KR101479225B1 - Method and apparatus for generating feature vector, and method and apparatus for Image Recognition using the same - Google Patents

Abstract

The present invention relates to a new method and an apparatus for generating a feature vector to express local binary patterns in eight-sided prism form as a technology of recognizing an image, and a method and an apparatus for recognizing an image using the same. The method for generating a feature vector according to the present invention includes: comparing a pixel value of a target pixel included in at least a partial region among input images with pixel values of pixels adjacent to the target pixel; and generating at least a three-dimensional feature vector according to the target pixel.

Description

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and apparatus for generating a feature vector,

The present invention relates to a method and apparatus for generating feature vectors for image recognition, and more particularly, to a method and apparatus for image recognition using Local Binary Patterns.

Face recognition technology refers to the identification of features by extracting features from the faces of people existing in the images or images and comparing them with the data stored in the existing ones.

Unlike other biometrics technologies, facial recognition technology does not require a part of the body to be in direct contact with the device, and the method of acquiring biometric information has a low degree of impossibility. However, changes in illumination, changes in facial expression and posture, aging, and change in size due to the perspective are distorting the facial image, which is a serious obstacle to stable recognition.

Particularly, the phenomenon that the face image is distorted by the illumination has a serious effect on the recognition rate. There are also cases where the changes to the same face in different lighting conditions are greater than those that occur between different faces. There are appearance-based, normalization-based, and feature-based techniques for solving the problem of face recognition due to illumination change, -based technique has been proposed in a number of techniques such as gray-level transformation, gradient and edge extraction, and face reflection field estimation, There is a problem that it is not easy to distinguish the face identity from the other face difference.

As a patent document, Korean Patent Laid-Open Publication No. 2011-67480 discloses a feature point detection method for detecting a face, in particular, a method for detecting a feature point, comprising: receiving n input images; Extracting feature points of each input image for each divided region, and converting a histogram, which is a distribution of feature points extracted for each region, into a dimension vector using a uniform pattern.

In the case of using the conventional local binary pattern, there is a problem that the local binary patterns of the regions similar to each other are similar and treated as similar images, and the local binary patterns of similar regions are not similar, It is an object of the present invention to provide a method and apparatus for generating a three-dimensional feature vector as a local binary pattern of a novel concept to alleviate this problem.

It is another object of the present invention to provide a method of generating the feature vector and an image recognition method and apparatus using the apparatus.

According to another aspect of the present invention,

Receiving an image; And comparing at least a pixel value of the target pixel included in the input image with a size of pixel values of adjacent pixels adjacent to the target pixel and generating at least three dimensional feature vectors according to the target pixel .

In the present invention, the feature vector is a three-dimensional feature vector, and the three-dimensional feature vector is used to determine a pixel value of the pixel and a pixel value of the adjacent pixels A vector selected from a three-dimensional structure in which local binary patterns defined through comparison of pixel values are arranged in three dimensions is preferable.

The local binary patterns are patterns in which information of N bits (where N is an integer of 8 or more) bits are expressed in the form of a ring, and the three-dimensional structure is a structure in which the patterns are positioned according to the morphological similarity between the patterns .

In the present invention, the three-dimensional structure preferably has a structure in which a spatial distance between patterns having a high similarity is smaller than a spatial distance between patterns having a low similarity.

Preferably, the local binary patterns are patterns in which the 8-bit information is expressed in the form of a ring, and the 3-dimensional structure is formed by arranging the patterns in accordance with the morphological similarity between the patterns expressed in the form of the ring It may be a structure of each column type. In addition, the three-dimensional structure may be arranged such that the same patterns in which the 8-bit information is expressed in the form of a ring are rotated in different directions.

In order to solve still another problem of the present invention,

An input unit for inputting an image; And comparing the pixel value of the target pixel included in at least a part of the input image with the size of the pixel values of the adjacent pixels adjacent to the target pixel and generating at least three dimensional feature vectors according to the target pixel, And a vector generating unit.

The feature vector generating unit may include a comparator for comparing the pixel value of the pixel with the pixel value of the adjacent pixel adjacent to the pixel with respect to each of pixels included in at least a part of the area included in the input image; And a generation unit that generates a feature vector of the pixel using the comparison result of the comparison unit. The generation unit may determine a three-dimensional feature vector of the pixel in the three-dimensional structure.

