WO2019085765A1 - 图像检索 - Google Patents

图像检索 Download PDF

Info

Publication number
WO2019085765A1
WO2019085765A1 PCT/CN2018/110865 CN2018110865W WO2019085765A1 WO 2019085765 A1 WO2019085765 A1 WO 2019085765A1 CN 2018110865 W CN2018110865 W CN 2018110865W WO 2019085765 A1 WO2019085765 A1 WO 2019085765A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
retrieved
feature
information entropy
entropy coding
Prior art date
Application number
PCT/CN2018/110865
Other languages
English (en)
French (fr)
Inventor
康丽萍
Original Assignee
北京三快在线科技有限公司
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 北京三快在线科技有限公司 filed Critical 北京三快在线科技有限公司
Priority to US16/762,139 priority Critical patent/US11281714B2/en
Priority to EP18874554.1A priority patent/EP3703061A4/en
Publication of WO2019085765A1 publication Critical patent/WO2019085765A1/zh

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
    • G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
    • G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
    • 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
    • G06F16/51—Indexing; Data structures therefor; Storage structures
    • 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
    • G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
    • G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
    • G06F16/5838—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content using colour
    • 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
    • G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
    • G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
    • G06F16/5854—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content using shape and object relationship
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/22—Matching criteria, e.g. proximity measures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Definitions

