KR101752742B1 - Method and apparatus for extracting key point based on simplifying generation of Gaussioan scale space - Google Patents
Method and apparatus for extracting key point based on simplifying generation of Gaussioan scale space Download PDFInfo
- Publication number
- KR101752742B1 KR101752742B1 KR1020150188231A KR20150188231A KR101752742B1 KR 101752742 B1 KR101752742 B1 KR 101752742B1 KR 1020150188231 A KR1020150188231 A KR 1020150188231A KR 20150188231 A KR20150188231 A KR 20150188231A KR 101752742 B1 KR101752742 B1 KR 101752742B1
- Authority
- KR
- South Korea
- Prior art keywords
- gaussian
- image
- octave
- feature point
- scale
- Prior art date
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
-
- G06K9/00973—
-
- G06K9/4671—
-
- G06K9/6211—
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
A feature point extraction method and apparatus based on simplification of Gaussian scale space generation are disclosed. A feature point extraction method based on simplification of Gaussian scale space generation is based on a step of determining a pre-octave based on a first Gaussian filter applying a first sigma value to an input image and a first Gaussian image constituting a dictionary octave And determining each of the N Gaussian scale spaces corresponding to each of octave 1 to octave N (where N is a natural number greater than 1) by applying a second Gaussian filter applying a second sigma value.
Description
The present invention relates to a feature point extraction method and apparatus, and more particularly, to a feature point extraction method and apparatus based on simplification of Gaussian scale space creation.
Recently, image processing technologies such as image recognition and object recognition have been widely applied in robots, security systems, mobile devices, etc. according to the development of computer vision technology.
Particularly, in the field of image processing, object recognition is mainly applied to special fields such as industry and military. However, since many products such as face recognition and panorama are applied to mobile devices and household appliances closely related to real life, It has been widely applied.
However, object recognition requires a lot of memory because algorithms are very complicated and require a lot of memory.
Object recognition is the process of receiving an image from a camera and finding out where the object is to be found and the location of the object. Since the types and characteristics of objects are so varied, many object recognition algorithms for each characteristic have been studied.
Among the object recognition algorithms, SIFT (scale invariant feature transform) and SURF (speed up robust feature) are typical algorithms for detecting feature points and using distribution based descriptors.
One aspect of the present invention provides a feature point extraction method based on simplification of Gaussian scale space generation.
Another aspect of the present invention provides a feature point extraction apparatus based on simplification of Gaussian scale space generation.
A feature point extraction method based on simplification of Gaussian scale space generation according to an aspect of the present invention includes a step of determining a pre-octave based on a first Gaussian filter applying a first sigma value to an input image, A second Gaussian filter applying a second sigma value based on the first Gaussian image of the first Gaussian image to determine N pieces of Gaussian scale spaces corresponding to
The first sigma value is twice the second sigma value, the pre-octave is a first Gaussian scale space including the first Gaussian image generated by downscaling the input image, The second Gaussian scale space including the image generated by upscaling until the scale of the first Gaussian image is doubled, and the octave N is a second Gaussian scale space including the N-1 Gaussian image included in the Gaussian scale space corresponding to the octave N-1. And may be a third Gaussian scale space including the image generated by upscaling until the scale of the image is doubled.
The step of determining the pre-octave by applying the first Gaussian filter to which the first sigma value is applied may further include filtering the input image based on the first Gaussian filter to which the first sigma value is applied, Generating a first Gaussian image by downscaling the filtered image to reduce a horizontal size of the input image by a factor of 1/2 and reducing a vertical size of the input image by a factor of 1/2; .
In addition, a feature point extraction method based on simplification of Gaussian scale space generation includes extracting feature point candidates from each of the N Gaussian scale spaces, removing feature point candidates that are not stable in matching based on the Taylor series, Determining a feature point descriptor for the feature point, and performing matching on the feature point based on the feature point descriptor to determine a final feature point for the input image.
