐｉ３、・・・の画素値ヒストグラムＰｈ１、Ｐｈ２、Ｐｈ３、・・・とを総当たりで比較する。図８は、公開画像Ｐｉｋの画素値ヒストグラムＰｈｋと、対象画像Ｏｉ１、Ｏｉ２、Ｏｉ３の画素値ヒストグラムＯｈ１、Ｏｈ２、Ｏｈ３とを比較する様子が示されている。対象画像Ｏｉと公開画像Ｐｉとの比較が双方向矢印により示されている。また、その比較により算出された第１類似度が閾値を越えたか否かが丸印とバツ印で示されている。丸印は第１類似度が閾値を越えたことを示し、バツ印は第１類似度が閾値以下であったことを示している。 FIG. 8 is an image diagram for explaining the first determination process. In the first determination process, the pixel value histograms Oh1, Oh2, Oh3,... Of the target images Oi1, Oi2, Oi3,..., And the pixel value histograms Ph1, Ph2 of the public images Pi1, Pi2, Pi3,. , Ph3,... FIG. 8 shows a state in which the pixel value histogram Phk of the public image Pik is compared with the pixel value histograms Oh1, Oh2, and Oh3 of the target images Oi1, Oi2, and Oi3. A comparison between the target image Oi and the public image Pi is indicated by a bidirectional arrow. In addition, whether or not the first similarity calculated by the comparison exceeds the threshold is indicated by a circle and a cross. A circle indicates that the first similarity exceeds the threshold, and a cross indicates that the first similarity is equal to or less than the threshold.
　図８の例では、画素値ヒストグラムＰｈｋと画素値ヒストグラムＯｈ１との第１類似度、および画素値ヒストグラムＰｈｋと画素値ヒストグラムＯｈ２との第１類似度が閾値を超え、画素値ヒストグラムＰｈｋと画素値ヒストグラムＯｈ３との第１類似度が閾値以下となっている。そのため、対象画像Ｏｉ１と公開画像Ｐｉｋの組合せと、対象画像Ｏｉ２と公開画像Ｐｉｋの組合せの情報は第１選択情報として記憶装置２３に記録される。一方、対象画像Ｏｉ３と公開画像Ｐｉｋの組合せの情報は第１選択情報として記憶装置２３に記録されない。 In the example of FIG. 8, the first similarity between the pixel value histogram Phk and the pixel value histogram Oh1 and the first similarity between the pixel value histogram Phk and the pixel value histogram Oh2 exceed the threshold value, and the pixel value histogram Phk and the pixel value The first similarity with the histogram Oh3 is equal to or less than the threshold value. Therefore, information on the combination of the target image Oi1 and the public image Pik and information on the combination of the target image Oi2 and the public image Pik are recorded in the storage device 23 as the first selection information. On the other hand, information on the combination of the target image Oi3 and the public image Pik is not recorded in the storage device 23 as the first selection information.
　図９は、第２判定処理のフローチャートである。図９を参照すると、第２判定部１４は、第１選択情報に対象画像Ｏｉと公開画像Ｐｉの組合せの情報が記録されている公開画像Ｐｉのそれぞれの特徴点を抽出する（ステップＳ４０１）。第２判定部１４は、抽出した特徴点の情報を記憶装置２３に記録する。なお、本実施形態では、ステップ２０３の特徴点抽出処理にて、予め対象画像およびその加工画像の特徴点を抽出しておくことにしているが、これに限定されることはない。他の例として、ステップＳ４０１にて、対象画像および加工画像の特徴点を抽出することにしてもよい。 FIG. 9 is a flowchart of the second determination process. Referring to FIG. 9, the second determination unit 14 extracts each feature point of the public image Pi in which information on the combination of the target image Oi and the public image Pi is recorded in the first selection information (step S401). The second determination unit 14 records the extracted feature point information in the storage device 23. In this embodiment, in the feature point extraction process in step 203, the feature points of the target image and the processed image are extracted in advance, but the present invention is not limited to this. As another example, in step S401, feature points of the target image and the processed image may be extracted.
　次に、第２判定部１４は、第１選択情報に記録されている対象画像Ｏｉと公開画像Ｐｉの各組合せについて、その対象画像Ｏｉに対応するグループに属する比較用画像の特徴点と、その公開画像Ｐｉの特徴点とを比較し、その対象画像Ｏｉとその公開画像Ｐｉとの第２類似度を算出する（ステップＳ４０２）。 Next, for each combination of the target image Oi and the public image Pi recorded in the first selection information, the second determination unit 14 includes the feature points of the comparison image belonging to the group corresponding to the target image Oi, The feature point of the public image Pi is compared, and the second similarity between the target image Oi and the public image Pi is calculated (step S402).