  • the present application relates to the field of computer technology, and in particular, to a method, an apparatus, and an electronic device for image retrieval.
  • Image retrieval is a process of finding the most similar image in a preset image database based on the image features of the image to be retrieved, and is widely used in related art.
  • a commonly used image retrieval method is to compare the image features of the image to be searched with the image features of the image stored in the image database to determine the image with the highest similarity.
  • the amount of data in image databases is getting larger and larger.
  • the retrieval efficiency of such image retrieval methods is extremely low.
  • the application provides an image retrieval method to improve the efficiency of image retrieval.
  • an embodiment of the present application provides an image retrieval method, including: determining, according to a first image feature of an image to be retrieved, a binarization feature of the image to be retrieved; a binarization feature, determining information entropy coding of the image to be retrieved; and searching for an image similar to the image to be retrieved in a preset image library based on information entropy coding of the image to be retrieved.
  • the image stored in the preset image library is indexed by information entropy coding.
  • an embodiment of the present application provides an image retrieval apparatus, including: a binarization feature acquisition module, configured to determine a binarization feature of the image to be retrieved based on a first image feature of an image to be retrieved; information entropy An encoding determining module, configured to determine information entropy encoding of the image to be retrieved based on a binarized feature of the image to be retrieved; an image retrieving module, configured to perform information entropy encoding based on the image to be retrieved, in a preset image library Searching for an image similar to the image to be retrieved.
  • the image in the preset image library is indexed by information entropy coding.
  • an embodiment of the present application further discloses an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, where the processor executes the computer program
  • the image retrieval method described in the embodiment of the present application is implemented.
  • an embodiment of the present application provides a computer readable storage medium, where a computer program is stored, and when the program is executed by a processor, the steps of the image retrieval method disclosed in the embodiment of the present application are implemented.
  • the image retrieval method disclosed in the embodiment of the present application determines the binarization feature of the image to be retrieved based on the first image feature of the image to be retrieved; and determines the to-be-searched based on the binarization feature of the image to be retrieved Entropy coding of the image; based on the information entropy coding of the image to be retrieved, retrieving an image similar to the image to be retrieved in the preset image library can effectively improve the efficiency of image retrieval.
  • the information entropy of the image-based binarization feature as the image coding, the image is indexed and retrieved, and the data amount of the comparison data is greatly reduced compared with directly comparing the binarization feature or the image feature, thereby effectively improving the image.
  • the efficiency of image retrieval is used as the image coding, the image is indexed and retrieved, and the data amount of the comparison data is greatly reduced compared with directly comparing the binarization feature or the image feature, thereby effectively improving the image. The efficiency of image retriev
  • FIG. 2 is a flowchart of an image retrieval method according to Embodiment 2 of the present application.
  • FIG. 3 is a schematic structural diagram of an image retrieval apparatus according to Embodiment 3 of the present application.
  • FIG. 4 is a second schematic structural diagram of an image retrieval apparatus according to Embodiment 3 of the present application.
  • FIG. 1 An image retrieval method disclosed in this embodiment is shown in FIG. 1 , and the method includes: Step 110 to Step 130 .
  • Step 110 Determine a binarization feature of the image to be retrieved based on a first image feature of the image to be retrieved.
  • Commonly used image features include texture features, color features, shape features, spatial relationship features, etc. Different features are applicable to different image content and different image recognition requirements.
  • Commonly used feature extraction methods include Fourier transform, Gabor transform, wavelet transform, and neural network model extraction.
  • the specific method for acquiring the image features of the image to be retrieved may be any method well known to those skilled in the art, such as extracting image features of the image to be retrieved by a convolutional neural network (CNN) model.
  • CNN convolutional neural network
  • the specific implementation method for obtaining image features of the image to be retrieved is not limited.
  • the image feature of the image to be retrieved is extracted by using a convolutional neural network model.
  • the convolutional neural network model extracts the image features of the image to be retrieved, that is, the convolutional neural network features (hereinafter also referred to as CNN features), which can better express the higher-level semantic features of the image, and has been in the field of image classification, recognition and detection. Get a good application.
  • CNN features the convolutional neural network features
  • the image retrieval technical solution is described in detail by taking the first image feature as the CNN feature as an example.
  • a binarization feature of the image to be retrieved is determined.
  • An image feature is usually a multidimensional vector.
  • the acquired first image feature of the image to be retrieved may be binarized by a binarization scheme well known to those skilled in the art to extract the binarized feature of the image to be retrieved. For example, traversing the feature value Xi of each dimension of the CNN feature, binarization of the convolutional neural network feature of the image to be retrieved according to the following rules:
  • TH is a threshold set according to experience, and the setting of the threshold is desirably such that the distribution of 0 and 1 in the binarized feature is as uneven as possible to increase the discrimination of the image, for example, may be 0.5.
  • Step 120 Determine information entropy coding of the image to be retrieved based on the binarization feature of the image to be retrieved.
  • Information entropy is often used to measure the distribution of information.
  • the information entropy of the binarization feature may be encoded as an information entropy of the image to be retrieved by the binarization feature. Then, a similar image is determined by encoding according to the information entropy.
  • Step 130 retrieve an image similar to the image to be retrieved in a preset image library based on information entropy coding of the image to be retrieved.
  • the preset image library may encode the image as an index of the image, and then, by comparing based on the information entropy coding, the similar image may be determined.
  • the information entropy coding reflects the binarization feature of the image to a certain extent, the information entropy coding is similar, and the corresponding image similarity is also high.
  • the similar image can be initially determined by comparing the information entropy coding. Further, the similarity judgment may be performed on the similar image based on the image feature, so that the image matching the image to be retrieved in the preset image library is relatively more accurately retrieved.
  • the image retrieval method disclosed in the embodiment of the present application determines the binarization feature of the image to be retrieved based on the first image feature of the image to be retrieved; and determines the binary image based on the binarization feature of the image to be retrieved.
  • Information entropy coding of the image to be retrieved searching for an image similar to the image to be retrieved in the preset image library based on the information entropy coding of the image to be retrieved, which can effectively improve the efficiency of image retrieval.
  • the image is indexed and retrieved, and the data amount of the comparison data is greatly reduced compared with directly comparing the binarization feature or the image feature, thereby effectively improving the image.
  • the efficiency of image retrieval is achieved.
  • an image retrieval method disclosed in another embodiment of the present application includes: Step 210 to Step 240.
  • Step 210 Determine a binarization feature of the image to be retrieved based on a first image feature of the image to be retrieved.
  • the first image feature is a convolutional neural network feature.
  • the convolutional neural network model extracts the first image feature of the image to be retrieved, that is, the CNN feature can better express the higher-level semantic features of the image, and has been well applied in the field of image classification, recognition and detection.
  • an embodiment of the image retrieval method is specifically described by taking the convolutional neural network feature of the image to be retrieved by the convolutional neural network model as an example.
  • ImageNet1000 is a computer vision system based on deep convolutional neural network, which trains a convolutional neural network model on 1000 class image classification problems.
  • InceptionV3 uses a pre-trained model on ImageNet1000, which is widely used in image processing.
  • the convolutional neural network feature of the image to be retrieved is extracted by InceptionV3 as the first image feature of the image to be retrieved.
  • the image to be retrieved is input into the InceptionV3 model, and the output parameter of the layer of the model "pool_8x8_s1" is obtained as a feature expression of the image to be retrieved.
  • the feature is a feature vector of 2048-dimensional (float type), which can be expressed as:
  • X [x 1 , x 2 , ..., x i , ..., x N ], i ⁇ [1, N].
  • the pool_8x8_s1 layer is the network layer closest to the loss layer in the InceptionV3 network structure. It is the most representative of the semantic features of the image and is more common in the field of image research.
  • a binarization feature of the image to be retrieved is determined.
  • this step will obtain an N-dimensional binarization feature, which can be expressed as:
  • Step 220 Determine information entropy coding of the image to be retrieved based on the binarization feature of the image to be retrieved.
  • Information entropy is often used to measure the distribution of information. Determining, according to the binarization feature of the image to be retrieved, information entropy coding of the image to be retrieved, comprising: determining a probability distribution of each feature value in the binarized feature of the image to be retrieved; determining based on the probability distribution Entropy of the binarization feature; discretizing the information entropy within a preset value range to obtain information entropy coding of the image to be retrieved.
  • a probability distribution of each feature value in the binarized feature of the image to be retrieved is determined, such as determining a probability distribution of 0 and 1 in the binarized feature.
  • an information entropy of the binarized feature is determined based on the probability distribution.
  • the information entropy of the binarization feature can be calculated by the following formula:
  • Entropy feature -p 0 log(p 0 )-p 1 log(p 1 ),Entropy feature ⁇ [0,1]
  • the distribution probability of 0 and 1 in the binarization feature X' can be calculated by the following formula:
  • N is the feature dimension of the binarized feature X'.
  • the information entropy is further discretized in a preset value range to determine information entropy coding of the image to be retrieved.
  • the information can be discretized by the following formula:
  • K is a preset value range, which can be determined according to the value range of the information entropy coding. For example, since the information entropy ranges from 0 to 1, if the information entropy coding is expected to range from 0 to 100, K is 100.
  • Information entropy is usually used to measure the distribution of information, while binarization features can express image features. Therefore, information entropy coding of binarized features can be used as a compressed expression of image features. For example, after information entropy coding of the binarization feature, the information entropy corresponding to the image to be retrieved can be obtained as 60. After determining the information entropy encoding of the image to be retrieved, an image similar to the image to be retrieved may be further retrieved in the preset image library based on the information entropy encoding.