According to another aspect of the present invention, a feature point extraction apparatus based on simplification of Gaussian scale space generation includes a processor, wherein the processor determines a pre-octave based on a first Gaussian filter applying a first sigma value to an input image And applying a second Gaussian filter applying a second sigma value based on the first Gaussian image constituting the pre-octave to obtain N Gaussian corresponding to each
The first sigma value is twice the second sigma value, the pre-octave is a first Gaussian scale space including the first Gaussian image generated by downscaling the input image, The second Gaussian scale space including the image generated by upscaling until the scale of the first Gaussian image is doubled, and the octave N is a second Gaussian scale space including the N-1 Gaussian image included in the Gaussian scale space corresponding to the octave N-1. And may be a third Gaussian scale space including the image generated by upscaling until the scale of the image is doubled.
The processor may filter the input image based on the first Gaussian filter to which the first sigma value is applied, downscale the filtered image to reduce the size of the input image by 1/2, The first Gaussian image may be generated by reducing the vertical size of the input image by ½ times.
The processor may extract feature point candidates from each of the N Gaussian scale spaces, determine feature points by removing unstable feature point candidates in matching based on the Taylor series among the feature point candidates, And determining a final feature point for the input image by performing matching on the feature point based on the feature point descriptor.
The feature point extraction method and apparatus based on the simplification of Gaussian scale space generation according to the embodiment of the present invention can reduce the computation amount and reduce the complexity of the SIFT algorithm based on the reduced image size and the initial sigma reduction. Therefore, the SIFT algorithm can be utilized in the fields such as mobile and drone which have not been utilized due to the computational complexity of the existing SIFT algorithm.
1 is a flowchart illustrating a SIFT algorithm according to an embodiment of the present invention.
2 is a flowchart showing a SIFT algorithm according to an embodiment of the present invention.
3 is a conceptual diagram illustrating a method of generating a Gaussian space according to an embodiment of the present invention.
4 is a conceptual diagram illustrating an apparatus for generating a Gaussian space according to an embodiment of the present invention.
The following detailed description of the invention refers to the accompanying drawings, which illustrate, by way of illustration, specific embodiments in which the invention may be practiced. These embodiments are described in sufficient detail to enable those skilled in the art to practice the invention. It should be understood that the various embodiments of the present invention are different, but need not be mutually exclusive. For example, certain features, structures, and characteristics described herein may be implemented in other embodiments without departing from the spirit and scope of the invention in connection with an embodiment. It is also to be understood that the position or arrangement of the individual components within each disclosed embodiment may be varied without departing from the spirit and scope of the invention. The following detailed description is, therefore, not to be taken in a limiting sense, and the scope of the present invention is to be limited only by the appended claims, along with the full scope of equivalents to which such claims are entitled, if properly explained. In the drawings, like reference numerals refer to the same or similar functions throughout the several views.
Hereinafter, preferred embodiments of the present invention will be described in more detail with reference to the drawings.
Feature point extraction extracts characteristic parts from input image coming from camera. Characteristic parts can be parts with complex shapes in the image, such as the edges of an object or the pattern existing on the surface of an object. These feature points can be used to generate the panoramic image by comparing the similarity with the feature points extracted from the next frame image of the camera.
In addition, the feature points can be utilized for object recognition by comparing similarities with feature points previously extracted from a previously stored image. In order for the extracted feature points to be utilized in various conditions and situations, the feature points should be extracted from the same position of the object even if the rotation, size, and brightness change of the image change. The feature points extracted by the scale-invariant feature transform (SIFT) algorithm are known to be the most robust to various image changes. Therefore, the SIFT algorithm can be most utilized for extracting feature points.