　処理の一例として、第２判定部１４は、公開画像の各特徴点に所定のスコアを持たせ、特徴点と一致する特徴点を有する比較用画像にスコアを付与し、その付与されたスコアの合計を比較用画像のスコアとし、グループに属する複数の比較用画像のスコアに基づいて、グループに紐づく対象画像と公開画像との第２類似度を算出することにしてもよい。各特徴点に持たせるスコアは、特に限定されないが、例えば、画像内のスコアの合計値が一定になるように、各特徴点にスコアを均等分配することにしてもよい。あるいは特徴点の種類により、その特徴点に持たせるスコアに重み付けしてもよい。さらに、グループに属する複数の比較用画像のスコアの合計値を第２類似度としてもよい。あるいは、グループに属する比較用画像のスコアの最高値を第２類似度としてもよい。 As an example of the processing, the second determination unit 14 gives a predetermined score to each feature point of the public image, gives a score to the comparison image having a feature point that matches the feature point, and the score of the given score The total may be used as the score of the comparison image, and the second similarity between the target image associated with the group and the public image may be calculated based on the scores of the plurality of comparison images belonging to the group. The score given to each feature point is not particularly limited. For example, the score may be equally distributed to each feature point so that the total value of the scores in the image is constant. Or you may weight the score given to the feature point by the kind of feature point. Furthermore, the total value of scores of a plurality of comparative images belonging to the group may be set as the second similarity. Or it is good also considering the highest value of the score of the image for a comparison which belongs to a group as the 2nd similarity.
　次に、第２判定部１４は、算出した第２類似度を所定の閾値と比較し、第２類似度が閾値を越えている対象画像Ｏｉと公開画像Ｐｉの組合せを選択する（ステップＳ４０３）。第２判定部１４は、第２類似度が閾値を越えた対象画像Ｏｉと公開画像Ｐｉとを対応づけた第２選択情報を記憶装置２３に記録する。 Next, the second determination unit 14 compares the calculated second similarity with a predetermined threshold, and selects a combination of the target image Oi and the public image Pi whose second similarity exceeds the threshold (step S403). . The second determination unit 14 records, in the storage device 23, second selection information in which the target image Oi whose second similarity exceeds the threshold and the public image Pi are associated with each other.
　第２類似度と比較する閾値は、予め定めた固定値を用いても良いし、ユーザが設定したり変更したりできる値であってもよい。第２判定処理では、対象画像Ｏｉを盗用したと推定される公開画像Ｐｉを抽出するのに適した閾値を設定するとよい。盗用の可能性が一定以上に高い画像のみを抽出するというポリシーであれば、第２類似度を比較する閾値を高めに設定すればよい。画像識別装置１０が抽出した画像を人間が目視で盗用の有無を確認することが前提となっていれば、第２類似度と比較する閾値を低めに設定し、疑いのある画像を画像識別装置１０でできるだけ見逃さないようにしてもよい。 The threshold value to be compared with the second similarity may be a predetermined fixed value, or may be a value that can be set or changed by the user. In the second determination process, a threshold value suitable for extracting the public image Pi estimated to have stolen the target image Oi may be set. If the policy is to extract only images with a possibility of theft being higher than a certain level, the threshold for comparing the second similarity may be set higher. If it is assumed that an image extracted by the image identification device 10 is visually confirmed by a human to be stolen, a threshold value to be compared with the second similarity is set low, and a suspicious image is identified by the image identification device. 10 may be as little as possible.
　図１０は、第２判定処理について説明するためのイメージ図である。第２判定処理では、第１判定処理で選択された対象画像Ｏｉと公開画像Ｐｉとの全ての組み合わせについて、対象画像Ｏｉのグループに属する全ての比較用画像の特徴点と公開画像Ｐｉの特徴点とを総当たりで比較する。図１０には、対象画像Ｏｉ１のグループに属する比較用画像の特徴点Ｏｆ１－１、Ｏｆ１－２、・・・・と、公開画像Ｐｉ１の特徴点Ｐｆ１とを総当たりで比較することが示されている。 FIG. 10 is an image diagram for explaining the second determination process. In the second determination process, for all combinations of the target image Oi and the public image Pi selected in the first determination process, the feature points of all the comparative images belonging to the group of the target image Oi and the characteristic points of the public image Pi And brute force. FIG. 10 shows that the feature points Of1-1, Of1-2,... Belonging to the group of the target image Oi1 are compared with the feature points Pf1 of the public image Pi1 in a brute force manner. ing.