  • the searching for an image similar to the image to be retrieved in the preset image library may include: first, determining a candidate image set in the preset image library based on the information entropy encoding; Then, by comparing the similarity between the image to be retrieved and each image in the candidate image set, one or more images matching the similarity of the image to be retrieved are determined.
  • Step 230 Determine a candidate image set in a preset image library based on the information entropy coding.
  • the image in the preset image library is indexed by information entropy coding.
  • the preset image library may be encoded as an index of the image with information entropy of the image.
  • the data format in the preset image library may be a key-value pair in the form of (information entropy coding, image).
  • the information entropy is encoded as an index of an image.
  • the method for acquiring the information entropy coding of the image in the preset image library is similar to the method for acquiring the information entropy encoding of the image to be retrieved, and details are not described herein again.
  • Determining a candidate image set in the preset image library based on information entropy encoding of the image to be retrieved comprising: in the preset image library, between information entropy coding and information entropy coding of the image to be retrieved An image whose difference value is smaller than a preset threshold is determined as a candidate matching image, and a plurality of candidate matching images are composed into a candidate image set.
  • the preset threshold may be, for example, 10, for the image to be retrieved with the information entropy code of 60, and the image with the information entropy coded between [50, 70] in the preset image library is selected as the matching candidate set of the image to be retrieved.
  • the retrieval amount can be reduced by 80%, which greatly improves the retrieval efficiency.
  • the preset threshold for the difference between the information entropy coding of the image in the image library and the information entropy coding of the image to be retrieved may be determined according to the experimental effect in a specific application scenario.
  • the difference between the information entropy coding of the image to be retrieved and the information entropy coding of all the images in the preset image library may be first calculated, and the information entropy code corresponding to the preset threshold (eg, 10) may be correspondingly encoded.
  • the image serves as a candidate matching image, and then all of the candidate matching images constitute a candidate image set.
  • the information entropy coding of the image can effectively express the feature distribution of the image.
  • Each image in the preset image library corresponds to one information entropy coding, and the image to be retrieved also corresponds to an information entropy coding.
  • the higher the similarity between the two images the closer their information entropy coding is. Therefore, the image entropy coding between the images within a certain range will have higher similarity.
  • Step 240 Perform similarity comparison between the image to be retrieved and the image in the candidate image set to determine an image that matches the image to be retrieved.
  • the image in the candidate image set is an image that is initially determined to have a high degree of similarity with the image to be retrieved.
  • each image in the candidate image set and the image to be retrieved may be further performed. Similarity comparison.
  • comparing the similarity between the image to be retrieved and the image in the candidate image set to determine an image that matches the image to be retrieved including: determining the image to be retrieved, and the candidate image set a second image feature of each image; calculating, based on the second image feature, a similarity score between the image to be retrieved and each image in the candidate image set; according to the similarity score from high to low An order of determining an image of the candidate image set that matches the image to be retrieved. First, the second image feature of the image to be retrieved and the second image feature of each image in the candidate image set are respectively acquired.
  • the second image feature may be a feature of the same type as the first image feature, such as the first image feature and the second image feature are CNN features; the second image feature may also be A feature of a different feature of the image feature, such as the first image feature being a CNN feature and the second image feature being a conventional image feature, such as a Gabor feature. Then, the Euclidean distance between the second image feature of the image to be retrieved and the second image feature of each image in the candidate image set is calculated to determine a similarity score for the two images.
  • the method for determining the similarity between the two images is not limited to the calculation of the Euclidean distance, and the similarity between the two images may be calculated by any method known to those skilled in the art, which is not limited in this application.
  • the images in the candidate image set are sorted according to the order of the similarity scores from high to low, and an image matching the image to be retrieved is determined.
  • image matching of the image to be retrieved in the candidate image set based on the second image feature may determine an image with the highest similarity as a final search result.
  • the image to be retrieved may be similarly matched with each image in the candidate image set, and the similarity between the image to be retrieved and each image in the candidate image set may be determined.
  • the scores are then sorted according to the similarity from high to low, and the sorted images in the candidate image set are fed back.
  • the second image feature may be a combined feature comprising at least two types of image features.
  • the first image feature may be included in the second image feature.
  • the second image feature can be a combined feature of CNN features and Gabor features.
  • the image retrieval method disclosed in the embodiment of the present application determines the binarization feature of the image to be retrieved based on the first image feature of the image to be retrieved; and determines the binary image based on the binarization feature of the image to be retrieved.
  • Entropy coding of the image to be retrieved determining the candidate image set in the preset image library based on the information entropy coding of the image to be retrieved, and finally performing feature matching of the image in the candidate image set, thereby effectively improving image retrieval s efficiency.
  • the image is indexed and retrieved, and the processing amount of the comparison data is greatly reduced compared with directly comparing the binarization feature or the image feature, thereby effectively improving the image.
  • the efficiency of image retrieval is greatly reduced.
  • a plurality of candidate matching images are initially determined according to the information entropy coding, and then the image to be retrieved and the candidate matching images are feature-matched one by one.
  • the information entropy coding is effective to reduce the image range of the feature matching, and the matching operation amount is reduced, thereby effectively improving the image retrieval efficiency.
  • the present application compares the original features of the entire image, the features are more comprehensive, and the retrieval effect is more accurate.
  • the convolutional neural network feature can better express the higher level semantic features of the image, and the image similarity matching based on the convolutional neural network feature can effectively ensure the accuracy of image matching.
  • An image retrieval device disclosed in this embodiment, as shown in FIG. 3, the device includes:
  • the binarization feature acquisition module 310 is configured to determine a binarization feature of the image to be retrieved based on the first image feature of the image to be retrieved;
  • the information entropy coding determining module 320 is configured to determine information entropy coding of the image to be retrieved based on the binarization feature of the image to be retrieved;
  • An image retrieval module 330 configured to retrieve an image similar to the image to be retrieved in a preset image library based on information entropy coding of the image to be retrieved; wherein the image in the preset image library is encoded by information entropy As an index.
  • the preset image library encodes the information entropy of the binarized feature of the image as an index of the image.
  • the image retrieval module 330 includes:
  • a candidate image set determining unit 3301 configured to determine a candidate image set in the preset image library based on the information entropy encoding determined by the information entropy encoding determining module 320;
  • the image matching unit 3302 is configured to compare the similarity between the image to be retrieved and the image in the candidate image set, and determine an image that matches the image to be retrieved.
  • the image matching unit 3302 is further configured to: determine the image to be retrieved, the second image feature of each image in the candidate image set; and calculate the to-be-study separately based on the second image feature Retrieving a similarity score between the image and each of the images in the candidate image set; determining an image of the candidate image set that matches the image to be retrieved in an order of the similarity score from high to low.
  • the second image feature is a combined feature comprising at least two types of image features.
  • the first image feature is included in the second image feature.
  • the candidate image set determining unit 3301 is further configured to: in the preset image library, an image in which an information entropy encoding and an information entropy encoding of the image to be retrieved are smaller than a preset threshold A candidate matching image is determined, and the candidate matching image is composed into a candidate image set.
  • the first image feature is a convolutional neural network feature.
  • the information entropy coding determining module 320 is further configured to: determine a probability distribution of each feature value in the binarized feature of the image to be retrieved; and determine a binary value of the image to be retrieved based on the probability distribution Information entropy of the feature; discretizing the information entropy within a preset value range to obtain information entropy coding of the image to be retrieved.
  • the image retrieval device disclosed in the embodiment of the present application determines the binarization feature of the image to be retrieved based on the first image feature of the image to be retrieved; and determines the binary image based on the binarization feature of the image to be retrieved.
  • Information entropy coding of the image to be retrieved searching for an image similar to the image to be retrieved in the preset image library based on the information entropy coding of the image to be retrieved, which can effectively improve the efficiency of image retrieval.
  • the image is indexed and retrieved, and the processing amount of the comparison data is greatly reduced compared with directly comparing the binarization feature or the image feature, thereby effectively improving the image.
  • the efficiency of image retrieval is greatly reduced.
  • a plurality of candidate matching images are initially determined according to the information entropy coding, and then the image to be retrieved and the candidate matching images are feature-matched one by one.
  • the information entropy coding is effective to reduce the image range of the feature matching, and the matching operation amount is reduced, thereby effectively improving the image retrieval efficiency.
  • the present application compares the original features of the entire image, the features are more comprehensive, and the retrieval effect is more accurate.
  • the convolutional neural network feature can better express the higher level semantic features of the image, and the image similarity matching based on the convolutional neural network feature can effectively ensure the accuracy of image matching.
  • the present application also discloses an electronic device including a memory, a processor, and a computer program stored on the memory and operable on the processor, the processor executing the computer program to implement the present application
  • the image retrieval method described in the first embodiment and the second embodiment can be a PC, a mobile terminal, a personal digital assistant, a tablet, or the like.
  • the present application also discloses a computer readable storage medium having stored thereon a computer program, the program being executed by the processor to implement the steps of the image retrieval method according to the first embodiment and the second embodiment of the present application.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Library & Information Science (AREA)
  • Databases & Information Systems (AREA)
  • Software Systems (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Mathematical Physics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Image Analysis (AREA)
  • Multimedia (AREA)