The SIFT algorithm can be divided into a part for finding the location of the minutiae in the image and a part for generating a descriptor for the surrounding area information of each minutiae. The SIFT algorithm can be robust against image size change by generating a Gaussian scale space from the input image and finding the location of the feature point. In addition, the SIFT algorithm can be robust to rotational changes by considering the directionality at each feature point position in the process of generating descriptors for feature points. Through this process, the SIFT algorithm is robust against various image changes, but the overall computational complexity is very high. Hereinafter, a method for reducing the complexity of the SIFT algorithm is disclosed in the embodiment of the present invention.
1 is a flowchart illustrating a SIFT algorithm according to an embodiment of the present invention.
Referring to FIG. 1, scale-space extreme detection is performed (step S100).
In the scale-space extreme-detection step, feature points that are presumed to be invariant to changes in scale and orientation can be detected. Scale space extreme detection can be performed using DoG (Difference of Gaussian). DoG is similar to LoG (Laplacian of Gaussian) but can be used because it is faster in terms of speed. Since the Gaussian convolution is calculated when constructing the pyramid image, DoG which calculates the difference operation of the pyramid image can obtain a similar effect as the second-differentiated LoG. Therefore, a large amount of computation can be saved.
Specifically, the Gaussian scale space can be generated to recognize objects of various sizes in the scale space extreme detection procedure. The Gaussian scale space is composed of several octaves and can have indices from 0 to N. The octave 0 can be composed of images of the same size as the input image. Each time the index of the octave increases, the size of the image is reduced by twice the width and the height. One octave can be composed of several Gaussian images. The first Gaussian image is a Gaussian filtered image with an initial sigma value. The second Gaussian image is an image filtered with a sigma value that is k times larger than the sigma value of the first Gaussian image. The rest of the Gaussian images are images filtered with sigma values that are k times larger than those in the previous sequence. That is, the nth Gaussian image may be an image filtered with a sigma value that is k n-1 times larger than the first Gaussian image.
A Gaussian filtering process is required to generate each Gaussian image from the input image. At this time, the size of the Gaussian filter may be proportional to the sigma value. Therefore, the size of the filter for generating the first Gaussian image is the smallest, and the larger the sigma value, the larger the size of the filter. As the size of the filter increases, the computational complexity increases proportionally and the processing speed decreases accordingly. With such a large number of Gaussian filter operations, the Gaussian scale space generation process is the most complex part of SIFT and takes the longest processing time.
When the initial sigma value is reduced, the Gaussian filter size also decreases proportionally, so the amount of computation is reduced. However, if the initial sigma value is lowered, the accuracy of the feature point extraction algorithm may be lowered. Conversely, if the image size is downscaled, the accuracy of the SIFT algorithm may drop.
Therefore, in the feature point extraction method based on the simplification of Gaussian scale space generation according to the embodiment of the present invention, the operation amount of the SIFT algorithm can be reduced based on the principle of Gaussian scale space. Gaussian images should be generated for every octave, but the amount of computation occupied by octave 0 is about 75% of the total Gaussian scale space creation. If the initial sigma value is halved in the octave (
That is, the feature point extraction method based on the simplification of the Gaussian scale space generation according to the embodiment of the present invention can reduce the data throughput of the SIFT algorithm based on the simplification of the Gaussian scale space when performing the SIFT algorithm.
The feature point extraction method based on the simplification of Gaussian scale space generation according to the embodiment of the present invention can reduce the size of the processed image and reduce the size of the Gaussian filter in the Gaussian scale space generation of the SIFT algorithm. This can greatly reduce the data throughput of the SIFT algorithm as well as the Gaussian scale space generation process. The feature point extraction method based on simplification of generation of a Gaussian scale space according to an embodiment of the present invention will be described in more detail in the following embodiments of the present invention.
And performs keypoint localization (step S110).
In the minutiae localization step, stable points can be removed and matching points can be precisely positioned in the continuous space rather than in the integer domain in matching among the points made as candidates of the minutiae using the Taylor series. When the DoG image is energized by the Taylor series, the binary image can be approximated to the continuous image. Since the approximated image has a different pole than the binary image, the adjustment of the feature point can be performed to determine the feature point of the feature point candidate.