　なお、本実施形態では、グループ作成部１１が予め対象画像Ｏｉの特徴点を抽出することとしたが、これに限定されることはない。他の例として、グループ作成部１１は対象画像Ｏｉの特徴点の抽出を予め行わず、第２判定部１４が、第１判定部１３により選択された対象画像と公開画像の組合せのみについて、その対象画像のグループに属する比較用画像の特徴点を抽出し、その比較用画像の特徴点とその公開画像の特徴点とを比較することにしてもよい。第１判定部１３で選択された対象画像と公開画像の組合せに挙がった対象画像のみについてその特徴点を抽出するので、処理を削減することができる。 In the present embodiment, the group creation unit 11 extracts feature points of the target image Oi in advance. However, the present invention is not limited to this. As another example, the group creation unit 11 does not extract feature points of the target image Oi in advance, and the second determination unit 14 determines only the combination of the target image and the public image selected by the first determination unit 13. The feature points of the comparison image belonging to the target image group may be extracted, and the feature points of the comparison image may be compared with the feature points of the public image. Since the feature points are extracted only for the target image listed in the combination of the target image selected by the first determination unit 13 and the public image, the processing can be reduced.
　また、本実施形態では、盗用時に用いられることが想定される加工方法で対象画像を加工した比較用画像を公開画像と比較することとしたが、それとは異なる観点で比較用画像の加工方法を決めてもよい。例えば、第２類似度は、対象画像に対応する比較用画像の特徴点と公開画像の特徴点との距離が小さいほど類似度が高くなるような算出方法で算出するものとし、加工画像には、９０度回転画像、１８０度回転画像、および２７０度回転画像を含めることにしてもよい。特徴点の距離から算出される第２類似度は、比較的簡易な処理で画像の類似を検知できるが画像を回転して盗用されると類似度が低下し検出が難しくなる。そこで、典型的に想定される回転角度の回転画像を用意し、公開画像との比較に用いることで、少ない処理量で盗用の検出精度を確保することができる。 Further, in this embodiment, the comparison image obtained by processing the target image with the processing method assumed to be used at the time of theft is compared with the public image. However, the comparison image processing method is different from the public image. You may decide. For example, the second similarity is calculated by a calculation method in which the similarity is higher as the distance between the feature point of the comparison image corresponding to the target image and the feature point of the public image is smaller. 90 degree rotated image, 180 degree rotated image, and 270 degree rotated image may be included. The second similarity calculated from the distance between the feature points can detect the similarity of the image by a relatively simple process. However, if the image is rotated and stolen, the similarity decreases and it is difficult to detect. Therefore, by preparing a rotation image with a rotation angle that is typically assumed and using it for comparison with a public image, it is possible to ensure the detection accuracy for theft with a small amount of processing.
　また、本実施形態では、図２に示したように、画像識別装置１０が単体の装置により構成される例を示したが、これに実施形態が限定されることはない。図１に示した画像識別装置１０の各部の処理を複数の装置により分担する構成も可能である。例えば、第１判定部１３と第２判定部１４が別個のコンピュータにより実現されてもよい。 In the present embodiment, as shown in FIG. 2, the example in which the image identification device 10 is configured by a single device has been described, but the embodiment is not limited thereto. A configuration in which processing of each unit of the image identification device 10 illustrated in FIG. 1 is shared by a plurality of devices is also possible. For example, the first determination unit 13 and the second determination unit 14 may be realized by separate computers.
　上述した本実施形態は、本発明の説明のための例示であり、本発明の範囲をこの実施形態のみに限定する趣旨ではない。当業者は、本発明の範囲を逸脱することなしに、他の様々な態様で本発明を実施することができる。 This embodiment described above is an example for explaining the present invention, and is not intended to limit the scope of the present invention to this embodiment alone. Those skilled in the art can implement the present invention in various other modes without departing from the scope of the present invention.
１０…画像識別装置、１１…グループ作成部、１２…公開画像収集部、１３…第１判定部、１４…第２判定部、１５…表示部、２１…処理装置、２２…メインメモリ、２３…記憶装置、２４…通信装置、２５…入力装置、２６…表示装置、２７…バス、９０…通信ネットワーク、９１…ウェブサーバ DESCRIPTION OF SYMBOLS 10 ... Image identification apparatus, 11 ... Group preparation part, 12 ... Public image collection part, 13 ... 1st determination part, 14 ... 2nd determination part, 15 ... Display part, 21 ... Processing apparatus, 22 ... Main memory, 23 ... Storage device, 24 ... communication device, 25 ... input device, 26 ... display device, 27 ... bus, 90 ... communication network, 91 ... web server