Abstract

本申请提供了一种图像检索方法。根据所述方法的一个例子，可基于待检索图像的第一图像特征确定所述待检索图像的二值化特征，并进一步基于所述待检索图像的二值化特征确定所述待检索图像的信息熵编码。然后，基于所述待检索图像的信息熵编码，可在预设图像库中检索与所述待检索图像相似的图像。

Description

图像检索
相关申请的交叉引用
本专利申请要求于2017年11月6日提交的、申请号为201711078740.7、发明名称为“一种图像检索方法及装置，电子设备”的中国专利申请的优先权，该申请的全文以引用的方式并入本文中。
技术领域
本申请涉及计算机技术领域，特别是涉及一种图像检索的方法、装置及电子设备。
背景技术
图像检索是基于待检索图像的图像特征在预设图像数据库中寻找最相似的图像的过程，在相关技术中有着广泛应用。常用的图像检索方法是将待搜索图像的图像特征与图像数据库中存储图像的图像特征进行一一比对，从而确定相似度最高的图像。然而，随着互联网技术的发展以及存储技术的发展，图像数据库的数据量越来越大，在海量的图像数据库中进行图像搜索时，这种图像检索方法的检索效率极低。
发明内容
本申请提供一种图像检索方法，以提高图像检索的效率。
为了解决上述问题，第一方面，本申请实施例提供了一种图像检索方法包括：基于待检索图像的第一图像特征，确定所述待检索图像的二值化特征；基于所述待检索图像的二值化特征，确定所述待检索图像的信息熵编码；基于所述待检索图像的信息熵编码，在预设图像库中检索与所述待检索图像相似的图像。其中，所述预设图像库中存储的图像以信息熵编码作为索引。
第二方面，本申请实施例提供了一种图像检索装置，包括：二值化特征获取模块，用于基于待检索图像的第一图像特征确定所述待检索图像的二值化特征；信息熵编码确定模块，用于基于所述待检索图像的二值化特征确定所述待检索图像的信息熵编码；图像检索模块，用于基于所述待检索图像的信息熵编码，在预设图像库中检索与所述待检索图像相似的图像。其中，所述预设图像库中的图像以信息熵编码作为索引。
第三方面，本申请实施例还公开了一种电子设备，包括存储器、处理器及存储在所述存储器上并可在所述处理器上运行的计算机程序，所述处理器执行所述计算机程序时实现本申请实施例所述的图像检索方法。
第四方面，本申请实施例提供了一种计算机可读存储介质，其上存储有计算机程序，该程序被处理器执行时实现本申请实施例公开的图像检索方法的步骤。
本申请实施例公开的图像检索方法，通过基于待检索图像的第一图像特征，确定所述待检索图像的二值化特征；基于所述待检索图像的二值化特征，确定所述待检索图像的信息熵编码；基于所述待检索图像的信息熵编码，在预设图像库中检索与所述待检索图像相似的图像，可有效提高图像检索的效率。通过以基于图像的二值化特征的信息熵作为图像编码，对图像进行索引和检索，相对于直接比较二值化特征或者图像特征而言，大量降低了比对数据的数据量，有效提升了图像检索的效率。
附图说明
为了更清楚地说明本申请实施例的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动性的前提下，还可以根据这些附图获得其他的附图。
图1是本申请实施例一图像检索方法流程图；
图2是本申请实施例二图像检索方法流程图；
图3是本申请实施例三图像检索装置结构示意图之一；
图4是本申请实施例三图像检索装置结构示意图之二。
具体实施方式
下面将结合本申请实施例中的附图，对本申请实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例是本申请一部分实施例，而不是全部的实施例。基于本申请中的实施例，本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其他实施例，都属于本申请保护的范围。
本实施例公开的一种图像检索方法，如图1所示，该方法包括：步骤110至步骤130。
步骤110，基于待检索图像的第一图像特征，确定所述待检索图像的二值化特征。
首先，获取待检索图像的第一图像特征。
常用的图像特征有纹理特征、颜色特征、形状特征、空间关系特征等，不同的特征适用于不同的图像内容以及不同的图像识别需求。常用的特征提取方法有Fourier变换法、Gabor变换法、小波变换法、神经网络模型提取法等。具体实施时，获取待检索图像的图像特征的具体方法可以采用本领域技术人员熟知的任意方法，如可通过卷积神经网络(CNN)模型提取待检索图像的图像特征。本申请对获取待检索图像的图像特征的具体实施方法不做限定，优选的，采用卷积神经网络模型提取待检索图像的图像特征。
卷积神经网络模型提取待检索图像的图像特征、即卷积神经网络特征(以下也可简称为CNN特征)能够较好地表达图像更高层次的语义特征，在图像分类、识别及检测领域已经得到很好的应用。本申请实施例中，以第一图像特征为CNN特征为例，详细说明图像检索技术方案。
然后，基于所述待检索图像的第一图像特征，确定所述待检索图像的二值化特征。
图像特征通常是一个多维的向量。具体实施时，可以通过本领域技术人员熟知的二值化方案对获取的待检索图像的第一图像特征进行二值化处理，以提取所述待检索图像的二值化特征。例如，遍历CNN特征每个维度的特征值Xi，根据以下规则对待检索图像的卷积神经网络特征进行二值化：
If Xi>TH:B(Xi)＝1；
else:B(Xi)＝0；
其中，TH为根据经验设定的阈值，该阈值的设置希望使得二值化后特征中的0和1的分布尽可能不均匀，以增加图像的区分度，例如可以取值为0.5。
步骤120，基于所述待检索图像的二值化特征，确定所述待检索图像的信息熵编码。
信息熵通常用于衡量信息的分布。本申请具体实施时，确定所述二值化特征之后，可以将所述二值化特征的信息熵作为所述二值化特征所属待检索图像的信息熵编码。然后，通过根据所述信息熵编码确定相似图像。
步骤130，基于所述待检索图像的信息熵编码，在预设图像库中检索与所述待检索图像相似的图像。
具体实施时，所述预设图像库可以以图像的信息熵编码作为图像的索引，然后，通 过基于信息熵编码进行比较，可以确定相似图像。具体实施时，由于信息熵编码在一定程度上反映了图像的二值化特征，因此，信息熵编码相近，对应的图像相似度也较高，通过比较信息熵编码可以初步确定相似图像。进一步的，可以通过基于图像特征对相似图像进行相似度判断，从而相对更准确地检索到预设图像库中与待检索图像匹配的图像。
本申请实施例公开的图像检索方法，通过基于所述待检索图像的第一图像特征，确定所述待检索图像的二值化特征；基于所述待检索图像的二值化特征，确定所述待检索图像的信息熵编码；基于所述待检索图像的信息熵编码，在预设图像库中检索与所述待检索图像相似的图像，可有效提高图像检索的效率。通过以基于图像的二值化特征的信息熵作为图像编码，对图像进行索引和检索，相对于直接比较二值化特征或者图像特征而言，大量降低了比对数据的数据量，有效提升了图像检索的效率。
如图2所示，本申请另一实施例公开的一种图像检索方法，包括：步骤210至步骤240。
步骤210，基于待检索图像的第一图像特征，确定所述待检索图像的二值化特征。
