KR101847175B1 - Method for object recognition and apparatus thereof - Google Patents

Abstract

An object recognition method using local significance pattern cluster code and an object recognition apparatus using the same are disclosed. An object recognition apparatus according to the present invention includes: a pattern coding unit for generating a lookup table for converting N local patterns (LTP) extracted from an arbitrary learning image into M codes; And a learning unit for generating a classifier for classifying a class of a feature vector by converting a local pattern extracted from an arbitrary learning image into a predetermined code using the lookup table and inputting a feature vector corresponding to the converted code, .

Description

METHOD FOR OBJECT RECOGNITION AND APPARATUS THEREOF FIELD OF THE INVENTION [0001]

The present invention relates to an object recognition method and an object recognition apparatus using the same, and more particularly, to an object recognition method using a local significance pattern cluster code and an object recognition apparatus using the same.

Recently, in the field of computer vision, many researches have been conducted to recognize and detect various objects such as a vehicle, a pedestrian, and a face in order to be applied to the fields of intelligent automobiles, robots, surveillance cameras, and games.

In order to detect a specific object in an image, as shown in FIG. 1, a process of recognizing an area corresponding to the object while searching the entire image using a previously learned model is performed. A general object recognition method as shown in FIG. 1 is divided into a learning step and a test step.

In the learning phase, feature vectors are extracted for positive learning images and negative learning images, and the parameter of the classifier that maximizes the recognition rate is estimated.

In the test step, a feature vector is extracted for an arbitrary test image, and a positive image or a negative image is determined according to the output value of the generated classifier.

For robust object recognition, feature vectors with superior discrimination are extracted from training images, and machine learning such as support vector machine (SVM), boosting, and random forest is used , It is essential to learn the optimal classifier.

In the conventional object recognition techniques, gradient-based feature vectors and texture-based feature vectors are widely used. The Histogram of Oriented Gradient (HOG) is a representative gradient-based feature vector, which extracts feature vectors by accumulating directional components of a gradient in a cell to provide robust object recognition performance under various external environmental conditions. However, since a large amount of computation is required to extract the gradient magnitude and the direction component, there is a problem that is difficult to apply to a real-time object detection system.

The local binary pattern (LBP), which is a typical feature extraction method based on texture, is capable of extracting feature vectors by a simple arithmetic processing and has advantages of being robust against illumination change. However, there is a disadvantage that recognition performance is somewhat lower than that of HOG due to weakness of image noise. Although the local ternary pattern (LTP) method has been proposed to overcome these disadvantages, the problem that the dimensionality of the feature vector dimension is increased when the block histogram is generated by the sudden increase in the number of patterns Occurs.

SUMMARY OF THE INVENTION The present invention has been made in order to solve the problems of the conventional art described above by extracting a texture-based feature vector capable of improving object recognition performance and operation speed even under various external environmental conditions, : LIPCC) and an object recognition apparatus using the same.

The objects of the present invention are not limited to the above-mentioned objects, and other objects not mentioned can be clearly understood by those skilled in the art from the following description.

According to an aspect of the present invention, there is provided an object recognition apparatus including: a lookup table for converting N local patterns (LTP) extracted from an arbitrary learning image into M codes; A pattern coding unit for generating a pattern; And a learning unit for generating a classifier for classifying a class of a feature vector by converting a local pattern extracted from an arbitrary learning image into a predetermined code using the lookup table and inputting a feature vector corresponding to the converted code, .

Wherein the pattern coding unit converts a positive learning image to be a target of object recognition and a negative learning image to be a control group for the positive learning image into an LTP (Local Ternary Pattern) image, divides the LTP image into n blocks, and n And generates an LTP block histogram for each block.

The pattern coding unit cumulatively averages the LTP block histograms for each of the n blocks and calculates the importance of the k-th local pattern for the j-th block to determine M-1 most significant local patterns.

In addition, the pattern coding unit may assign a local significance pattern (LIPCC) code to each of the M-1 highest significant local patterns sequentially, The LIPCC corresponding to the remaining local patterns is allocated based on the distance between the most significant local code and the remaining local patterns.

