TWI419058B - Image recognition model and the image recognition method using the image recognition model - Google Patents

Description

Method for establishing image recognition model and image recognition method using the image recognition model

The invention relates to image recognition, in particular to the establishment and application of an image recognition model.

According to the face recognition system, the biometric feature recognition technology can use statistical methods to construct the facial feature model for the comparison of the face images to be tested. The overall face recognition technology is divided into two stages: training. In the stage and test phase, a set of face feature models is used in the training phase to distinguish and characterize different face images in the feature space of the model. Therefore, in the training phase, the face feature model used affects the face. The accuracy of the identification.

In view of the prior art efforts to find the most perfect face recognition model, according to a paper published by He et al. in IEEE Trans. On Pattern Analysis and Machine Intelligence, vol. 27, pp. 328-340, 2005. The article "face recognition using laplacianfaces" proposes a face recognition system based on the feature space method to establish a Laplacian matrix to preserve the local structure of training samples. The method is to preserve the locality preserving projection (LPP) of the flow structure in the face feature space in the nearest-neighbor graph. However, according to Fig. 1, it is a method of transforming the feature space of the training sample by the method proposed by He et al. As shown in the figure, the prior art is still subject to external disturbance factors: expression, light and angle. The result of poor facial image discrimination is not good. However, the locally popular structure produced by the method proposed by He et al. is more efficient than other prior art techniques using principal component analysis (PCA) or linear discriminant analysis (LDA) of the euclidean structure.

In the prior art, many algorithms use classification tags to perform discriminant analysis procedures, using Fisher criterion to optimize intra-group tightness and inter-group separability, and calculate the modified based on the region preservation projection method. Intra- and inter-group variation to preserve local structure, and to describe samples using orthogonal neighborhood preserving discriminant analysis (ONPDA) or marginal Fisher analysis (MFA) A similarity matrix between adjacent relationships. Here are three representative papers: 1. Cai et al. published "Othogonal Laplacianfaces for Face Recognition" in the same journal in 2006. The spirit of the Laplacian matrix continues. The difference is also the most important. Cai converts the feature space into an orthogonal base. 2.Yan published "Graph embedding and extensions: General framework for dimensionality reduction" in the same journal in 2007. This paper is also a continuation of Laplace's method. The difference is that the concept of category information is added, and When calculating different categories of information, only the information of the nearest neighbor is considered, not all points. 3. Yan et al. proposed a general architecture for chart embedding, which uses two graphs, one is an intrinsic graph and the other is a penalty graph to describe the adjacent points in the group. The adjacent relationship of the border points between the groups, so these approximation methods can retain the structure within the local group and the adjacent groups.

In another stage of the face recognition-testing phase, the image is converted into the constructed feature conversion space, and compared with the training sample of the formed feature conversion space. The approximation method in the prior art focuses on Distance measurement and comparison algorithm design, in which the nearest feature line (NFL) method was first proposed by SZLi and J.Lu in the IEEE Transactions on Pattern Analysis and Machine Intelligence papers, while SZLi and J. Lu The two publications are "Face recognition using the nearest feature line method" and "Performance evaluation of the nearest feature line method in the image classification and compensation" (Performance evaluation of the nearest feature line method in "Image classification and retrieval"" refers to the feature line generated from two feature points and the feature line can linearly insert or extrapolate each pair of feature points in the same group, each group has an infinite number of pseudo-prototypes (pseudo prototypes) ) is generated by linear interpolation, and the classification work is completed by selected input images and features The shortest distance between the two approaches a linear characteristic line is human facial image variation amount.

Furthermore, according to the "Discriminant wavelet faces and nearest feature classifiers for face recognition" published by Chien et al., IEEE Transactions on Pattern Analysis and Machine Intelligence vol. 24, pp. 1644-1649, 2002. It extends the technology of the recent NFL-based classifier to test comparison of face images in the manner of nearest feature space.

Compared with the traditional nearest neighbor classifier (NN-based classifier), the feature line classifier is much better than the traditional point-to-point comparison method. However, recently the feature line classifier In the comparison, since the sample pairs combined in a straight line are selected, when the sample size is large, a large amount of comparison time is required, which is a disadvantage.