Claims (13)

  1. 　複数の公開画像の各々と対象画像とを画素値の統計量および／または特徴点の統計量により比較することにより算出される第１類似度が所定の閾値を超える対象画像と公開画像の組合せを選択する第１判定部と、
    　前記選択された対象画像と公開画像の組合せについて、該公開画像の特徴点を、前記対象画像の特徴点および前記対象画像に所定の加工を施した加工画像の特徴点と比較することにより算出される第２類似度が所定の閾値を超える対象画像と公開画像の組合せを選択する第２判定部と、
    を有する画像識別装置。     A combination of a target image and a public image in which the first similarity calculated by comparing each of the plurality of public images and the target image with a statistic of pixel values and / or a statistic of feature points exceeds a predetermined threshold. A first determination unit to select;
    The combination of the selected target image and the public image is calculated by comparing the feature point of the public image with the feature point of the target image and the feature point of the processed image obtained by performing predetermined processing on the target image. A second determination unit that selects a combination of a target image and a public image whose second similarity exceeds a predetermined threshold;
    An image identification device.
  2. 　前記第１判定部は、前記公開画像に含まれる画素値のヒストグラムの情報を前記対象画像の画素値のヒストグラムの情報と比較することにより前記第１類似度を算出する、
    請求項１に記載の画像識別装置。     The first determination unit calculates the first similarity by comparing information of a histogram of pixel values included in the public image with information of a histogram of pixel values of the target image.
    The image identification device according to claim 1.
  3. 　特徴点には複数の種類があり、
    　前記第１判定部は、前記公開画像の特徴点から種類毎に算出される統計量を、前記対象画像の特徴点から種類毎に算出される統計量と比較することにより前記第１類似度を算出する、
    請求項１に記載の画像識別装置。     There are several types of feature points.
    The first determination unit compares the statistic calculated for each type from the feature points of the public image with the statistic calculated for each type from the feature points of the target image, thereby calculating the first similarity. calculate,
    The image identification device according to claim 1.
  4. 　複数の前記対象画像のそれぞれについて、前記複数の対象画像に共通する所定の加工方法で所定個の加工画像を作成し、前記対象画像の同一画像および前記加工画像を比較用画像としてグループ化して、前記複数の対象画像のそれぞれと紐づけるグループ作成部を更に有し、
    　前記第２判定部は、前記公開画像の特徴点と、前記対象画像に紐づくグループに含まれる前記比較用画像の特徴点との比較に基づき、前記公開画像と前記対象画像との前記第２類似度を算出する、
    請求項１に記載の画像識別装置。     For each of the plurality of target images, create a predetermined number of processed images by a predetermined processing method common to the plurality of target images, group the same image of the target image and the processed image as a comparison image, A group creation unit associated with each of the plurality of target images;
    The second determination unit is configured to compare the feature point of the public image with the feature point of the comparison image included in the group associated with the target image, based on a comparison between the public image and the target image. Calculate similarity,
    The image identification device according to claim 1.
  5. 　前記第２判定部は、前記公開画像と前記グループに含まれる前記複数の比較用画像のそれぞれとの特徴点に関する類似度を加算していくことにより、前記公開画像の前記対象画像との第２類似度を算出する、
    請求項４に記載の画像識別装置。     The second determination unit adds a similarity degree regarding feature points between the public image and each of the plurality of comparison images included in the group, thereby obtaining a second image of the public image with the target image. Calculate similarity,
    The image identification device according to claim 4.
  6. 　前記第２判定部は、前記公開画像と前記グループに含まれる前記複数の比較用画像のそれぞれとの特徴点に関する類似度のうち最大値に基づいて、前記公開画像の前記対象画像との第２類似度を算出する、
    請求項４に記載の画像識別装置。     A second determination unit configured to determine a second value of the target image of the public image based on a maximum value of similarities relating to feature points between the public image and each of the plurality of comparison images included in the group; Calculate similarity,
    The image identification device according to claim 4.
  7. 　前記公開画像を予め収集しておく公開画像収集部を更に有し、
    　前記グループ作成部は、前記各対象画像のグループの前記比較用画像を予め作成しておき、
    　前記第２判定部は、予め収集された前記公開画像の特徴点と予め作成された前記比較用画像の特徴点とを比較する、
    請求項４に記載の画像識別装置。     A public image collection unit for collecting the public image in advance;
    The group creation unit creates in advance the comparison image of the group of the target images,
    The second determination unit compares feature points of the public image collected in advance with feature points of the comparison image created in advance.
    The image identification device according to claim 4.
  8. 　前記第２判定部は、
    　前記公開画像の各特徴点に所定のスコアを持たせ、当該特徴点と一致する特徴点を有する比較用画像に当該スコアを付与し、
    　前記付与されたスコアの合計を当該比較用画像のスコアとし、
    　前記グループに属する複数の比較用画像のスコアに基づいて、当該グループに紐づく対象画像と前記公開画像との前記第２類似度を算出する、