An orientation assignment is performed (step S120).
In the orientation assignment stage, the orientation, which is the direction and size of the feature point, can be determined. Since the SIFT algorithm basically needs to be able to detect feature points even in the rotation change, when the feature points are determined, it is possible to determine in which direction the feature points are based on the orientation allocation step.
A minutiae descriptor is determined (step S130).
In the step of determining the feature point descriptor, the orientation of the neighboring grafient values around the feature point can be obtained. The minutiae descriptor may be substantially information for indicating the feature of the minutiae. The surrounding gradient values may be weighted using a Gaussian window. Gaussian windows can be used to avoid sudden changes in the window and to relatively less emphasize distances far from the center of the descriptor.
Minutiae matching is performed (step S140).
When the template image and the environment image are matched, the feature point database of the two images can be constructed, and all the feature points of the template image can be compared with the feature points of the environment image. The Euclidean distance between the feature points can be obtained and if the distance between the closest matching feature point and the second closest matching feature point is equal to or greater than a certain ratio, matching can be excluded.
2 is a flowchart showing a SIFT algorithm according to an embodiment of the present invention.
FIG. 2 illustrates a method for generating a Gaussian scale space in the step of detecting a scale space extreme among SIFT algorithms.
The most basic way to analyze an image over multiple scales is to perform the necessary analysis while gradually changing the size of the input image. The set of images generated in this way can be expressed by the term image pyramid.
For example, when a pedestrian is detected in an image, it is possible to determine whether the area in the window is a pedestrian by first moving the fixed size window in each scale image after generating the image pyramid. As the size of the image is reduced, the scale of the image may have a larger value in inverse proportion.
For purposes of detecting objects of various sizes, the image pyramid can be constructed by increasing the scale by 1.1 times (or 1.05 times). In this case, the size of the image may be reduced by 1 / scale times. That is, if the size of the first image is 1, then the size of the next image may be 1 / 1.1.
When creating an image pyramid, a pyramid may be created by reducing the image by a factor of 1/2 through a series of blurring and sub-sampling. Gaussian filter is used for blurring and downsampling can be performed by discarding even-numbered pixels and taking only odd-numbered pixels. The resulting image pyramid can be called the Gaussian pyramid. In Gaussian pyramids, the size of the image is rapidly reduced, so the time and memory requirement for building the pyramid is low and then the image can be analyzed quickly. However, since the Gaussian pyramid in the scale axis is an inexact sampling of the continuous scale changes that an object can have, it may have a drawback that it becomes difficult to algorithmically compare or match objects on the scale axis.
Thus, the Gaussian pyramid can be used in a coarse-to-fine form that calculates a more accurate feature from the original scale, gradually after detecting the desired feature or object in the reduced image.
The Gaussian scale space may be a space generated along the scale axis. That is, the Gaussian scale space may be a space for expressing a range of various scales that a subject can have at one time, rather than seeing only one scale or a current scale when an object is viewed. The Gaussian scale space is a space for dealing with the object structure over several scales and can be represented by a point in space with the scale axis as a parameter. The Gaussian scale space representation for image f (x, y) is a series of smoothed images fs (x, y) generated through Gaussian blurring (convolution with two dimensional Gaussian filter) . ≪ / RTI >
That is, the Gaussian scale space for an image is composed of a series of images generated through Gaussian blurring. A method of changing the scale of an image is to enlarge or reduce the image as it is shuffled. A method of changing the scale without enlarging or reducing the image directly is to blur the image. As you blur the image, the details of the image disappear and you can see the structure of the image on a larger scale.
The scale of the blurred image may be proportional to the sigma (sigma) of the Gaussian filter used. That is, if σ2 = kσ1, the scale of fσ2 (x, y) can be k times the scale of fσ1 (x, y). The point at which the scale is doubled may be an important point for generating the Gaussian pyramid described above. This is because if the size of the image is reduced to 1/2 at the point where the scale is doubled, the image structure of the scale can be maintained while minimizing loss of information due to downsampling.