In addition, face recognition technology can use the characteristics of the image itself to design the classifier. Its technical features are usually simple, such as color information. However, such line features are also subject to large changes in noise such as light changes. The ability to classify is not dominant, but it is also limited in practice.

In view of the above, the present invention provides an image recognition model establishing method and an image recognition method using the image recognition model, which directly integrates the method of the feature line classifier into the Laplacian matrix in the training stage of the previous stage. In order to improve the accuracy of image recognition and speed up the recognition.

The main object of the present invention is to provide a method for establishing an image recognition model and an image recognition method using the image recognition model, which is to integrate a concept embedded in the feature space into a Laplacian matrix to establish a new feature space transformation matrix. Not only the topology data of the training samples in the original space can be retained, but also the linear relationship of the training samples can be saved, and the face image recognition can be performed in the test phase by the fast comparison method.

The second object of the present invention is to provide a method for establishing an image recognition model and an image recognition method using the image recognition model. The mathematical model established by the present invention is not limited to the recent feature line method, and is also compatible with various spatial classifications. For example, the nearest feature point, the nearest feature surface, and the nearest high-dimensional feature space classifier to create more mathematical models with recognition to improve the accuracy of face recognition.

A further object of the present invention is to provide a method for establishing an image recognition model and an image recognition method using the image recognition model. In the training phase, a nearest feature point classifier can be used for face recognition to improve recognition efficiency.

In order to achieve the above object, the method for establishing an image recognition model disclosed by the present invention and the image recognition method using the image recognition model firstly collect a plurality of training samples of different groups and determine the closest distance from the point to the feature subspace. The method performs spatial transformation; the continuation, the main transformation matrix of the training samples is calculated by principal component analysis method, and the training samples are projected to the dimensionality reduction space; then the group of all samples is calculated by the determined distance from the point to the closest distance of the feature subspace. The inner projection point and the inter-group projection point and the intra-group vector and inter-group vector formed by the projection points, and the vectors are sequentially arranged; and then, according to the intra-group vector and the inter-group vector, the intra-group is calculated. Mutation matrix and inter-group variation matrix; the variation transformation matrix is obtained from the intra-group variation matrix and the inter-group variation matrix; then the main transformation matrix inner product variation transformation matrix can be used to obtain the feature space transformation matrix.

Therefore, the comparison between the test image and the training sample can be performed in the feature space of the feature space conversion matrix to identify the group to which the test image belongs.

The purpose, technical contents, features and effects achieved by the present invention will be more readily understood by the detailed description of the embodiments and the accompanying drawings.

At present, the common face recognition methods focus on three major categories: feature space conversion method, image processing and algorithm. However, these methods often fail to preserve the characteristics of the image in the original space after dimensionality reduction, or are easily interfered by external factors. For example, light and shadow changes make the efficiency of classification greatly reduced. Therefore, the present invention preliminarily processes images by principal component analysis in the training phase, and then establishes a set of feature space transformation matrices by nearest feature space embedding, that is, recent feature space characteristics. It is integrated into the Laplacian matrix and improved on the feature space transformation method. Therefore, the technique of feature space transformation and the nearest feature space classifier is integrated, which is determined by point-to-line, point-to-surface, or point-to-high-dimensional space. Feature space conversion matrix.

Referring to FIG. 2, as shown in step S10, first, a plurality of face images of different groups are collected as training samples for establishing an image recognition model, and each training sample data is normalized to the same size. Window image and align each face image. At the same time, the p-value is determined, and the p-value represents the form of the feature subspace. For example, p=1 represents the feature subspace system as a point, p=2 represents a feature subspace system as a line, and so on. Therefore, according to the p value, we can know that the strategy of the feature transformation space is the shortest distance method from point to point, point to line, point to plane or point to high dimensional space, and the feature subspace represents point, line, plane or high dimensional space.