According to another aspect of the present invention, there is provided an image recognition method comprising: receiving an image; And comparing pixel values of the pixels and pixel values of adjacent pixels adjacent to the pixel with respect to each of pixels included in at least a part of the input image and generating feature vectors of the compared pixels; And comparing the generated feature vectors with training feature vectors generated from previously prepared training images.

An image recognition apparatus of the present invention includes: an input unit for inputting an image; And a feature vector generation unit for comparing the pixel value of the pixel with the pixel values of the adjacent pixels adjacent to the pixel with respect to each of the pixels included in at least a part of the input image and generating feature vectors of the compared pixels, part; And a feature vector comparing unit for comparing the generated feature vectors with training feature vectors generated from a training image prepared in advance.

According to the present invention, a problem that a local binary pattern of regions that are not similar to each other is treated as a similar image, a problem of image recognition error due to a problem in which local binary patterns of similar regions are treated as non- . There is also a strong effect on the change of illumination.

1 is a flowchart illustrating a method of generating a feature vector according to an embodiment of the present invention.
2 is a block diagram illustrating an apparatus for generating feature vectors according to an embodiment of the present invention.
3 is a diagram for explaining a conventional local binary pattern generation method.
4 is a view for explaining a problem of a conventional local binary pattern.
5 is a diagram for explaining the difference of the histogram intersection under different illumination.
FIG. 6A is a diagram illustrating a conventional one-dimensional local binary pattern, FIG. 6B is a diagram illustrating a local binary pattern rearranged to have rotation-invariant characteristics, FIG. 6C is a diagram illustrating an octagonal local binary pattern array newly defined in the present invention It is a picture.
7 is a diagram showing a three-dimensional structure in which a local binary pattern of the present invention is arranged in space.
Fig. 8 shows the arrangement of Fig. 7 as a look-up table of coordinate values.
9 is a flowchart illustrating an image recognition method using a local binary pattern according to an embodiment of the present invention.
10 is a block diagram illustrating an image recognition apparatus using a local binary pattern according to an embodiment of the present invention.
FIG. 11 shows an embodiment for weighting face images.

The following merely illustrates the principles of the invention. Therefore, those skilled in the art will be able to devise various apparatuses which, although not explicitly described or shown herein, embody the principles of the invention and are included in the concept and scope of the invention. It is also to be understood that all conditional terms and examples recited in this specification are, in principle, expressly intended for the purpose of enabling the inventive concept to be understood and are not to be construed as limited to such specifically recited embodiments and conditions do. It is also to be understood that the detailed description, as well as the principles, aspects and embodiments of the invention, as well as specific embodiments thereof, are intended to cover structural and functional equivalents thereof. It is also to be understood that such equivalents include all elements contemplated to perform the same function irrespective of the currently known equivalents as well as the equivalents to be developed in the future, i.e., the structure.

In the claims of the present specification, the elements represented as means for performing the functions described in the detailed description may include, for example, a combination of circuit elements performing the function or any type of software including firmware / And is coupled with appropriate circuitry for executing the software to perform the function. It is to be understood that the invention as defined by the appended claims is not to be interpreted as limiting the scope of the invention as defined by the appended claims, as the functions provided by the various enumerated means are combined and combined with the manner in which the claims require. .

It is to be understood that when an element is referred to as being "connected" or "connected" to another element, it may be directly connected or connected to the other element, . On the other hand, when an element is referred to as being "directly connected" or "directly connected" to another element, it should be understood that there are no other elements in between. Unless defined otherwise, all terms (including technical and scientific terms) used herein may be used in a sense commonly understood by one of ordinary skill in the art to which this invention belongs. Also, commonly used predefined terms are not ideally or excessively interpreted unless explicitly defined otherwise.

In this specification, a singular form may include plural forms unless specifically stated in the phrase. It is noted that the terms "comprises" and / or "comprising" used in the specification are intended to be inclusive in a manner similar to the components, steps, operations, and / Or additions.

BRIEF DESCRIPTION OF THE DRAWINGS The above and other objects, features and advantages of the present invention will become more apparent from the following detailed description taken in conjunction with the accompanying drawings, in which: . In the following description, a detailed description of known technologies related to the present invention will be omitted when it is determined that the gist of the present invention may be unnecessarily blurred. Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings.