In addition, the pattern encoding unit allocates LIPCCs for the highest significant local pattern at a distance closest to the remaining local patterns.

The M-th LIPCC is allocated to the remaining local patterns in which the distance between the M-1 most significant local patterns and the remaining local patterns is equal to or greater than a preset threshold value.

The learning unit compares N LTPs extracted from the positive learning image, which is a target of object recognition, and the negative learning image, which is a control group of the positive learning image, with a LUT (Local Important Pattern Clustering Code) And the feature vector is learned by using the LIPCC block histogram generated for each image block.

According to an aspect of the present invention, there is provided an object recognition apparatus for generating an LIPCC block histogram for each image block using the lookup table in an input test image, and the feature vector corresponding to the LIPCC block histogram using the classifier is a positive An object recognition unit for determining whether the feature vector of the image is a feature vector or a feature vector of the negative image.

According to another aspect of the present invention, there is provided a method of recognizing an object, comprising the steps of: (a) extracting N local patterns extracted from a negative learning image, which is a control group for the positive learning image, LTPCC look-up table for converting LTPCC to L local critical pattern clustering codes (LIPCC); (b) converting the N LTPs extracted from the positive learning image and the negative learning image into M LIPCCs using the LIPCC lookup-table; And (c) generating a LIPCC block histogram for each of the image blocks using the M LIPCCs and generating a classifier for learning a feature vector corresponding to the LIPCC block histogram.

The step (a) includes the steps of: converting the positive learning image and the negative learning image into LTP images, dividing the LTP image into n blocks, generating an LTP block histogram for each of the n blocks Calculating an M-1 highest significant local pattern by cumulatively averaging the LTP block histograms for each of the n blocks to calculate a significance of a k-th local pattern for a j-th block, (LIPCC) is assigned to each of the most significant local patterns of the M-1 highest significant local patterns, and the distance between the M-1 most significant local patterns and the remaining local patterns calculated based on the clustering technique And allocating the LIPCC corresponding to the remaining local patterns.

The step (a) further includes the steps of: allocating a LIPCC for a most significant local pattern at a distance closest to the remaining local patterns; and determining a distance between the M-1 most significant local patterns and the remaining local patterns, The M-th LIPCC is allocated to the remaining local patterns that are equal to or greater than the predetermined threshold value T.

According to another aspect of the present invention, there is provided a method for recognizing an object, comprising the steps of: (d) generating a LIPCC block histogram for each image block using the LIPCC lookup table in an input test image; And determining whether the feature vector is a feature vector of the positive image or a feature vector of the negative image.

As described above, according to the present invention, it is possible to effectively overcome the problem of texture-based feature vectors, in particular, the problem of increasing the number of dimensions of LTP feature vectors, and to provide recognition performance superior to gradient-based feature vectors, Can be improved at the same time.

1 illustrates an object recognition method of an object recognition system according to the prior art.
2 is a block diagram illustrating an object recognition apparatus according to an embodiment of the present invention.
FIG. 3 is a flowchart illustrating a process of generating a local significance pattern grouping code lookup table in the pattern encoding unit of FIG. 2;
FIG. 4 is a diagram illustrating a process of generating the local significance pattern grouping code lookup table in the pattern encoding unit of FIG. 2;
FIG. 5 is a diagram illustrating a process of generating a classifier for classifying two classes from the learning unit image of FIG. 2;
FIG. 6 is an exemplary diagram illustrating a process of recognizing an object from a test image of the object recognition unit of FIG. 2;

BRIEF DESCRIPTION OF THE DRAWINGS The advantages and features of the present invention, and the manner of achieving them, will be apparent from and elucidated with reference to the embodiments described hereinafter in conjunction with the accompanying drawings. The present invention may, however, be embodied in many different forms and should not be construed as being limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art. Is provided to fully convey the scope of the invention to those skilled in the art, and the invention is only defined by the scope of the claims. It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. In the present specification, the singular form includes plural forms unless otherwise specified in the specification.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. In the drawings, the same reference numerals are used to designate the same or similar components throughout the drawings. In the following description of the present invention, a detailed description of known functions and configurations incorporated herein will be omitted when it may make the subject matter of the present invention rather unclear.