Next, as shown in step S12, two values of K ₁ and K _{2 are} determined, which are representative of the number of vectors in the group and the number of vectors between groups in each training sample when performing Fisher's Criterion calculation. Then, in step S14, the main transformation matrix w _{PCA of} the collected training samples is established by principal component analysis, and the training samples are projected to the dimensionality reduction space. Then, the projection points of each sample point in the same group and the projection points of each sample point between the clusters can be calculated. As shown in step S16, the point-to-point, point-to-line point-to-surface and point-to-point are obtained according to the p-value of step S10. A key vector that points to a high-dimensional space, where the key vector contains the intra-group vector formed by the intra-group projection points, and the inter-group vector formed by the inter-group projection points, and the intra-group vector and the inter-group vector are small and large according to the distance. Arranged in a way.

As shown in step S18 in FIG. 2, from the aligned intra-group vectors, the first K ₁ intra-group vectors are selected, and the K ₁ value has been determined in step S10, and the corresponding two weight matrices W and M are set. To calculate the within-class scatter S _W , as shown in Figure 3(a), and also from the inter-group vector, select the first K ₂ inter-group vectors and set the corresponding two weights. The matrices W and M are used to calculate the between-class scatter S _B as shown in Fig. 3(b).

Proceeding to the next step S20, the Fishery's Criterion transformation is used to obtain the variation transformation matrix w ^* obtained by the intra-group variation matrix S _B and the inter-group variation matrix S _W , and the relationship is The feature vector with the largest eigenvalue is selected to form the mutation transformation matrix. Finally, after the main transformation matrix w _PCA and the mutation transformation matrix w ^{* are} inner product, the feature space transformation matrix w is obtained. As shown in step S22, the relationship is w = w _PCA w ^* .

Step S18 may adopt another method for screening the intra-group vector and the inter-group vector, first setting the first feature threshold, and then selecting the intra-group vector of the group lower than the first feature threshold, and setting the corresponding two The weight matrix W and M are calculated to obtain the intra-group variation matrix S _W . And, the second feature threshold is also set, and the inter-group vector smaller than the second feature threshold in the different types of groups is selected, and the corresponding two weight matrices W and M are set to obtain the inter-group variation matrix S _B . If the calculation is performed in the manner of this paragraph in step S18, it is not necessary to consider the number of K ₁ and K ₂ in advance, so step S12 may be omitted, and step S14 may be directly performed from step S10, and the remaining processes are unchanged.

The feature space transformation matrix established by the invention is a feature space obtained by embedding a Laplacian matrix strategy in the nearest feature space, and the feature space transformation matrix constructed by the invention is used as an image recognition model, which not only retains training. The original topology information of the sample, and at the same time, can show the linear change between the training samples and the group information, as shown in Fig. 4, which is the training sample distribution map of the feature space obtained by the method of the present invention. Comparing Fig. 4 with Fig. 1, it can be seen that the image recognition model established by the present invention effectively classifies different types of training samples, and the recognition ability is better than the prior art.

The present invention also proposes an image recognition device using an image recognition model. As shown in FIG. 5, the device obtains a test image 12 from the image capture device 10, and then uses a image simulation processor 14 to test the image. 12 is projected in the image recognition model 16 established by the present invention, that is, in the feature space of the feature space conversion matrix, since the feature space is constructed according to a set of different types of training samples 18, the image simulation processor 14 The test image 12 will be compared with the training sample 18 by using the point-to-point, point-to-point, point-to-point or point-to-high-dimensional space closest to each other, so the calculator 10 calculates the group to which the test image 12 belongs, and finally The identification result of the test image is output through a display 20. If there is a time to shorten the test phase, that is, the face recognition time, the test image can be compared with the training sample by selecting the closest point-to-point distance in the feature space, thereby saving the time for face recognition.

Further, the feature space conversion matrix construction process of the present invention first considers a specific training sample point y _i in the feature space, and the distance from the point to the feature space can be defined as ∥ y _i - f ^{( P )} ( y _i )∥, where f ^(P) is a feature subspace composed of p samples, for example, when p=2, f ⁽²⁾ is a line, as shown in Fig. 6, when p=3, f ^{(3) )} , as shown in Figure 7. And f ^{( P )} ( y _i ) represents the projection point of y _i on the feature subspace, then there is a total for y _i Pairing can be projected. The vector formed by the projection onto the feature subspace can be used to calculate the scatter matrix. This method is called neaping feature space embedding. We can establish two kinds of target functions as follows. Equation (1) and formula (2):