The present invention discloses a method and apparatus for generating a three-dimensional feature vector by using a three-dimensional octagonal prism (LBP) and a method and apparatus for recognizing an image using the same. The local binary pattern is encoded in coordinates on a three-dimensional space, thereby reducing degradation of image recognition performance due to illumination changes compared to scalar value encoding of a localized binary pattern according to conventional methods. As the image described in this specification, various types of images such as a human face or an image including the human face can be used. Hereinafter, the invention disclosed in this specification will be described in detail with reference to the drawings.

1 is a flowchart illustrating a method of generating a feature vector according to an embodiment of the present invention.

The feature vector generation method shown in FIG. 1 includes the following steps, which are performed in a time-series manner in the feature vector generation apparatus 200.

In step S101, the input unit 201 receives a target image for image recognition and image classification. The image received in this step is an object image to be recognized or classified, for example, a face image, a fingerprint image, and the like.

In step S102, the preprocessing unit 202 performs preprocessing including image normalization on the input image. This step is performed before the feature vector generation, and is performed to reduce the illumination effect, etc., and can be selectively performed as needed. In addition, the image normalization in this step may be implemented to selectively perform scaling on the image and perform normalization on the scaled image.

In step S103, the feature vector generation unit 203 generates a three-dimensional feature vector of the preprocessed image. The feature vector generation unit 203 compares the pixel values of each pixel included in a part of the input image with the pixel values of adjacent pixels adjacent to the pixel, And generates a feature vector. Herein, the pixel value is a concept of a signal value of a pixel, which will be apparent to those skilled in the art.

In step S103, the feature vector is generated for each pixel, and a three-dimensional vector is preferable. In general, the feature vector is used for classifying or judging similarity of a technical image through a computer vision technique. In the present embodiment, a feature vector is a three-dimensional vector of existing one-dimensional local binary patterns. That is, it is characterized in that existing local binary patterns are expressed by three-dimensional coordinates or three-dimensional vectors specified by a three-dimensional structure.

In this embodiment, the three-dimensional feature vector corresponds to a specific point in which the local binary patterns according to the known concept are arranged at regular intervals along the respective vertexes of the three-dimensional regular polygonal column and the side faces of the regular polygonal column. In other words, in the present embodiment, existing local binary patterns are expressed by coordinates specified through three-dimensional regular polygonal columns or the like. Of course, a cylindrical structure can also be used.

The localized binary patterns disposed adjacent to each other are preferably uniform patterns. Here, the uniform pattern means that the transition between 0 and 1 in the binary pattern consisting of 0 and 1, that is, the number of spatial conversions is less than 2 times. The local binary pattern may be a pattern composed of 3 x 3 pixels, and may be composed of 4 x 4 or more pixels according to an embodiment. Also, the shapes are similar on the plane coordinates of the three-dimensional columns, and the distances between adjacent local binary patterns may be equal to each other. In this case, it may be a regular polygonal or cylindrical structure, or it may be regular octagonal when the local binary pattern is a pattern composed of 3 x 3 pixels. The distance between adjacent local binary patterns on the height coordinates of the three-dimensional column may be the same. By doing so, similar patterns in different illumination environments can be placed close to each other on three-dimensional coordinates, and illumination change robust image recognition is possible.

2 is a diagram for explaining a feature vector generating apparatus according to an embodiment of the present invention.

Referring to FIG. 2, the feature vector generation apparatus 200 of the present embodiment includes an input unit 201, a preprocessing unit 202, and a feature vector generation unit 203. In addition, the feature vector generation unit 203 of the embodiment compares the pixel value of the pixel with the pixel value of the adjacent pixel adjacent to the pixel, with respect to each pixel included in at least a part of the input image, (Not shown) for generating a three-dimensional feature vector of the pixel using the comparison result of the comparison unit. In addition, the feature vector generation apparatus 200 may further include a storage unit (not shown) for storing the generated feature vectors.

In the present embodiment, the feature vector is a three-dimensional feature vector selected from a predefined three-dimensional structure. Here, the predefined three-dimensional structure is a three-dimensional structure in which, in order to perform classification or similar determination of an image on a computer vision, a pixel value of an arbitrary pixel and a size of pixel values of adjacent pixels adjacent to the arbitrary pixel Dimensional structure defined by associating at least a part of local binary patterns defined on a three-dimensional plane with a three-dimensional structure.