In addition to maintaining the image structure, the Gaussian filter has a very useful property for generating Gaussian scale space. Convolution of a small Gaussian filter can have the same effect as applying a large Gaussian filter.
The Gaussian scale space and the octave can have the following relationship. Although the Gaussian scale space itself is defined continuously, it is practically impossible to analyze all the scales. Therefore, in order to utilize the Gaussian scale space for image processing, it is necessary to create a sampled scale space at regular intervals along the scale axis. The sampling interval at this time is called the scale step, and the commonly used scale step may be the value of root 2.
Also, Gaussian blurring and downsampling can be used in combination to create memory space efficiency and computational efficiency when creating Gaussian scale space. That is, the Gaussian filter may be applied step by step to increase the scale of the image, and then the downsampling may be performed to a half size when the scale is doubled. The Gaussian filter can then be applied again in the downsampled image. In this case, the Gaussian scale space generated by Gaussian blurring until the scale is doubled is defined as an octave, and the smaller the scale step, the more the number of images constituting the octave can be increased.
Referring to FIG. 2, when the scale step has a value of root2, three images may be included in one octave. Three images may be included in the
In the embodiment of the present invention, it is possible to generate the
3 is a conceptual diagram illustrating a method of generating a Gaussian space according to an embodiment of the present invention.
Referring to FIG. 3, pre-octave generation is performed (step S300).
In the Gaussian space generation method according to the embodiment of the present invention, one octave generation step may be added instead of generating octave 0.
The pre-octave generation step may include a Gaussian filtering step and a downscaling step.
In the Gaussian filtering step, the Gaussian filter can filter the input image with the initial sigma value. The filtered image may be reduced in size by a factor of two in the horizontal and vertical directions through a downscaling step. The reduced size image is the first Gaussian image of octave one.
After the dictionary octave generation step, a Gaussian scale space corresponding to
2, the initial sigma value used for generating the Gaussian scale space can be reduced to half (1/2). The feature point extraction and descriptor generation process of the SIFT algorithm can be performed using the generated Gaussian scale space.
When the Gaussian space generation method according to the embodiment of the present invention is used, the effect of the reduction of the processing image size and the initial sigma reduction can be obtained. First, when the octave 0 is not generated in the conventional Gaussian scale space generation process, the size of the processed image in the Gaussian scale space generation process can be greatly reduced. In addition, functions such as DoG and feature extraction performed subsequently do not process octave 0, so the amount of computation can be reduced by about 75%.
Also, as the initial sigma value is reduced by a factor of 1/2, the size of the Gaussian filter required for Gaussian scale space generation can be reduced by a factor of two. This reduction of the initial sigma value can further reduce the amount of computation of the Gaussian scale space with the reduction of the processing image size described above. Also, the size of the local patch needed to generate descriptors for each feature point may be proportional to the square of the initial sigma, and as the size of the peripheral region is reduced, the computation amount of this portion is also reduced by about 75% . That is, by reducing the initial sigma value to twice without generating octave 0 in the Gaussian scale space generation process, the computation amount of all functions of the SIFT algorithm as well as the Gaussian scale space can be reduced by 75% or more.
Due to the computational complexity of the SIFT algorithm, the SIFT algorithm can be applied to mobile devices and drones that have not been properly applied. Mobile and drone are good portability and cameras are always attached. In such devices, applications such as camera stabilization, panorama image generation, object recognition, and background separation can be performed using a feature point extraction algorithm. Previously, we used algorithms with lower computational complexity than SIFT algorithms, or extracted feature points by greatly reducing the size of the input image. However, the algorithms with a low computational complexity to replace the SIFT algorithms were limited in their application. When the feature point extraction method and apparatus based on the simplification of Gaussian scale space generation according to the embodiment of the present invention is used, the computation processing time can be greatly reduced while maintaining the accuracy of the SIFT algorithm. From this, applications with fast and accurate SIFT algorithms can be used in devices such as mobile or drone.