Wherein, the above-mentioned binary weight value w ^{( P )} ( y _i ) can be used to represent the connection state of the projection point on the point-to-feature subspace, w ^{( P )} ( y _i ) is 1 or 0, and 1 represents the projection vector. Incorporate the calculation, 0 means that the projection vector is not included in the calculation. The equation (1) represents that all the points are squared with the vector formed by the corresponding projection points, and the equation (2) represents that the squares of all the points formed by their corresponding projection points are added. For example, when p=1, (1) is the local linear embedding (LLE) target function, and (2) is the laplacian matrix preserving projection (LPP). The objective function. To simplify the representation, the weight value w ^{( P )} ( y _i ) can be further represented by a matrix of N × N ^p , and at p = 2, y _i can be found Point-to-line projection vector with weight versus Is not the same, and It does not exist.

Continued, showing how to construct a Laplacian matrix algorithm from point-to-feature subspace:

1. Nearly feature point embedding, p=1

Under this condition, f ⁽¹⁾ converges to a point, if the selection of the K nearest neighbors, the formula (1) is compared with _a standard F. LLE functions: Since the LLE function is equal to the LPP function, Equation (1) can be expressed as w ^T XLX ^T w , and w ^{( P )} ( y _i ) in Equation (2) can be regarded as a similarity function of LPP, ie And the formula (2) can also be expressed as w ^T XLX ^T w .

2, nearest feature line embedding, p=2

As shown in Figure 6, based on the point-to-line distance, y _i to the feature line Distance can Representation, where the projection point It can be represented by a linear combination of y _m and y _n : Where the parameter t _m,n =( y _i - y _m ) ^T ( y _m - y _n )/( y _m - y _n ) ^T ( y _m - y _n ), and y _i to The resulting vector can be expressed by the following formula (3):

And 1- t _m,n = t _n,m , the objective function F ₁ can be expressed as shown in equation (4):

The W and M systems in equation (4) are:

Where t _m,n =( x _i - x _m ) ^T ( x _m - x _n )/( x _m - x _n ) ^T ( x _m - x _n ), Σ _j M _i,j =1, W _{i , j} = ( M + M ^T - M ^T M ) _{i , j} , then the formula (1) can be expressed in the form of a Laplacian matrix. In equation (2), K matrices can be decomposed first, and each M _{i , j} (1) represents the point X _i and its nearest neighbor feature , i, m, n can be a natural number from 1 to N. Two non-zero terms M _{i , n} (1)= t _{m , n} , M _{i , m} (1)= t _{n , m} is the value in M _{i , j} (1), and thus, equation (2) can Expressed as the following formula (7):

Where M _i,n ( k )= t _m,n , M _i,m ( k )= t _n,m , i ≠ m ≠ n , W _i,j ( k )=( M ( k )+ M ( k ) ^T + M ( k ) ^T M ( k )) _i,j , Therefore, the formula (2) can also be expressed in the form of a Laplacian matrix, so that it is only necessary to fill in the value of the corresponding matrix M content with the correct corresponding value.

3, recent feature plane / high-dimensional space embedding (nearest feature plane / space embedding), p = 3

under this condition, Pointing point y _i to the projection point on the feature plane E _{q , m , n} a vector formed, wherein the plane E _{q , m , n} is a plane formed by three feature points y _q , y _m and y _n as shown in Fig. 7, and the projection point on the plane It can be obtained by the formula (8), where the formula (8) is as follows:

Where Y _{q , m , n} =[( y _m - y _q )( y _n - y _q )] is a matrix of size d × 2, and three feature points y _q , y _m and y _{n are known} , so get

And you can put the projection point ( y _i ) is represented by a linear combination of three feature points y _q , y _m and y _n :

Where t _q + t _m + t _n =1. Similar to the case of p=2, the point y _i can be added to the vector adjacent to the K feature planes. Representation, and the weight value in the M matrix is [ t _q t _m t _n ] ^{T in} the above formula, and , t _q =1 - t _m - t _n , X _{q , m , n} = [( X _m - x _q )( x _n - x _q )]. When K is selected and the weight value in the M matrix is also given, F ₁ in the formula (1) can be expressed in the form of a Laplacian matrix: w ^T XLX ^T w , and F ₂ in the same formula (1) is also Can be expressed as w ^T XLX ^T w .