2 is a block diagram illustrating an object recognition apparatus according to an embodiment of the present invention.

Referring to FIG. 2, an object recognition apparatus according to the present invention includes a pattern coding unit 100, a learning unit 200, and an object recognition unit 300.

The pattern coding unit 100 generates a look-up table for converting N local patterns (LTP) extracted from an arbitrary learning image into M codes.

At this time, the pattern coding unit 100 can extract N local patterns based on the texture in the positive learning image, which is the target of object recognition, and the negative learning image, which is a control group for the positive learning image, and extracts the N local patterns (LIPCC lookup-table) for converting a Local Ternary Pattern (LTP) into M Local Important Pattern Clustering Codes (LIPCC).

FIGS. 3 and 4 illustrate a concrete procedure for the LIPCC lookup table generated by the pattern coding unit 100. FIG. Hereinafter, a process of generating the LIPCC lookup-table by the pattern encoding unit 100 (pattern encoding process) will be described with reference to FIGS. 3 and 4. FIG.

Pattern coding process

3 and 4, a positive learning image A and a negative learning image B are input to the pattern coding unit 100 (S110 and S210).

The pattern coding unit 100 converts the input positive learning image A and the input negative learning image B into LTP images A 'and B', and outputs the converted LTP images A 'and B' Are divided into n blocks (S120 and S220).

The pattern coding unit 100 generates an LTP block histogram for each of n blocks (S130 and S230). For example, as shown in FIG. 3, n blocks B1, B2, ... , Bn, an LTP block histogram for each of these blocks is generated.

Then, the pattern coding unit 100 cumulatively averages the LTP block histograms of the n blocks to calculate the importance of the kth local pattern for the jth block (S300).

라 한다면,

는 아래 수학식 1에서와 같이 연산될 수 있다. For example, if the importance of the k-th local pattern for the j-th block is

However,

Can be calculated as shown in Equation (1) below.

[Equation 1]

와

는 포지티브 및 네거티브 학습 영상의 개수를 나타내고,

와

는 i번째 포지티브 및 네거티브 학습 영상에서 j번째 블록에 대한 k번째 패턴의 누적된 개수를 나타낸다.

는 패턴의 전체 빈도수와 패턴 빈도수 차이간의 가중치를 결정하는 상수이다.here,

Wow

Represents the number of positive and negative learning images,

Wow

Represents the cumulative number of the kth pattern for the jth block in the i < th > positive and negative learning images.

Is a constant that determines the weight between the total frequency of the pattern and the pattern frequency difference.

When the importance of the local pattern is determined through the above-described process, the pattern coding unit 100 determines the highest significant local pattern according to the importance (S400). As described above, when it is desired to generate a look-up table for converting N local patterns into M codes, the pattern coding unit 100 determines M-1 most significant local significance patterns using the importance of local patterns .

In step S500, the pattern coding unit 100 assigns a code (e.g., a Local Important Pattern Clustering Code (LIPCC)) to each of the local patterns using a clustering technique.

First, the pattern coding unit 100 sequentially assigns a local significance pattern cluster code to each of the M-1 highest-order significant local patterns. For example, if the M local significance pattern cluster codes are 0, 1, ... , And M-1, the local significance pattern cluster codes of 0 to M-2 may be sequentially allocated to each of the M-1 most significant local patterns.

The pattern coding unit 100 allocates the local significance pattern cluster code to the remaining local patterns based on the distance calculated based on the clustering technique.

라 한다면, 나머지 국부 패턴들과 가장 가까운 거리에 있는 최상위 중요 국부 패턴에 대한 국부 중요 패턴 군집 코드가 해당 나머지 국부 패턴에 할당된다. For example, if the distance between the remaining local patterns and the most significant local patterns assigned to the existing code is

, A local significance pattern cluster code for the highest significant local pattern closest to the remaining local patterns is assigned to the corresponding remaining local pattern.