In addition, the Gaussian scale space creation simplification method according to the embodiment of the present invention can be used for a feature point extraction algorithm other than SIFT. Feature extraction algorithms with low complexity generate multiple octaves by downscaling an input image instead of creating a Gaussian scale space. In this case, the feature extraction algorithms have characteristics that are not as robust as the SIFT algorithm for image size variation. At this time, by using the proposed Gaussian scale space simplification method, a Gaussian scale space can be generated at a high speed.
4 is a conceptual diagram illustrating an apparatus for generating a Gaussian space according to an embodiment of the present invention.
4, the Gaussian space generating apparatus may include a pre-octave generating unit 400, a post-octave generating unit 410, and a processor 420.
The pre-octave generator 400 may perform pre-octave generation. The pre-octave generator 400 may perform pre-octave generation through a Gaussian filtering step and a downscaling step. In the Gaussian filtering step, the Gaussian filter can filter the input image with the initial sigma value. The filtered image may be reduced in size by a factor of two in the horizontal and vertical directions through a downscaling step. The reduced size image can be the first Gaussian image of octave one.
The post-octave generation unit 410 may generate a Gaussian scale space corresponding to
The processor 420 may be implemented to control operations of the pre-octave generator 400 and the post-octave generator 410.
The existing SIFT algorithm and the SIFT algorithm according to the embodiment of the present invention can be selectively utilized. For example, the existing SIFT algorithm or the SIFT algorithm according to the embodiment of the present invention can be selected and utilized in consideration of the performance of the processor, the required processing speed, and the quality of feature point extraction.
The feature point extraction method based on the simplification of Gaussian scale space creation can be implemented in an application or in the form of program instructions that can be executed through various computer components and recorded in a computer-readable recording medium. The computer-readable recording medium may include program commands, data files, data structures, and the like, alone or in combination.
The program instructions recorded on the computer-readable recording medium may be ones that are specially designed and configured for the present invention and are known and available to those skilled in the art of computer software.
Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks and magnetic tape, optical recording media such as CD-ROMs and DVDs, magneto-optical media such as floptical disks, media, and hardware devices specifically configured to store and execute program instructions such as ROM, RAM, flash memory, and the like.
Examples of program instructions include machine language code such as those generated by a compiler, as well as high-level language code that can be executed by a computer using an interpreter or the like. The hardware device may be configured to operate as one or more software modules for performing the processing according to the present invention, and vice versa.
While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it will be understood by those skilled in the art that various changes and modifications may be made therein without departing from the spirit and scope of the invention as defined in the appended claims. It will be possible.
S100: Performs Scale Space Extreme Detection
S110: Perform localization of feature points
S120: Perform orientation assignment
S130: Determine the minutiae descriptor
S140: Performing feature point matching
210: 1 st Octave
220: 2 nd Octave
S300: Create Pre-Octave
S310:
400: pre-octave generator
410: post-octave generating unit
420: processor
Claims (8)
Determining a pre-octave based on a first Gaussian filter applying a first sigma value to an input image; And
A second Gaussian filter applying a second sigma value based on the first Gaussian image constituting the dictionary octave is applied to obtain N Gaussian scale spaces corresponding to octave 1 to octave N (where N is a natural number greater than 1) And determining each of the first and second threshold values,
Wherein the first sigma value is twice the second sigma value,
Wherein the pre-octave is a first Gaussian scale space including the first Gaussian image generated by downscaling the input image,
Wherein the octave 1 is a second Gaussian scale space including an image generated by upscaling until the scale of the first Gaussian image is doubled,
And the octave N is a third Gaussian scale space including an image generated by upscaling until the scale of the (N-1) -th Gaussian image included in the Gaussian scale space corresponding to the octave N-1 is doubled Way.