The local binary patterns may be defined as a pattern in which information of N bits (where N is an integer equal to or greater than 8) bits are expressed in the form of a ring, and the three- Can be defined as a structure positioned in the structure. Although it is also possible to make all of the local binary patterns correspond to the three-dimensional structure, the three-dimensional feature vector according to this embodiment is a vector defined by associating at least a part of the local binary patterns with the three-dimensional structure. It is preferable that the number of spatial transitions is equal to or less than a predetermined value in the local binary pattern corresponding to the three-dimensional structure. Particularly, it is preferable that the spatial transition is 0 or 2.

Here, the spatial transition means that the value of the local binary pattern changes from 0 to 1, or 1 to 0. The value of the 3 * 3 size local binary pattern is 0 to 255, and the total number of local binary patterns is 256, but 58 uniform characteristic vectors having linear characteristics are selected as shown in FIG. It is possible to associate the feature vectors with the three-dimensional structure of the octagonal columnar shape. In this embodiment, the three-dimensional vector value according to the corresponding position is defined as a three-dimensional feature vector. Referring to 6b and 6c to be described later, the predefined three-dimensional structure is defined so that the spatial distance between local binary patterns having high morphological similarity is smaller than the spatial distance between the local binary patterns having low morphological similarity desirable. In the eight columnar structure, it is preferable that the patterns having the same number of 0 or 1 are arranged in a certain direction (clockwise or counterclockwise).

The operation contents of the feature vector generating apparatus described in FIG. 2 are duplicated with the feature vector generating method of FIG. 1, so that the description thereof will be omitted.

The preprocessing unit 202 may perform an image normalization procedure or the like. The image normalization procedure can be performed in advance to reduce the image recognition rate degradation due to illumination before proceeding with the local binary pattern encoding. In particular, the preprocessing unit can perform gamma correction, DoG (Difference of Gaussian) filtering, and equalization of variation. Through such a preprocessing process, degradation of the image recognition rate due to illumination can be reduced.

3 is a diagram for explaining the basic concept of a local binary pattern. A local binary pattern is calculated for each pixel belonging to an image. In this embodiment, a pixel for which a local binary pattern value is to be calculated is called a target pixel, and a pixel adjacent to the pixel is called an adjacent pixel. Referring to FIG. 3, the local binary pattern compares the values (g ₁ to g ₈ ) of the 3 × 3 neighboring pixels around the pixel with the pixel value (g _c ) of the target pixel located at the center, The number of 8 bits constituted by the allocation becomes the local binary pattern value for the center pixel g _c .

The histogram of local binary pattern values within the local region may be a feature descriptor for that region. All pixels of the image are coded with an 8-bit local binary pattern value (LBP value) by the following equations (1) and (2).

[Equation 1]

&Quot; (2) "

4 is a diagram for explaining a problem of a conventional local binary pattern encoding method. The conventional encoding method has a problem in that it can generate a local binary pattern value with a large difference even in a similar structure and this difference can be larger than the value of the local binary pattern between structures which are completely different from each other.

Referring to FIG. 4A, if two localized binary patterns are extracted from the selected image, it can be seen that the localized binary pattern values are similar to each other, although the two localized binary patterns are completely different from each other. When comparing two patterns from these local binary pattern values, it can be determined that the two patterns are very similar to each other.

4 (b) is a view for explaining a case opposite to that of FIG. 4 (a). Referring to FIG. 4B, when two localized binary patterns are extracted from the selected image, although the two localized binary patterns are similar to each other, the local binary pattern values according to the existing definition are different from each other, Can be determined to be different.

The problem of the conventional local binary pattern encoding method described in FIGS. 4 (a) and 4 (b) is a major cause of deterioration in recognition performance in illumination change.

Most methods that use the distribution characteristics of localized binary patterns use bin-to-bin distance measurements to exploit the histogram difference. It is not easy to avoid the problems of Figs. 4 (a) and 4 (b) described above when using this binocular distance measurement. Although the localized binary pattern feature is invariant to a particular monotonic gray scale transformation, it is not invariant to the illumination change. However, in most cases the illumination change still produces a similar localized binary pattern. That is, even though the localized binary pattern is not invariant under illumination changes, the overall structure may still be similar.

5 is a diagram for explaining a difference of a histogram intersection under different illumination. Figures 5 (a) and 5 (b) show the lips of the same person under different illumination and their corresponding local binary patterns.