The pattern coding unit 100 applies the M-1 highest significant local stationary patterns to the remaining local patterns in which the distance between the M-1 most significant local patterns and the remaining local patterns is equal to or greater than a preset threshold T, And allocates a remaining code not assigned to the pattern, for example, code of M-1.

That is, patterns that are not important and whose degree of similarity between existing top patterns is low are assigned to one code. As a result, for all local patterns, a LIPCC look-up table C allocated with M codes is generated (S600).

Referring again to FIG. 2, the learning unit 200 estimates an optimal hyperplane capable of classifying the two classes using the feature vectors extracted from an arbitrary learning image.

That is, the learning unit 200 converts the local pattern extracted from an arbitrary learning image into a predetermined code using the LIPCC lookup-table generated in the pattern coding unit 100, and outputs a characteristic corresponding to the converted code And a classifier for classifying the class of the feature vector is generated.

Hereinafter, a process of generating a classifier for classifying two classes from the learning image by the learning unit 200 will be described with reference to FIG.

First, a positive learning image A and a negative learning image B are input to the learning unit 100. The input positive learning image A and the input negative learning image B are input to the LTP images A 'and B ').

Then, the learning unit 100 divides each of the converted LTP images A 'and B' into n blocks, and outputs the local pattern LTP extracted from each of the n blocks to the pattern coding unit 100 And converts it into M local significance pattern clustering codes (LIPCC) using the generated LIPCC lookup table.

Then, the learning unit 100 generates LIPCC block histograms for each of the n blocks, and concatenates the LIPCC block histograms sequentially, thereby extracting feature vectors for expressing the input learning image.

At this time, the feature vectors are used as inputs to the classifier, and the learning unit 100 determines the parameters of the classifier capable of classifying the two classes to the maximum. Here, the classifier can be support vector matching (SVM), Adaboost (adaptive boosting), multi-layer perceptron (MLP), and random forest. Equation (2) below is a formula for estimating a classifier having a maximum margin when a linear SVM is used.

&Quot; (2) "

는 분류기의 매개 변수를 나타내며,

는 i번째 샘플에 대한 클래스 라벨(label),

는 i번째 샘플에 대한 특징 벡터, b는 분류기의 바이어스(bias)를 나타낸다. here,

Represents the parameter of the classifier,

Is the class label for the i < th > sample,

Is the feature vector for the i-th sample, and b is the bias of the classifier.

In this way, a parameter capable of classifying the two classes to the maximum is determined, so that the classifier D in which the feature vector of the learning image is learned optimally is generated.

6, the object recognizer 300 converts the input test image T into an LTP image T 'and outputs LTP extracted from the converted LTP image T' to the LIPCC lookup table The LIPCC block histogram is generated for each of the image blocks and the feature vector corresponding to the LIPCC block histogram is determined as a feature vector of the positive image or a feature vector of the negative image using the classifier.

Meanwhile, the object recognition method according to the present invention described above can be implemented as a computer-readable code on a computer-readable recording medium. The computer-readable recording medium includes all kinds of recording media storing data that can be decoded by a computer system. For example, there may be a ROM (Read Only Memory), a RAM (Random Access Memory), a magnetic tape, a magnetic disk, a flash memory, an optical data storage device and the like. The computer-readable recording medium may also be distributed and executed in a computer system connected to a computer network and stored and executed as a code that can be read in a distributed manner.

 It will be understood by those skilled in the art that the present invention may be embodied in other specific forms without departing from the spirit or essential characteristics thereof. It is therefore to be understood that the above-described embodiments are illustrative in all aspects and not restrictive. The scope of the present invention is defined by the appended claims rather than the detailed description, and all changes or modifications derived from the scope of the claims and their equivalents should be construed as being included within the scope of the present invention.