首先，获取待检索图像的第一图像特征。
优选的，所述第一图像特征为卷积神经网络特征。卷积神经网络模型提取待检索图像的第一图像特征、即CNN特征能够较好地表达图像更高层次的语义特征，在图像分类、识别及检测领域已经得到很好的应用。本实施例中，以通过卷积神经网络模型提取待检索图像的卷积神经网络特征为例，具体描述图像检索方法的实施方案。
ImageNet1000是基于深度卷积神经网络的计算机视觉***，其在1000类图像分类问题上训练了一个卷积神经网络模型。而InceptionV3使用的是在ImageNet1000上预训练过的模型，在图像处理领域应用比较广泛。本申请具体实施时，通过InceptionV3提取待检索图像的卷积神经网络特征作为待检索图像的第一图像特征。具体实施时，把待检索图像输入InceptionV3模型，获取模型“pool_8x8_s1”层的输出参数作为所述待检索图像的特征表达，该特征为2048维(float类型)的特征向量，可以表示为:
X＝[x 1,x 2,…,x i,…,x N],i∈[1,N]。
其中，N表示特征的维度，例如N＝2048。pool_8x8_s1层是InceptionV3网络结构中最接近loss(损失)层的网络层，最能代表图像的语义特征，且在图像研究领域较为通用。
然后，基于所述待检索图像的第一图像特征，确定所述待检索图像的二值化特征。
基于所述待检索图像的第一图像特征，确定所述待检索图像的二值化特征的具体实施方法可参见实施例一，此处不再赘述。
通过对获取的待检索图像的第一图像特征X进行二值化处理，本步骤将得到一个N维的二值化特征，可以表示为：
X’＝[x′ 1,x′ 2,…,x′ i,…,x′ N],i∈[1,N],
其中，N表示特征的维度，例如N＝2048。
步骤220，基于所述待检索图像的二值化特征，确定所述待检索图像的信息熵编码。
信息熵通常用于衡量信息的分布。基于所述待检索图像的二值化特征，确定所述待检索图像的信息熵编码，包括：确定所述待检索图像的二值化特征中各特征值的概率分布；基于所述概率分布确定所述二值化特征的信息熵；将所述信息熵在预设取值范围内进行离散化处理，得到所述待检索图像的信息熵编码。
具体实施时，首先确定所述待检索图像的二值化特征中的各特征值的概率分布，如确定二值化特征中0和1的概率分布。然后，基于所述概率分布确定所述二值化特征的信息熵。具体实施时，可以通过如下公式计算二值化特征的信息熵：
Entropy feature＝-p 0log(p 0)-p 1log(p 1),Entropy feature∈[0,1]
其中，p 0为二值化特征中特征值0的分布概率，p 1为二值化特征中特征值1的分布概率。具体实施时，可以通过如下公式统计二值化特征X’中的0和1的分布概率：
Figure PCTCN2018110865-appb-000001
其中，N为二值化特征X’的特征维度。
确定所述待检索图像的二值化特征的信息熵之后，进一步对该信息熵在预设取值范围内进行离散化编码，以确定所述待检索图像的信息熵编码。具体实施时，可以通过以下公式对信息上进行离散化编码：
Q(Entropy feature)＝int(K*Entropy feature),Q(Entropy feature)∈[0,K]；
其中，K为预设取值范围，可根据信息熵编码的取值范围确定。例如，因为信息熵的取值范围为0到1，如果期望信息熵编码的取值范围为0到100之间，则K取值为100。
信息熵通常用于衡量信息的分布，而二值化特征可以表达图像特征，因此，二值化特征的信息熵编码可以作为图像特征的一种压缩的表达方式。例如，经过对二值化特征 进行信息熵编码，可以得到待检索图像对应的信息熵编码为60。在确定了待检索图像的信息熵编码之后，可以进一步基于所述信息熵编码，在预设图像库中检索与所述待检索图像相似的图像。具体实施时，基于所述信息熵编码，在预设图像库中检索与所述待检索图像相似的图像可包括：首先，基于所述信息熵编码，在预设图像库中确定候选图像集合；然后，通过将所述待检索图像与所述候选图像集合中各图像进行相似度比较，确定与所述待检索图像相似度匹配的一幅或多幅图像。
步骤230，基于所述信息熵编码，在预设图像库中确定候选图像集合。
其中，所述预设图像库中的图像以信息熵编码作为索引。
本申请具体实施时，首先需要预设图像库作为待检索图像检索的对象。具体实施时，预设图像库可以以图像的信息熵编码作为所述图像的索引。例如，所述预设图像库中的数据格式可为(信息熵编码，图像)形式的键值对。其中，所述信息熵编码作为图像的索引。预设图像库中图像的信息熵编码的获取方法和待检索图像的信息熵编码的获取方法相似，此处不再赘述。
基于所述待检索图像的信息熵编码，在预设图像库中确定候选图像集合，包括：在所述预设图像库中，将信息熵编码与所述待检索图像的信息熵编码之间的差值小于预设阈值的图像确定为候选匹配图像，并将多个候选匹配图像组成候选图像集合。具体实施时，预设阈值可以取例如10，针对信息熵编码为60的待检索图像，选取预设图像库中信息熵编码在[50,70]之间的图像作为待检索图像的匹配候选集。这样，相比图像库中信息熵编码为[0,100]而言，检索量可减少了80％，极大提升了检索效率。需要说明的是，针对图像库中图像的信息熵编码与待检索图像的信息熵编码之间的差值的预设阈值，在具体的应用场景中可依据实验效果确定。该阈值越小，索引的信息熵编码范围越小，对应的候选图像集合越小，检索的效率越高，但是精确度可能较低；相反，该阈值越大，索引的信息熵编码范围越大，对应的候选图像集合越大，检索的效率越低，但是精确度可能较高。
具体实施时，可以首先计算待检索图像的信息熵编码与预设图像库中所有图像的信息熵编码之间的差值，并将差值小于预设阈值(如10)的信息熵编码对应的图像作为候选匹配图像，然后由所有候选匹配图像构成候选图像集合。图像的信息熵编码可以有效表达图像的特征分布，预设图像库中每幅图像均对应一个信息熵编码，待检索图像而言也对应一个信息熵编码。并且，两个图像之间的相似度越高，它们的信息熵编码越接近。因此，相互之间的信息熵编码差值在一定范围内的图像，相似性会较高。通过设置对一 定范围内的信息熵编码进行匹配，可大幅减小待检索图像的范围。
步骤240，将所述待检索图像与所述候选图像集合中的图像进行相似度比较，确定与所述待检索图像匹配的图像。
候选图像集合中的图像是初步判断与待检索图像相似度较高的图像，为了提高检索结果的准确性，本申请具体实施时，可进一步对所述候选图像集合中各图像与待检索图像进行相似度比较。
具体实施时，将所述待检索图像与所述候选图像集合中的图像进行相似度比较，确定与所述待检索图像匹配的图像，包括：确定所述待检索图像、所述候选图像集合中的各图像的第二图像特征；基于所述第二图像特征，分别计算所述待检索图像和所述候选图像集合中各图像之间的相似度得分；按照所述相似度得分由高到低的顺序，确定所述候选图像集中与所述待检索图像匹配的图像。首先，分别获取待检索图像的第二图像特征和候选图像集合中各图像的第二图像特征。其中，所述第二图像特征可以为与所述第一图像特征相同类别的特征，如第一图像特征和第二图像特征均为CNN特征；所述第二图像特征也可以为与所述第一图像特征不同类别的特征，如第一图像特征为CNN特征，第二图像特征为传统图像特征，如Gabor特征。然后，计算待检索图像的第二图像特征与候选图像集合中各图像的第二图像特征之间的欧式距离，以确定两幅图像的相似度得分。具体实施时，确定两幅图像之间的相似度的方法不限于计算欧式距离，还可以采用本领域技术人员熟知的任意方法计算两幅图像之间的相似度，本申请对此不作限定。