    請求項７に記載の画像識別装置。     The second determination unit includes
    Give each feature point of the public image a predetermined score, give the score to a comparative image having a feature point that matches the feature point,
    The total of the given scores is used as the score of the comparison image,
    Based on the scores of a plurality of comparison images belonging to the group, the second similarity between the target image associated with the group and the public image is calculated.
    The image identification device according to claim 7.
  9. 　前記第２判定部は、前記第１判定部により選択された対象画像と公開画像の組合せのみについて、該対象画像のグループに属する比較用画像の特徴点を抽出し、該比較用画像の特徴点と前記公開画像の特徴点とを比較する、
    請求項４に記載の画像識別装置。     The second determination unit extracts feature points of the comparison image belonging to the group of the target images for only the combination of the target image and the public image selected by the first determination unit, and the feature points of the comparison image And the feature point of the published image,
    The image identification device according to claim 4.
  10. 　前記加工画像には、前記対象画像の画像サイズを拡大した拡大画像、前記対象画像の画像サイズを縮小した縮小画像、前記対象画像を所定角度だけ回転させた回転画像、および前記対象画像の一部を切り出した分割画像、の少なくとも１つが含まれる、
    請求項１に記載の画像識別装置。     The processed image includes an enlarged image obtained by enlarging the image size of the target image, a reduced image obtained by reducing the image size of the target image, a rotated image obtained by rotating the target image by a predetermined angle, and a part of the target image. At least one of the divided images obtained by cutting out
    The image identification device according to claim 1.
  11. 　前記第２類似度は、前記対象画像に対応する比較用画像の特徴点と前記公開画像の特徴点との距離が小さいほど類似度が高くなり、
    　前記加工画像には、９０度回転画像、１８０度回転画像、および２７０度回転画像が含まれる、
    請求項１に記載の画像識別装置。     The second similarity is higher as the distance between the feature point of the comparison image corresponding to the target image and the feature point of the published image is smaller,
    The processed image includes a 90-degree rotated image, a 180-degree rotated image, and a 270-degree rotated image.
    The image identification device according to claim 1.
  12. 　複数の公開画像の中から対象画像に依拠する画像を発見するための画像識別方法であって、
    　前記複数の公開画像の中から、該画像に含まれる画素値のヒストグラムを前記対象画像のヒストグラムと比較することにより算出される第１類似度が所定値以上の公開画像を選択し、
    　前記選択された公開画像の中から、該公開画像の特徴点を、前記対象画像の特徴点および前記対象画像に所定の加工を施した加工画像の特徴点と比較することにより算出される第２類似度に基づき、公開画像を選択する、
    ことをコンピュータが実行する画像識別方法。     An image identification method for finding an image depending on a target image from a plurality of public images,
    From among the plurality of public images, a public image having a first similarity calculated by comparing a histogram of pixel values included in the image with a histogram of the target image is greater than or equal to a predetermined value,
    A second calculated from the selected public image by comparing a feature point of the public image with a feature point of the target image and a feature point of a processed image obtained by performing predetermined processing on the target image. Select public images based on similarity,
    An image identification method in which a computer executes this.
  13. 　複数の公開画像の中から対象画像に依拠する画像を発見するための画像識別プログラムであって、
    　前記複数の公開画像の中から、該画像に含まれる画素値のヒストグラムを前記対象画像のヒストグラムと比較することにより算出される第１類似度が所定値以上の公開画像を選択し、
    　前記選択された公開画像の中から、該公開画像の特徴点を、前記対象画像の特徴点および前記対象画像に所定の加工を施した加工画像の特徴点と比較することにより算出される第２類似度に基づき、公開画像を選択する、
    ことをコンピュータに実行させるための画像識別プログラム。     An image identification program for finding an image depending on a target image from a plurality of public images,
    A public image having a first similarity calculated by comparing a histogram of pixel values included in the image with a histogram of the target image is selected from the plurality of public images,
    A second calculated from the selected public image by comparing a feature point of the public image with a feature point of the target image and a feature point of a processed image obtained by performing predetermined processing on the target image. Select public images based on similarity,
    An image identification program for causing a computer to execute this.
PCT/JP2019/015064 2018-04-25 2019-04-05 Image identification device, image identification method, and image identification program WO2019208158A1 (en)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
JP2018083891A JP7248968B2 (en) 2018-04-25 2018-04-25 Image identification device, image identification method, and image identification program
JP2018-083891 2018-04-25