100: pattern coding unit
200:
300: Object recognition unit

Claims (12)

A pattern coding unit for generating a lookup table for converting N local patterns (LTP) extracted from an arbitrary learning image into M codes; And
A learning unit for converting a local pattern extracted from an arbitrary learning image into a predetermined code using the look-up table, receiving a feature vector corresponding to the converted code, and generating a classifier for classifying the class of the feature vector Including,
Wherein the pattern encoding unit comprises:
A negative learning image as a target of object recognition and a negative learning image as a control group for the positive learning image is converted into an LTP (Local Ternary Pattern) image, the LTP image is divided into n blocks, Generates an LTP block histogram,
The LTP block histograms for each of the n blocks are cumulatively averaged to calculate the importance of the k-th local pattern for the j-th block to determine M-1 highest significant local patterns,
The importance of the local pattern is calculated based on the weight between the number of positive and negative learning images, the cumulative number of kth local patterns for the jth block in the positive and negative learning images, the total frequency of the local patterns, Lt; / RTI >
delete delete The apparatus of claim 1, wherein the pattern coding unit comprises:
(M-1) -th highest significant local patterns and (M-1) -th highest significant local patterns, which are calculated based on the clustering technique, And allocates an Mth LIPCC to the remaining local patterns based on the distance between the patterns.
The apparatus of claim 4, wherein the pattern coding unit comprises:
And assigns LIPCCs for the highest significant local pattern at a distance closest to the remaining local patterns.
The apparatus of claim 4, wherein the pattern coding unit comprises:
And a remaining local pattern in which the distance between the M-1 most significant local pattern and the remaining local pattern is equal to or greater than a preset threshold value T is allocated to the remaining local patterns not allocated to the M-1 most significant local pattern, Gt; LIPCC. ≪ / RTI >
The apparatus according to claim 1,
The N LTPs extracted from the negative learning image as a target of object recognition and the negative learning image as a control group for the positive learning image are classified into M LTPCCs using a LIPCC lookup table LIPCC, and the feature vector is learned using the LIPCC block histogram generated for each image block.
The method according to claim 1,
A LIPCC block histogram is generated for each of the image blocks using the lookup table in the input test image, and a feature vector corresponding to the LIPCC block histogram is determined as a feature vector of a positive image or a feature vector of a negative image And an object recognizing unit for recognizing the object.
(a) Local temporal patterns (LTP) extracted from a negative learning image, which is a target of object recognition, and a negative learning image, which is a control group for the positive learning image, are divided into M local significance patterns Clustering Code (LIPCC);
(b) converting the N LTPs extracted from the positive learning image and the negative learning image into M LIPCCs using the LIPCC lookup-table; And
(c) generating a LIPCC block histogram for each of the image blocks using the M LIPCCs, and generating a classifier for learning a feature vector corresponding to the LIPCC block histogram
The step (a)
A negative learning image, which is a control group for the positive learning image, into an LTP (Local Ternary Pattern) image, and dividing the LTP image into n blocks;
Generating an LTP block histogram for each of the n blocks;
Accumulating a LTP block histogram for each of the n blocks to calculate an importance of a k-th local pattern for a j-th block to determine M-1 highest significant local patterns;
The importance of the local pattern is calculated based on the weight between the number of positive and negative learning images, the cumulative number of kth local patterns for the jth block in the positive and negative learning images, the total frequency of the local patterns, Being
In object recognition method.
10. The method of claim 9, wherein the step (a)
(M-1) -th highest significant local patterns and (M-1) -th highest significant local patterns, which are calculated based on the clustering technique, And allocating a LIPCC corresponding to the remaining local patterns based on the distance between the patterns
In object recognition method.
11. The method of claim 10, wherein the step (a)
Assigning a LIPCC for a most significant local pattern at a distance closest to the remaining local patterns;
And a remaining local pattern in which the distance between the M-1 most significant local pattern and the remaining local pattern is equal to or greater than a preset threshold value T is allocated to the remaining local patterns not allocated to the M-1 most significant local pattern, Lt; RTI ID = 0.0 > LIPCC < / RTI >
In object recognition method.
10. The method of claim 9,
(d) generating a LIPCC block histogram for each of the image blocks using the LIPCC look-up table in the input test image, and using the classifier, determining whether the feature vector corresponding to the LIPCC block histogram is a feature vector of the positive image, Determining whether the feature vector is a feature vector
The object recognition method further comprising:

Images