Referring to Figures 5 (a) and 5 (b), the diagonal edge structure is similar in spite of the illumination variation according to the illumination change in the local binary pattern. In this case, the local binary pattern values are similar to each other, but the encoded local binary pattern values are calculated differently. Thus, when using the binovision distance measure, even these similar patterns will result in a lower score in the matching procedure.

For example, when the local binary pattern histogram is generated from FIGS. 5A and 5B, the two histogram distributions may be similar to each other, as shown in FIG. 5C. However, applying the histogram intersection to these two histograms, the similarity score can be calculated to be low despite a similar histogram distribution. That is, while illumination changes cause a change in small local patterns, empty distance measurements produce a low similarity score.

Therefore, when encoding an image using the conventional local binary pattern concept, local binary pattern values are encoded in a one-dimensional scalar value, so that even though they are similar patterns, a large difference in local binary pattern values do.

Accordingly, in order to alleviate this problem, the present invention proposes a method of encoding a local binary pattern value on three dimensions. This approach makes it possible to assign similar local binary pattern values for similar structures.

6A to 6C are diagrams for explaining a relationship between a conventional local binary pattern and a local binary pattern of the present invention. 6A to 6C illustrate three-dimensional coordinates of an octagonal column with respect to a 3X3 localized binary pattern, but the present invention is not limited thereto, and the present invention can be applied to a polygonal or circular column with N X N (where N is an integer of 2 or more).

6 (a) is a diagram for explaining a conventional one-dimensional encoding method of a local binary pattern. It has a scalar value from 0 to 255 for each binary pattern.

6 (b) illustrates the local binary pattern arrangement disclosed herein. Each local binary pattern shown is a uniform pattern. As shown, in the 3 X 3 block, a total of 58 patterns can be arranged. The uniform pattern belonging to Fig. 6 (b) means a localized binary pattern having a maximum of two spatial transitions. An example of this technique can be found in Ojala et al, "Multiresolution Gray-Scale and Rotational Invariant Texture Classification with Local Binary Patterns ".

Here, the uniform pattern is defined as a local binary pattern with a maximum of two spatial transitions. For example, in a face image, most of the pixels (about 90% or more of the pixels) . ≪ / RTI > In FIG. 6 (a), the first pattern (0) has no spatial transition, the second pattern (1) has two spatial transitions, and the sixth pattern (5) has four spatial transitions. 6 (b) is a pattern in which the switching from 0 to> 1 or 1 to> 0 in the patterns 0 to 255 in FIG. 6 (a) occurs twice or less.

More specifically, the following example will be described in connection with the spatial transition. The pattern below shows an example of a conventional localized binary pattern.

0 (starting point) 0 0

0 1

0 0 0

In this case, when starting from the upper left point and returning to the original position again, 0-> 1 conversion occurs once and 1 -> 0 conversion occurs once, resulting in a total of 2 conversions. So this corresponds to the uniform pattern. Consider the following cases.

0 (starting point) 0 0

1 1

1 0 0

In this case, 0-> 1 conversion occurs twice and 1-to-0 conversion occurs twice, resulting in a total of 4 conversions, which is not a uniform pattern. Of the total 256 patterns, there are only 58 in-form patterns. In the case of all 0's and 1's in the 58 uniform patterns shown in FIG. 6 (b), no conversion occurs, and in the remaining 56 patterns, a total of two conversions occur.

When the conventional local binary pattern belonging to FIG. 6A does not belong to FIG. 6B, that is, when the corresponding local binary pattern is not a uniform pattern, the corresponding pixel is defined to be excluded in a comparison process between subsequent patterns for image recognition . Or the uniformity degree of the image, or the uniformity degree of the pattern, or a uniformity pattern shown in FIG. 6B.

6C is an octagonal structure in which the eight patterns included in the first column in FIG. 6B are arranged at a constant distance from each other. Referring to FIG. 6C, the local binary patterns are arranged in a circular shape at the vertexes of the octagonal shape, and the distances between the adjacent patterns can be arranged to be the same. In Fig. 6C, each pattern is arranged at the same distance (k = 1) from the center. The plane coordinates of each pattern are shown in Fig.

FIGS. 7A and 7B are diagrams for explaining a three-dimensional structure in which feature vectors of the present invention, that is, local binary patterns are arranged on three dimensions.