Publications (1)

Publication Number Publication Date
WO2019208158A1 true WO2019208158A1 (en) 2019-10-31

Family

ID=68294075

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/JP2019/015064 WO2019208158A1 (en) 2018-04-25 2019-04-05 Image identification device, image identification method, and image identification program

Country Status (2)

Country Link
JP (1) JP7248968B2 (en)
WO (1) WO2019208158A1 (en)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2009282883A (en) * 2008-05-26 2009-12-03 Fujifilm Corp Image retrieval system, crawling device, and image retrieval device
JP2015022631A (en) * 2013-07-22 2015-02-02 富士ゼロックス株式会社 Information processor, information processing system and program
JP2018022293A (en) * 2016-08-02 2018-02-08 キヤノン株式会社 Image detection program and control method thereof, and program

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2004258729A (en) * 2003-02-24 2004-09-16 Casio Comput Co Ltd Image search system and program
JP5096776B2 (en) * 2007-04-04 2012-12-12 キヤノン株式会社 Image processing apparatus and image search method

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2009282883A (en) * 2008-05-26 2009-12-03 Fujifilm Corp Image retrieval system, crawling device, and image retrieval device
JP2015022631A (en) * 2013-07-22 2015-02-02 富士ゼロックス株式会社 Information processor, information processing system and program
JP2018022293A (en) * 2016-08-02 2018-02-08 キヤノン株式会社 Image detection program and control method thereof, and program

Also Published As

Publication number Publication date
JP2019191923A (en) 2019-10-31
JP7248968B2 (en) 2023-03-30

Similar Documents

Publication Publication Date Title
JP4545641B2 (en) Similar image retrieval method, similar image retrieval system, similar image retrieval program, and recording medium
US9576218B2 (en) Selecting features from image data
JP5241954B2 (en) Image search based on shape
CN108881947B (en) Method and device for detecting infringement of live stream
EP2172856A2 (en) Image processing apparatus, image processing method and program
EP2742442B1 (en) A method for detecting a copy of a reference video, corresponding apparatus for extracting a spatio-temporal signature from video data and corresponding computer readable storage medium
JP2019520662A (en) Content-based search and retrieval of trademark images
JP2008225692A (en) Image retrieval method, device and program
WO2021060147A1 (en) Similar region detection device, similar region detection method, and program
Xu et al. Robust image copy detection using multi-resolution histogram
JP5430636B2 (en) Data acquisition apparatus, method and program
EP3301603A1 (en) Improved search for data loss prevention
JP2018526754A (en) Image processing apparatus, image processing method, and storage medium
JP2011243007A (en) Image processing device, and processing method and program thereof
CN109697240A (en) A kind of image search method and device based on feature
JP5526209B2 (en) Forensic system, forensic method, and forensic program
An et al. Benchmarking the robustness of image watermarks
CN112990792A (en) Method and device for automatically detecting infringement risk and electronic equipment
WO2019208158A1 (en) Image identification device, image identification method, and image identification program
CN110083731B (en) Image retrieval method, device, computer equipment and storage medium
JP2013152543A (en) Image storage program, method and device
JP2012003358A (en) Background determination device, method, and program
JP6244887B2 (en) Information processing apparatus, image search method, and program
CN112100670A (en) Big data based privacy data grading protection method
JP7367540B2 (en) Image evaluation device and image evaluation program

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19793859

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19793859

Country of ref document: EP

Kind code of ref document: A1