Furthermore, when P>3, the projection point of the feature point y _i onto the subspace substrate can be expressed as:

Wherein Y = [ y ₂ - y ₁ , y ₃ - y ₁ , ..., y _P - y ₁ ] T is a matrix of size d × ( P -1), and . The vector formed by the feature point y _i and the projection point can be expressed as: y _i - f ^{( P )} ( y _i ) = y _{i - Σ j} M _{i , j} y _j , therefore, the last F ₁ and F ₂ can be derived In the form of a Laplacian matrix.

According to the above-mentioned Laplacian matrix, the nearest variation space condition can be obtained, and the variation matrix of the difference information can be further obtained. Finally, the feature space transformation matrix can be obtained, and the different types of images can be segmented, as shown in FIG. The samples of each category also retain the topology information and linear relationship of the training samples in the original space. Therefore, we only need to use the feature space corresponding to the feature space transformation matrix, and the comparison with the nearest neighbor point method has a fairly good recognition ability, which can greatly save the time required for the conventional comparison.

Please refer to Figures 8-11, which are the results of the face recognition accuracy of the image recognition model of the present invention in a low dimensional space, compared with four prior art techniques, as shown in the prior art. They are: PCA+MFA stands for Principal Component Analysis with Border Fisher Analysis; PCA+ONPDA stands for Principal Component Analysis with Orthogonal Base Proximity Discriminant Analysis; PCA+LPPface stands for Principal Component Analysis plus Region Preservation Projection; and PCA+OLPPface stands for The principal component analysis method is used in conjunction with the orthogonal base region to preserve the projection method.

Moreover, the image recognition model of the present invention sets the p-value to 2 (in point-to-line closest distance mode), and the K ₁ value and the K ₂ value are all 10 to establish a feature conversion space matrix. Among them, Figures 8-9 show the results obtained from the training samples provided by the CMU database (Figure 8 shows 6 training samples for each group, and Figure 9 shows 9 training samples for each group), 10-11 The plan is the result of the training samples provided by the IIS database (Figure 10 shows 6 training samples for each group and Figure 11 shows 7 training samples for each group).

According to the test result of the present invention, it can be seen that the method for establishing the image recognition model of the present invention has excellent recognition. Since the present invention changes the concept of the nearest feature space embedded in the test phase of the face recognition, the recent feature is Spatial embedding is integrated into the feature space transformation matrix constructed in the training phase, and the invention can choose to construct the feature space transformation by using points to different feature subspaces (such as: point, line, plane or high dimensional space). And it is expressed in the Laplacian matrix, so not only the topology information in the original space can be retained, but also the linear change relationship of the training samples in the original space can be further preserved, and the recognition degree can be improved. Moreover, in the test phase, only a simple point-to-point classifier is used to identify the face to be tested, thereby reducing the time it takes to identify the face generally during the test phase.

The embodiments described above are merely illustrative of the technical spirit and the features of the present invention, and the objects of the present invention can be understood by those skilled in the art, and the scope of the present invention cannot be limited thereto. That is, the equivalent variations or modifications made by the spirit of the present invention should still be included in the scope of the present invention.

10. . . Image capture device

12. . . Test image

14. . . Image simulation processor

16. . . Image recognition model

18. . . Training samples

20. . . monitor

FIG. 1 is a schematic diagram showing a feature space conversion distribution of a training sample of the prior art.

FIG. 2 is a flow chart showing the steps of the method for establishing the image recognition model of the present invention.

Figure 3(a) is a schematic diagram of intra-group variation.

Figure 3(b) is a schematic diagram of variation between groups.

Figure 4 is a schematic diagram showing the feature space conversion distribution of the training samples of the present invention.

Figure 5 is an image recognition device of the present invention.

Figure 6 is a schematic diagram of the feature subspace from point to p=2.

Figure 7 is a schematic diagram of the feature subspace from point to p=3.

Figures 8-11 are diagrams showing the results of the face recognition accuracy of the present invention and the prior art, respectively.