FIG. 7A shows a local binary pattern arranged in two adjacent plane coordinates. This is the same as the local binary pattern of the first and second columns in Fig. 6B. The first and second columns have a distance difference of h on the height coordinates of the three-dimensional coordinates.

FIG. 7B shows that the local binary patterns arranged in FIG. 6B are all arranged on the three-dimensional coordinates of the octagonal columns. The local binary pattern placed on the plane coordinate may have the same height (h) for the height coordinate. The coordinates of each local binary pattern are arranged on the side of the octagonal column.

FIG. 8 shows a lookup table for specifying a coordinate value in the three-dimensional structure of FIG. 7B. Referring to FIG. 8, x-y denotes plane coordinates and z denotes height coordinates.

The local binary patterns of the similar structure are arranged at positions close to each other by the three-dimensional arrangement of the local binary patterns, and when the input image is compared with the preliminarily prepared training image, the distance between the encoded local binary patterns is calculated, It is possible to judge whether or not they are similar to each other. Euclidean distance calculation can be used as the distance calculation method.

The feature vector having the three-dimensional dimension described above can be calculated for every pixel of the input image, but is preferably calculated for a pixel having a uniform pattern. In consideration of image recognition performed later, it is preferable to generate a three-dimensional feature vector for pixels corresponding to a specific region used for image recognition. For example, it is implemented to be performed only on feature points included in an area important for image recognition after preprocessing (scaling, normalization, etc.) on the image, for example, face recognition, jaw line, eye, nose, and mouth .

9 is a flowchart illustrating an image recognition method using a feature vector according to an embodiment of the present invention. The image recognition method of this embodiment includes the following steps performed in the image recognition apparatus 400 in a time series manner.

First, in step S301, the input unit 401 receives a target image for image recognition.

In step S302, the feature vector generation unit 402 generates a feature vector by using the three-dimensional structure in which the local binary patterns specified according to the magnitude comparison between the pixel value of a certain pixel and the pixel values of adjacent pixels adjacent to the pixel are arranged in three dimensions A feature vector to be selected is generated for each pixel.

In step S303, the feature vector comparison unit 403 compares the feature vectors generated in step S302 with previously prepared stored images or feature vectors generated from the training images.

Since the image recognition method using the three-dimensional feature vector described in FIG. 9 is substantially the same as the description of FIG. 1, a common description will be omitted.

10 is a block diagram illustrating an image recognition apparatus using a feature vector according to an embodiment of the present invention.

The input unit 401 receives an image to be recognized.

The feature vector generating unit 402 generates a feature vector for each of the pixels included in at least a part of the input image by using the pixel value and the local binary patterns specified by the magnitude comparison between the pixel values of the adjacent pixels adjacent to the pixel A feature vector selected from three-dimensional structures arranged on three dimensions is generated for each pixel. The feature vector generating unit 402 includes a comparator (not shown) for comparing the pixel value of its own with the pixel value of the adjacent pixel, and a generating unit (Not shown). Although not shown in FIG. 10, a pre-processing unit may be further included between the feature vector generating unit 402 and the input unit 401 to perform the pre-processing operation described above.

The feature vector comparison unit 403 compares the feature vector of the input image generated by the feature vector generation unit 402 with the feature vector of the already stored image or the training image. The feature vector comparing unit 403 can determine that the images to be compared are similar if the similarity between the feature vectors satisfies a predetermined criterion, according to the comparison result.

The storage unit 404 stores feature vectors previously extracted from the stored images or training images.

For example, when the image recognition apparatus of the present embodiment is a face recognition system for access control, the storage unit stores face images of people who are allowed to enter and leave, and three-dimensional features extracted in advance according to the definition of the present invention from the face images Information about the vector can be stored.

The feature vector comparing unit 403 does not need to compare the feature vectors according to all the pixels included in the contrast images with the feature vectors according to all the pixels included in the stored images. Preferably, the feature vector comparing unit 403 can compare only the feature points of the selected feature points according to a predetermined reference, or the feature vectors of the pixels included in the specific region. In order to compare an input image with a previously stored image or a training image, each pixel to be compared must be specified. This can be performed through various conventional methods for extracting image feature points.