最后，按照所述相似度得分由高到低的顺序，对所述候选图像集合中的图像进行排序，确定与所述待检索图像匹配的图像。具体实施时，基于所述第二图像特征在所述候选图像集合中对所述待检索图像进行图像匹配，可以确定相似度最高的图像作为最终的检索结果。在某些具体应用中，还可以将所述待检索图像与所述候选图像集合中各图像进行相似度匹配，确定所述待检索图像与所述候选图像集合中每幅图像之间的相似度得分，然后，按照相似度由高到低进行排序，将排序后的所述候选图像集合中的图像进行反馈。
优选的，所述第二图像特征可为包括至少两类图像特征的组合特征。例如，所述第二图像特征中可包括所述第一图像特征。为了提高图像检索的准确度，在进行相似度比较时，可以提取图像的更细腻、更丰富的特征。例如，所述第二图像特征可为CNN特征和Gabor特征的组合特征。
本申请实施例公开的图像检索方法，通过基于所述待检索图像的第一图像特征，确定所述待检索图像的二值化特征；基于所述待检索图像的二值化特征，确定所述待检索图像的信息熵编码；基于所述待检索图像的信息熵编码，在预设图像库中确定候选图像集合，最后在所述候选图像集合中进一步进行图像的特征匹配，可有效提高图像检索的效率。通过以基于图像的二值化特征的信息熵作为图像编码，对图像进行索引和检索，相对于直接比较二值化特征或者图像特征而言，大量降低了比对数据的处理量，有效提升了图像检索的效率。
基于相似的图像对应相似的信息熵编码这一特性，首先根据信息熵编码初步确定多个候选匹配图像，然后将待检索图像与候选匹配图像一一进行特征匹配。这样，进行信息熵编码比较有效缩小了特征匹配的图像范围，减小了匹配运算量，从而有效提升了图像检索效率。并且，相对于直接比较图像二值化特征的方法，本申请比较的是整幅图像的原始特征，特征更全面，检索效果更准确。进一步的，卷积神经网络特征能够较好地表达图像更高层次的语义特征，基于卷积神经网络特征进行图像相似度匹配，可以有效保证图像匹配的准确性。
本实施例公开的一种图像检索装置，如图3所示，所述装置包括：
二值化特征获取模块310，用于基于待检索图像的第一图像特征，确定所述待检索图像的二值化特征；
信息熵编码确定模块320，用于基于所述待检索图像的二值化特征，确定所述待检索图像的信息熵编码；
图像检索模块330，用于基于所述待检索图像的信息熵编码，在预设图像库中检索与所述待检索图像相似的图像；其中，所述预设图像库中的图像以信息熵编码作为索引。
具体实施时，预设图像库以图像的二值化特征的信息熵编码作为图像的索引。
可选的，如图4所示，所述图像检索模块330包括：
候选图像集合确定单元3301，用于基于所述信息熵编码确定模块320确定的信息熵编码，在预设图像库中确定候选图像集合；
图像匹配单元3302，用于将所述待检索图像与所述候选图像集合中的图像进行相似度比较，确定与所述待检索图像匹配的图像。
可选的，所述图像匹配单元3302，进一步用于：确定所述待检索图像、所述候选图 像集合中的各图像的第二图像特征；基于所述第二图像特征，分别计算所述待检索图像和所述候选图像集合中各图像之间的相似度得分；按照所述相似度得分由高到低的顺序，确定所述候选图像集合中与所述待检索图像匹配的图像。
优选的，所述第二图像特征为包括至少两类图像特征的组合特征。例如，所述第二图像特征中包括所述第一图像特征。
可选的，所述候选图像集合确定单元3301进一步用于：在所述预设图像库中，将信息熵编码与所述待检索图像的信息熵编码之间的差值小于预设阈值的图像确定为候选匹配图像，并将所述候选匹配图像组成候选图像集合。
可选的，所述第一图像特征为卷积神经网络特征。
可选的，所述信息熵编码确定模块320，进一步用于：确定所述待检索图像的二值化特征中各特征值的概率分布；基于所述概率分布确定所述待检索图像的二值化特征的信息熵；将所述信息熵在预设取值范围内进行离散化处理，得到所述待检索图像的信息熵编码。
本申请实施例公开的图像检索装置，通过基于所述待检索图像的第一图像特征，确定所述待检索图像的二值化特征；基于所述待检索图像的二值化特征，确定所述待检索图像的信息熵编码；基于所述待检索图像的信息熵编码，在预设图像库中检索与所述待检索图像相似的图像，可有效提高图像检索的效率。通过以基于图像的二值化特征的信息熵作为图像编码，对图像进行索引和检索，相对于直接比较二值化特征或者图像特征而言，大量降低了比对数据的处理量，有效提升了图像检索的效率。
基于相似的图像对应相似的信息熵编码这一特性，首先根据信息熵编码初步确定多个候选匹配图像，然后将待检索图像与候选匹配图像一一进行特征匹配。这样，进行信息熵编码比较有效缩小了特征匹配的图像范围，减小了匹配运算量，从而有效提升了图像检索效率。并且，相对于直接比较图像二值化特征的方法，本申请比较的是整幅图像的原始特征，特征更全面，检索效果更准确。进一步的，卷积神经网络特征能够较好地表达图像更高层次的语义特征，基于卷积神经网络特征进行图像相似度匹配，可以有效保证图像匹配的准确性。
相应的，本申请还公开了一种电子设备，包括存储器、处理器及存储在所述存储器上并可在处理器上运行的计算机程序，所述处理器执行所述计算机程序时实现如本申请实施例一和实施例二所述的图像检索方法。所述电子设备可以为PC机、移动终端、个 人数字助理、平板电脑等。
本申请还公开了一种计算机可读存储介质，其上存储有计算机程序，该程序被处理器执行时实现如本申请实施例一和实施例二所述的图像检索方法的步骤。
本说明书中的各个实施例均采用递进的方式描述，每个实施例重点说明的都是与其他实施例的不同之处，各个实施例之间相同相似的部分互相参见即可。对于装置实施例而言，由于其与方法实施例基本相似，所以描述的比较简单，相关之处参见方法实施例的部分说明即可。
以上对本申请提供的一种图像检索方法及装置进行了详细介绍，本文中应用了具体个例对本申请的原理及实施方式进行了阐述，以上实施例的说明只是用于帮助理解本申请的方法及其核心思想；同时，对于本领域的一般技术人员，依据本申请的思想，在具体实施方式及应用范围上均会有改变之处，综上所述，本说明书内容不应理解为对本申请的限制。
通过以上的实施方式的描述，本领域的技术人员可以清楚地了解到各实施方式可借助软件加必需的通用硬件平台的方式来实现，当然也可以通过硬件实现。基于这样的理解，上述技术方案本质上或者说对现有技术做出贡献的部分可以以软件产品的形式体现出来，该计算机软件产品可以存储在计算机可读存储介质中，如ROM/RAM、磁碟、光盘等，包括若干指令用以使得一台计算机设备(可以是个人计算机，服务器，或者网络设备等)执行各个实施例或者实施例的某些部分所述的方法。