Claims (18)

A method for establishing an image recognition model includes: (a) collecting a plurality of training samples of a plurality of clusters, and determining a shortest distance from the feature subspace to perform feature space conversion; (b) establishing the training samples a primary transformation matrix, and projecting the training samples into the dimensionality reduction space; (c) calculating a plurality of projection points and complex numbers of the training samples according to the determined shortest distance from the point to the feature subspace Projective points between groups, obtain a plurality of intra-group vectors and a plurality of inter-group vectors; (d) use an equation for calculating intra-group variation matrices and inter-group variation matrices according to the intra-group vectors and the inter-group vectors Calculating a set of internal variation matrices and an inter-group variation matrix, the calculation of the intra-group variation matrices and the inter-group variation matrix system can be expressed in the form of a Laplacian matrix, wherein the equation is as follows: Where y _i is the training sample point; f ^{( P )} ( y _i ) is the projection point of y _i on the feature subspace; w ^{( P )} ( y _i ) is 1 or 0, 1 indicates the projection vector Including calculation; (e) obtaining a mutation transformation matrix according to the intra-group variation matrix and the inter-group variation matrix; and (f) obtaining a feature space transformation matrix according to the main transformation matrix and the mutation transformation matrix The image recognition model is used. The method for establishing an image recognition model according to claim 1, wherein the training samples or the test images are facial images. The method for establishing an image recognition model according to claim 1, wherein the step (c) Calculating the projection point between the projection points of the training samples in the group and the groups in the different groups according to the shortest distance from the point to the feature subspace, and obtaining the A plurality of intra-group vectors formed by projection points in the group and a plurality of inter-group vectors formed by the projection points between the groups. The method for establishing an image recognition model according to claim 1, wherein the step (a) further comprises a step of: determining a number of vectors K ₁ in the group and a number K ₂ between groups. The method for establishing an image recognition model according to claim 4, wherein the step (c) further comprises: arranging the intra-group vectors and the inter-group projection vectors in a large and small manner according to the distance. The method for establishing an image recognition model according to claim 5, wherein in the step (d), in the group of vectors in which the distance is arranged from large to small, the first K ₁ inner vector of the group is selected. And setting corresponding at least one weight matrix to calculate the intra-group variation matrix. The method for establishing an image recognition model according to claim 5, wherein in the step (d), among the inter-group vectors whose distances are arranged from large to small, the first K ₂ inter-group vectors are selected. And corresponding at least one weight matrix is set to calculate the variation matrix between the groups. The method for establishing an image recognition model according to claim 1, wherein in the step (d), selecting the intra-group vector of the first feature threshold value of the vector group in the same group, and setting the corresponding at least one weight A matrix to calculate the variation matrix within the group. The method for establishing an image recognition model according to claim 1, wherein in the step (d), the inter-group vectors in which the vector distances in the different groups are smaller than the second characteristic threshold are selected, and the corresponding at least A weight matrix is used to calculate the variation matrix between groups. The method for establishing an image recognition model according to claim 1, wherein the feature subspace in the step (a) comprises a point, a line, a surface or a high dimensional space. The method for establishing an image recognition model according to claim 1, wherein the step (e) is performed by calculating the Fisher's Criterion conversion between the obtained intra-group variation matrix and the inter-group variation matrix. The mutation transformation matrix. The method for establishing an image recognition model according to claim 1, wherein the mutation transformation matrix w ^* in the step (e) is obtained by the intra-group variation matrix S _B and the inter-group variation matrix S _w , Its relationship is . The method for establishing an image recognition model according to claim 1, wherein the feature space conversion matrix in the step (f) is an inner variation matrix of the main conversion matrix. The method for establishing an image recognition model according to claim 1, wherein the main conversion matrix is obtained by principal component analysis. The method for establishing an image recognition model according to claim 1, wherein the image recognition model provides at least one test image projection, and identifies the test image according to the closest distance of the test image from the training sample. Class group. An image recognition method using the image recognition model of claim 1 is to project the training samples and at least one test image into the image recognition model, and compare the test images with the training in the nearest distance manner. Sample to correctly identify the group to which the test image belongs. For example, in the image recognition method described in claim 16, wherein the system uses the shortest distance method of the point-to-feature subspace to correctly identify the group of the test images. The image recognition method according to claim 16, wherein the feature subspace is a point, a line, a surface or a high dimensional space.