In this embodiment, since only the uniform pattern is defined as a three-dimensional feature vector, which is a vector specified in the three-dimensional structure, it can be implemented not to perform comparison with respect to a pattern that is not a uniform pattern. This can improve the processing speed and have a strong noise effect. If the scales of the contrast images are not the same, a method of matching the size of the image with the size of the stored image by interpolating the image of a small size image may be used in order to compare the feature vectors. Preferably, The characteristic vectors of the image may be prepared in advance for several scale criteria, and the feature vectors may be compared after comparing the scales.

Hereinafter, the concept of defining the feature vector of the present embodiment and comparing feature vectors will be described through a mathematical definition.

For example, according to the above-described embodiment, the local binary pattern value at the X position can be expressed by a three-dimensional feature vector (OPR: Octagonal Prism Representation) on a three-dimensional space as shown in Equation (3).

&Quot; (3) "

,

,

은 국부이진패턴 값으로부터 생성된 3차원 벡터를 생성하는 함수를 의미한다. 이러한 함수는 도 8에 의해 정의될 수 있다. h는 1 또는 임의의 값으로 정의할 수 있다.Here, l (X) means a local binary pattern value at the X position,

,

,

Means a function for generating a three-dimensional vector generated from a local binary pattern value. Such a function can be defined by Fig. h can be defined as 1 or any value.

)과 레퍼런스 영상의 픽셀(

) 간의 OPR 값의 유사도는 [수학식 4]와 같이 유클리디안 거리로 계산될 수 있다.The pixel of the test image at an arbitrary X position

) And the pixel of the reference image

) Can be calculated as an Euclidean distance as shown in Equation (4).

&Quot; (4) "

Since the localized binary pattern is placed on an octagonal prism (column), the Euclidean distance can be equivalent to the circular distance measurement. The similarity of the octagonal prism representation (SOPR) between the reference image and the test image can be calculated by Equation (5).

&Quot; (5) "

Here, N is the number of pixels that are uniform patterns in the reference image and the test image, S is the uniform local pattern set, and B is the comparison area.

When the invention disclosed in this specification is applied to face recognition, it is the most important feature part in distinguishing the eyes from the face of a person compared to the chin and the ball. Therefore, weights can be given to the corresponding regions of the image to emphasize these eye and mouth features.

11 shows an embodiment in which weights are given according to regions in a face image.

Referring to Fig. 11, the eyes and mouth have high weights, and the jaws and balls have low weights. When this weight is given to the SOPR, the following equation (6) is obtained.

&Quot; (6) "

는 위치 X에 대한 가중치를 의미한다. here,

Is a weight for position X. [

In the case of face recognition, the recognition result appears as the person whose SOPR value is minimum.

The feature vector generation method and the image recognition method described above can be configured as a computer-readable recording medium.

A computer-readable recording medium includes all kinds of recording media in which data that can be read by a computer system is stored. Examples of the computer-readable recording medium include a RAM, a ROM, a CD-ROM, a magnetic tape, an optical data storage device, a floppy disk, and the like.

In addition, the computer-readable recording medium may be distributed over network-connected computer systems so that computer readable codes can be stored and executed in a distributed manner. And, functional programs, codes, and code segments for implementing the recording method can be easily deduced by programmers in the technical field to which the present invention belongs.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is to be understood that the invention is not limited to the disclosed exemplary embodiments, but, on the contrary, is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims. It can be understood that the modification and application of the present invention are possible. For example, each component specifically shown in the embodiments can be modified and implemented. It is to be understood that the present invention may be embodied in many other specific forms without departing from the spirit or essential characteristics thereof.

201: input unit
202: Feature vector generating unit
200: Feature vector generation device

Claims (20)