Claims (10)

  1. 一种图像检索方法，包括：
    基于待检索图像的第一图像特征，确定所述待检索图像的二值化特征；
    基于所述待检索图像的所述二值化特征，确定所述待检索图像的信息熵编码；
    基于所述待检索图像的所述信息熵编码，在预设图像库中检索与所述待检索图像相似的图像；其中，所述预设图像库中的图像以所述信息熵编码作为索引。
  2. 根据权利要求1所述的方法，其特征在于，基于所述待检索图像的所述信息熵编码，在所述预设图像库中检索与所述待检索图像相似的图像的步骤，包括：
    基于所述待检索图像的所述信息熵编码，在所述预设图像库中确定候选图像集合；
    将所述待检索图像与所述候选图像集合中的各图像进行相似度比较，确定与所述待检索图像匹配的图像。
  3. 根据权利要求2所述的方法，其特征在于，将所述待检索图像与所述候选图像集合中的各图像进行相似度比较，确定与所述待检索图像匹配的图像的步骤，包括：
    确定所述待检索图像以及所述候选图像集合中的各图像的第二图像特征；
    基于所述第二图像特征，分别计算所述待检索图像和所述候选图像集合中各图像之间的相似度得分；
    按照所述相似度得分由高到低的顺序，确定所述候选图像集合中与所述待检索图像匹配的图像。
  4. 根据权利要求3所述的方法，其特征在于，
    所述第二图像特征为包括至少两类图像特征的组合特征，
    所述第二图像特征中包括所述第一图像特征。
  5. 根据权利要求2所述的方法，其特征在于，基于所述待检索图像的所述信息熵编码，在所述预设图像库中确定所述候选图像集合的步骤，包括：
    在所述预设图像库中，将所述信息熵编码与所述待检索图像的所述信息熵编码之间的差值小于预设阈值的图像确定为候选匹配图像，
    将所述候选匹配图像组成所述候选图像集合。
  6. 根据权利要求1至5任一项所述的方法，其特征在于，所述第一图像特征为卷 积神经网络特征。
  7. 根据权利要求1至5任一项所述的方法，其特征在于，基于所述待检索图像的所述二值化特征，确定所述待检索图像的所述信息熵编码的步骤，包括：
    确定所述待检索图像的所述二值化特征中各特征值的概率分布；
    基于所述概率分布确定所述二值化特征的信息熵；
    将所述信息熵在预设取值范围内进行离散化处理，得到所述待检索图像的信息熵编码。
  8. 一种图像检索装置，包括：
    二值化特征获取模块，用于基于待检索图像的第一图像特征，确定所述待检索图像的二值化特征；
    信息熵编码确定模块，用于基于所述待检索图像的所述二值化特征，确定所述待检索图像的信息熵编码；
    图像检索模块，用于基于所述待检索图像的所述信息熵编码，在预设图像库中检索与所述待检索图像相似的图像；其中，所述预设图像库中的图像以所述信息熵编码作为索引。
  9. 一种电子设备，包括存储器、处理器及存储在所述存储器上并可在所述处理器上运行的计算机程序，其特征在于，所述处理器执行所述计算机程序时实现权利要求1至7任意一项所述的图像检索方法。
  10. 一种计算机可读存储介质，其上存储有计算机程序，其特征在于，该程序被处理器执行时实现权利要求1至7任意一项所述的图像检索方法的步骤。
PCT/CN2018/110865 2017-11-06 2018-10-18 图像检索 WO2019085765A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
US16/762,139 US11281714B2 (en) 2017-11-06 2018-10-18 Image retrieval
EP18874554.1A EP3703061A4 (en) 2017-11-06 2018-10-18 IMAGE RECOVERY