Wherein the step of determining the pre-octave by applying the first Gaussian filter applying the first sigma value to the input image comprises:
Filtering the input image based on the first Gaussian filter to which the first sigma value is applied; And
Generating a first Gaussian image by downscaling the filtered image to reduce the horizontal size of the input image by a factor of 1/2 and reducing the vertical size of the input image by a factor of 1/2; Lt; / RTI >
Extracting feature point candidates from each of the N Gaussian scale spaces;
Determining feature points by removing candidate feature points that are not stable in matching based on the Taylor series among the feature point candidates;
Determining a minutiae descriptor for the minutiae; And
And performing matching on the feature points based on the feature point descriptor to determine final feature points for the input image.
Wherein the feature point extracting apparatus includes a processor,
Wherein the processor determines a pre-octave based on a first Gaussian filter applying a first sigma value to an input image,
A second Gaussian filter applying a second sigma value based on the first Gaussian image constituting the dictionary octave is applied to obtain N Gaussian scale spaces corresponding to octave 1 to octave N (where N is a natural number greater than 1) Wherein the feature point extracting unit is configured to determine each of the feature point extracting units,
Wherein the first sigma value is twice the second sigma value,
Wherein the pre-octave is a first Gaussian scale space including the first Gaussian image generated by downscaling the input image,
Wherein the octave 1 is a second Gaussian scale space including an image generated by upscaling until the scale of the first Gaussian image is doubled,
And the octave N is a third Gaussian scale space including an image generated by upscaling until the scale of the (N-1) -th Gaussian image included in the Gaussian scale space corresponding to the octave N-1 is doubled Feature point extraction device.
Wherein the processor filters the input image based on the first Gaussian filter to which the first sigma value is applied,
Wherein the first Gaussian image is generated by downscaling the filtered image to reduce the horizontal size of the input image by ½ times and reducing the vertical size of the input image by ½ times. .
The processor extracts feature point candidates from each of the N Gaussian scale spaces,
Determining feature points by removing candidate feature points that are not stable in matching based on Taylor series among the feature point candidates,
Determining a minutiae descriptor for the minutiae,
Wherein the matching unit is configured to perform matching on the feature points based on the feature point descriptor to determine final feature points for the input image.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| KR1020150188231A KR101752742B1 (en) | 2015-12-29 | 2015-12-29 | Method and apparatus for extracting key point based on simplifying generation of Gaussioan scale space |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| KR1020150188231A KR101752742B1 (en) | 2015-12-29 | 2015-12-29 | Method and apparatus for extracting key point based on simplifying generation of Gaussioan scale space |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| KR101752742B1 true KR101752742B1 (en) | 2017-06-30 |
Family
ID=59279813
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| KR1020150188231A KR101752742B1 (en) | 2015-12-29 | 2015-12-29 | Method and apparatus for extracting key point based on simplifying generation of Gaussioan scale space |
Country Status (1)
| Country | Link |
|---|---|
| KR (1) | KR101752742B1 (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109872293A (en) * | 2019-03-12 | 2019-06-11 | 深圳微品致远信息科技有限公司 | A kind of method, system and storage medium for intelligently modifying dimension of picture based on image recognition |
| KR20200043680A (en) * | 2018-10-18 | 2020-04-28 | 삼성전자주식회사 | Method of extracting features from image, image matching method using the same and image processing method using the same |
| WO2020175729A1 (en) * | 2019-02-26 | 2020-09-03 | 연세대학교 산학협력단 | Apparatus and method for detecting facial feature point by using gaussian feature point map and regression scheme |