Receiving an image; And
Comparing a pixel value of an object pixel included in at least a part of the input image with a size of pixel values of adjacent pixels adjacent to the object pixel,
At least some of the local binary patterns defined through comparison of pixel values of an arbitrary pixel according to the target pixel and pixel values of neighboring pixels adjacent to the arbitrary pixel, And generating a three-dimensional feature vector selected from a defined three-dimensional structure. delete The method according to claim 1,
The local binary patterns are patterns in which information of N bits (where N is an integer of 8 or more) is expressed in the form of a ring,
Wherein the three-dimensional structure is located on the three-dimensional structure according to morphological similarity between the localized binary patterns. The method according to claim 1,
At least some of the localized binary patterns include
Wherein the spatial transition of the localized binary patterns includes a number of times that the spatial transition of the localized binary patterns changes from 0 to 1 or 1 to 0. The method of claim 3,
Wherein the predefined three-dimensional structure is a structure defined such that a spatial distance between local binary patterns with high morphological similarity is smaller than a spatial distance between local binary patterns with low morphological similarity. Generation method. The method of claim 3,
Wherein the localized binary patterns are 8-bit information and are represented in a ring form, and the predefined three-dimensional structure comprises a plurality of local binary patterns corresponding to the local similarity between the local binary patterns expressed in the form of the ring Wherein the shape of the feature vector is defined by an octagonal columnar structure. The method according to claim 6,
The octagonal column structure in which the localized binary patterns are located according to the morphological similarity,
Wherein the patterns having the same number of 0 or 1 out of the patterns expressed in the ring form are sequentially positioned in a three-dimensional space so as to be parallel to the bottom face of the octagonal columnar structure, . 8. The method of claim 7,
Wherein the patterns having the same number of 0 or 1 are arranged clockwise or counterclockwise in the octagonal columnar structure. An input unit for inputting an image; And
Comparing a pixel value of an object pixel included in at least a part of the input image with a size of pixel values of adjacent pixels adjacent to the object pixel,
At least some of the local binary patterns defined through comparison of pixel values of an arbitrary pixel according to the target pixel and pixel values of neighboring pixels adjacent to the arbitrary pixel, And a feature vector generation unit for generating a feature vector of three dimensions selected from the defined three-dimensional structure. delete The apparatus according to claim 9,
A comparison unit for comparing the pixel value of the pixel and the pixel values of adjacent pixels adjacent to the pixel with respect to each of pixels included in at least a part of the area included in the input image; And a generation unit that generates a feature vector of the pixel using the comparison result of the comparison unit,
Wherein the generation unit determines a three-dimensional feature vector of the pixel in the three-dimensional structure. 10. The method of claim 9,
Wherein the local binary patterns are patterns in which information of N bits (where N is an integer equal to or greater than 8) bits are expressed in the form of a ring, and the three-dimensional structure is a pattern in which the patterns are positioned according to the morphological similarity between the localized binary patterns Wherein the three-dimensional structure is a structure defined such that a spatial distance between patterns having high morphological similarity is smaller than a spatial distance between patterns having low morphological similarity. 13. The method of claim 12,
Wherein the local binary patterns are patterns in which 8-bit information is expressed in the form of a ring, and the 3-dimensional structure is an octagonal columnar shape in which the patterns are arranged according to the morphological similarity between patterns expressed in the form of the ring And a feature vector generating unit for generating the feature vector. Receiving an image;
Comparing a pixel value of an object pixel included in at least a part of the input image with a size of pixel values of adjacent pixels adjacent to the object pixel and comparing the pixel value of an arbitrary pixel with the target pixel, Dimensional feature vector corresponding to at least a part of the local binary patterns defined through comparison of pixel values of neighboring pixels adjacent to the neighboring pixels and a predefined three-dimensional structure step; And
And comparing the generated feature vectors with training feature vectors generated from previously prepared training images. delete 15. The method of claim 14,
At least some of the Local Binary Patterns may be < RTI ID = 0.0 >
Wherein the spatial transition of the localized binary patterns includes a number of times that the spatial transition is 0 or 1, and the number of times is less than or equal to 2. 15. The method of claim 14,
Wherein comparing the generated feature vectors with training feature vectors generated from a training image prepared in advance is performed using three-dimensional distances between feature vectors corresponding to each other. An input unit for inputting an image;
Comparing a pixel value of an object pixel included in at least a part of the input image with a size of pixel values of adjacent pixels adjacent to the object pixel,
At least some of the local binary patterns defined through comparison of pixel values of an arbitrary pixel according to the target pixel and pixel values of neighboring pixels adjacent to the arbitrary pixel, A feature vector generation unit for generating a three-dimensional feature vector selected from a defined three-dimensional structure; And
And a feature vector comparing unit for comparing the generated feature vectors with training feature vectors generated from a training image prepared in advance. delete 19. The apparatus according to claim 18,
A comparison unit for comparing the pixel value of the pixel with the pixel values of the adjacent pixels adjacent to the pixel, with respect to each of the pixels included in at least a part of the input image; And a generation unit that generates a feature vector of the pixel using the comparison result of the comparison unit,
Wherein the generation unit generates the three-dimensional feature vector of the pixel in the three-dimensional structure.

Images