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201711078740.7A CN108090117B (zh) 2017-11-06 2017-11-06 一种图像检索方法及装置，电子设备
CN201711078740.7 2017-11-06

Publications (1)

Publication Number Publication Date
WO2019085765A1 true WO2019085765A1 (zh) 2019-05-09

Family

ID=62170467

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/110865 WO2019085765A1 (zh) 2017-11-06 2018-10-18 图像检索

Country Status (4)

Country Link
US (1) US11281714B2 (zh)
EP (1) EP3703061A4 (zh)
CN (1) CN108090117B (zh)
WO (1) WO2019085765A1 (zh)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108090117B (zh) 2017-11-06 2019-03-19 北京三快在线科技有限公司 一种图像检索方法及装置，电子设备
CN111324819B (zh) * 2020-03-24 2021-07-30 北京字节跳动网络技术有限公司 一种媒体内容搜索的方法、装置、计算机设备及存储介质
CN114077681B (zh) * 2022-01-19 2022-04-12 腾讯科技（深圳）有限公司 一种图像数据处理方法、装置、计算机设备及存储介质

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102419820A (zh) * 2011-08-18 2012-04-18 电子科技大学 一种快速检测视频与图像中汽车车标的方法
JP2015056077A (ja) * 2013-09-12 2015-03-23 Kddi株式会社 画像に基づくバイナリ特徴ベクトルを用いた画像検索装置、システム、プログラム及び方法
CN108090117A (zh) * 2017-11-06 2018-05-29 北京三快在线科技有限公司 一种图像检索方法及装置，电子设备

Family Cites Families (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR100448111B1 (ko) * 1998-06-09 2004-09-10 마츠시타 덴끼 산교 가부시키가이샤 화상 부호화 장치, 화상 복호화 장치 및 데이터 기억 매체
US6285995B1 (en) * 1998-06-22 2001-09-04 U.S. Philips Corporation Image retrieval system using a query image
CN101639983B (zh) * 2009-08-21 2011-02-02 任雪梅 一种基于图像信息熵的多车道车流量检测方法
US9940575B2 (en) 2015-06-04 2018-04-10 Yahoo Holdings, Inc. Image searching
CN106445939B (zh) 2015-08-06 2019-12-13 阿里巴巴集团控股有限公司 图像检索、获取图像信息及图像识别方法、装置及***
US10628736B2 (en) * 2015-09-24 2020-04-21 Huron Technologies International Inc. Systems and methods for barcode annotations for digital images
CN106354850A (zh) * 2016-08-31 2017-01-25 广州精点计算机科技有限公司 一种基于k‑近邻分类的图像识别方法
CN106874421A (zh) * 2017-01-24 2017-06-20 聊城大学 基于自适应矩形窗口的图像检索方法
CN106909924B (zh) * 2017-02-18 2020-08-28 北京工业大学 一种基于深度显著性的遥感影像快速检索方法
CN107122809B (zh) * 2017-04-24 2020-04-28 北京工业大学 基于图像自编码的神经网络特征学习方法

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102419820A (zh) * 2011-08-18 2012-04-18 电子科技大学 一种快速检测视频与图像中汽车车标的方法
JP2015056077A (ja) * 2013-09-12 2015-03-23 Kddi株式会社 画像に基づくバイナリ特徴ベクトルを用いた画像検索装置、システム、プログラム及び方法
CN108090117A (zh) * 2017-11-06 2018-05-29 北京三快在线科技有限公司 一种图像检索方法及装置，电子设备

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
See also references of EP3703061A4 *
WANG, SONG: "Entropy Based Image Retrieval", MASTER THESIS, 15 September 2003 (2003-09-15), pages 1 - 45, XP009520801, ISSN: 1671-6779 *

Also Published As

Publication number Publication date
CN108090117A (zh) 2018-05-29
US20200364259A1 (en) 2020-11-19
EP3703061A4 (en) 2020-12-30
EP3703061A1 (en) 2020-09-02
US11281714B2 (en) 2022-03-22
CN108090117B (zh) 2019-03-19

Similar Documents

Publication Publication Date Title
CN109783655B (zh) 一种跨模态检索方法、装置、计算机设备和存储介质
CN105912611B (zh) 一种基于cnn的快速图像检索方法
US10796196B2 (en) Large scale image recognition using global signatures and local feature information
WO2018001384A1 (zh) 数据处理、数据识别方法和装置、计算机设备
US8498455B2 (en) Scalable face image retrieval
WO2021056710A1 (zh) 多轮问答识别方法、装置、计算机设备及存储介质
WO2019085765A1 (zh) 图像检索
CN112434553B (zh) 一种基于深度字典学习的视频鉴别方法及***
US20200151435A1 (en) Automated and unsupervised curation of image datasets
US20220215205A1 (en) Robust content fingerprinting for image attribution
CN110858217A (zh) 微博敏感话题的检测方法、装置及可读存储介质
Qazanfari et al. Content-based image retrieval using HSV color space features
Sundara Vadivel et al. An efficient CBIR system based on color histogram, edge, and texture features
Garcia-Fidalgo et al. On the use of binary feature descriptors for loop closure detection
CN114461839A (zh) 基于多模态预训练的相似图片检索方法、装置及电子设备
CN109597906B (zh) 图像检索方法及装置
CN108694411B (zh) 一种识别相似图像的方法
Sun et al. Search by detection: Object-level feature for image retrieval
JP6017277B2 (ja) 特徴ベクトルの集合で表されるコンテンツ間の類似度を算出するプログラム、装置及び方法
CN108763265B (zh) 一种基于分块检索的图像识别方法
CN115909381A (zh) 一种文本图像识别方法、***及相关装置
Chen et al. Neighborhood-exact nearest neighbor search for face retrieval
US20210158082A1 (en) Duplicate image detection based on image content
Wu et al. Similar image retrieval in large-scale trademark databases based on regional and boundary fusion feature
Cai et al. Semantic edge detection based on deep metric learning

Legal Events

Date Code Title Description
NENP Non-entry into the national phase

Ref country code: DE

ENP Entry into the national phase

Ref document number: 2018874554

Country of ref document: EP

Effective date: 20200528

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

Ref document number: 18874554

Country of ref document: EP

Kind code of ref document: A1