| CN111739073A (en) * | 2020-06-24 | 2020-10-02 | 刘秀萍 | Efficient and rapid image registration optimization method for handheld device |
| CN117251087A (en) * | 2023-11-17 | 2023-12-19 | 济宁市金桥煤矿 | Coal mine safety simulation interaction method based on virtual reality |
-
2015
- 2015-12-29 KR KR1020150188231A patent/KR101752742B1/en active IP Right Grant
Non-Patent Citations (1)
| Title |
|---|
| 김응섭 외 1인, "메모리 크기 감소를 위한 SIFT 하드웨어 구조", 대한전자공학회 학술대회 , 2009.11, 75-76 (2 pages).* |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20200043680A (en) * | 2018-10-18 | 2020-04-28 | 삼성전자주식회사 | Method of extracting features from image, image matching method using the same and image processing method using the same |
| KR102528453B1 (en) * | 2018-10-18 | 2023-05-02 | 삼성전자주식회사 | Method of extracting features from image, image matching method using the same and image processing method using the same |
| WO2020175729A1 (en) * | 2019-02-26 | 2020-09-03 | 연세대학교 산학협력단 | Apparatus and method for detecting facial feature point by using gaussian feature point map and regression scheme |
| CN109872293A (en) * | 2019-03-12 | 2019-06-11 | 深圳微品致远信息科技有限公司 | A kind of method, system and storage medium for intelligently modifying dimension of picture based on image recognition |
| CN111739073A (en) * | 2020-06-24 | 2020-10-02 | 刘秀萍 | Efficient and rapid image registration optimization method for handheld device |
| CN117251087A (en) * | 2023-11-17 | 2023-12-19 | 济宁市金桥煤矿 | Coal mine safety simulation interaction method based on virtual reality |
| CN117251087B (en) * | 2023-11-17 | 2024-02-09 | 济宁市金桥煤矿 | Coal mine safety simulation interaction method based on virtual reality |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Choi et al. | Context-aware deep feature compression for high-speed visual tracking | |
| KR101752742B1 (en) | Method and apparatus for extracting key point based on simplifying generation of Gaussioan scale space | |
| EP2989607B1 (en) | Method and device for performing super-resolution on an input image | |
| JP4613617B2 (en) | Image processing system, learning apparatus and method, and program | |
| CN111768432A (en) | Moving target segmentation method and system based on twin deep neural network | |
| US9530079B2 (en) | Point spread function classification using structural properties | |
| CN110334762B (en) | Feature matching method based on quad tree combined with ORB and SIFT | |
| WO2015148824A1 (en) | System for filtering, segmenting and recognizing objects in unconstrained environments | |
| US9196053B1 (en) | Motion-seeded object based attention for dynamic visual imagery | |
| KR102070956B1 (en) | Apparatus and method for processing image | |
| US10582179B2 (en) | Method and apparatus for processing binocular disparity image | |
| Patil et al. | Multi-frame recurrent adversarial network for moving object segmentation | |
| CN110516731B (en) | Visual odometer feature point detection method and system based on deep learning | |
| Noor et al. | Gradient image super-resolution for low-resolution image recognition | |
| Leal et al. | Smartphone camera document detection via Geodesic Object Proposals | |
| CN113129332A (en) | Method and apparatus for performing target object tracking | |
| JP2010271792A (en) | Image processing apparatus and method | |
| CN108182391B (en) | Method for automatically extracting nonrepetitive slide based on education video | |
| Chaudhary et al. | A comparative study of fruit defect segmentation techniques | |
| Vaidya et al. | Comparative analysis of motion based and feature based algorithms for object detection and tracking | |
| US10026181B2 (en) | Method and apparatus for detecting object | |
| Nguyen et al. | Unconstrained gender classification by multi-resolution LPQ and SIFT | |
| Do et al. | A single-column convolutional neural networks for crowd counting | |
| El-Shekheby et al. | Restoration of Spatially-Varying Motion-Blurred Images | |
| JP5995943B2 (en) | Image feature extraction apparatus, method, and program |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| E701 | Decision to grant or registration of patent right | ||
| GRNT | Written decision to grant |