Face recognition method based on multi-feature hybrid dictionary and probability prediction

Abstract

A face recognition method based on a multi-feature hybrid dictionary and a probability prediction mechanism includes: step 1, using a high-definition camera to collect face images under normal lighting conditions to ensure that the face is in the central position of the image; step 2, extracting local texture features and global features from the sampled data, and combining the two to construct a hybrid over-complete dictionary; step 3, to meet the N-K constraint condition of the over-complete dictionary, performing a PCA dimensionality reduction operation on the hybrid over-complete dictionary; step 4, using the constructed hybrid over-complete dictionary to perform sparsity decomposition on the test sample through the OMP algorithm to obtain a coefficient vector X; step 4, calculating the sample reconstructed by the sparse vector and the hybrid over-complete dictionary, comparing the original sample with the reconstructed sample, and calculating the error e; step 5, analyzing the coefficient distribution of each category, calculating the correlation score of the coefficient, and obtaining the prediction probability by using SoftMax for the correlation score; step 6, combining the error e and the prediction probability to obtain the final category.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence image processing technology, and more specifically to target recognition in the field of computer vision and sparsity representation methods in machine learning. Background Art

[0002] Face recognition technology is a type of human biometric recognition technology. It is an inherent attribute of human beings and has strong inherent stability and individual differences. Therefore, it is an ideal basis for identity verification. The face feature recognition system extracts biometric features and forms feature templates. When people use the recognition system for identity authentication, the recognition system extracts its features and compares them with the feature templates in the database to determine whether they match, thereby deciding whether to accept or reject the person.

[0003] Face recognition is the process of comparing the face images detected in static images or video images with the face images in the database to find the matching faces, so as to achieve the purpose of identity recognition and identification. It is a subject belonging to the fields of biometric recognition and artificial intelligence. Face recognition is one of the most successful applications of image analysis and understanding. Due to more than 30 years of research on the feasibility of face recognition technology and its wide application in many aspects such as business, security, identity authentication, and law enforcement, it has been increasingly valued. Today, the application background of face detection has far exceeded the scope of face recognition systems, and has important applications in content-based retrieval, digital video processing, visual detection and other directions.

[0004] The study of face recognition is of great significance both in theory and technology: first, it can advance the understanding of the human visual system itself; second, it can meet the needs of artificial intelligence applications. The use of face recognition technology, the establishment of an automatic face recognition system, and the use of computers to realize the automatic recognition of face images have broad application fields and attractive application prospects. At the same time, face recognition as a biometric recognition technology has the following three advantages compared with other recognition methods (such as fingerprints, irises, etc.). 1. Non-invasive, the acquisition of face images does not require physical contact with the person being detected, which is particularly important in the context of epidemic prevention and control; 2. Low cost and easy installation, the face recognition system only needs to use an ordinary camera, and there are no special installation requirements for users; 3. No human participation, the entire face recognition process does not require the user to actively participate, which has great research value in the field of security. Summary of the invention

[0005] The present invention overcomes the shortcomings of existing face recognition technologies, such as only using local features and ignoring global color features, and proposes a face recognition method based on a multi-feature hybrid dictionary and probability prediction.

[0006] The present invention provides a face recognition method based on global color features and local features, and the specific steps are as follows:

[0007] Step 1: Convert the captured image into a grayscale image and normalize it.

[0008] Step 2: Divide the original grayscale image into four parts, perform a separate singular value decomposition on each block, obtain the singular value feature vector of each block of the image, and then combine the singular value feature vectors of all local images with the singular value feature vector of the overall image to obtain a singular value feature vector that combines the local and the overall. The combination method is: arrange the singular value feature vectors of each local image in order from top to bottom and from left to right, and then add the singular value feature vector of the overall image to the end of the arranged vector, that is, to obtain a feature vector that combines the local and the overall.

[0009] Step 3: Use the Gabor filter group to convolve the input image with the original grayscale image, divide the filtered result image into equal and non-overlapping image sub-blocks, take the grayscale average value of each image sub-block, and connect the grayscale average values ​​of all image sub-blocks in series to form a feature vector representing the image.

[0010] Step 3: Fuse the feature map based on singular value decomposition and the Gist feature to construct a mixed feature column vector. Considering that when the feature dimension of the constructed atom increases, the number of atoms will not meet the constraint condition that is much larger than the atomic dimension, the PCA compression operation is performed on the constructed mixed feature here.

[0011] Step 4: After converting all images of the dataset into mixed feature column vectors according to step 3, divide the dataset into training set, validation set, and test set in a ratio of 7:2:1. Construct a mixed feature dictionary D using the mixed feature vectors generated by the images of the training set.

[0012] Step 5: Sparse representation studies how to reconstruct the signal in a concise way. For example, for a signal Y∈R m*1 , how to use X∈R n , Y=DX. This problem can be solved by solving L 0 The norm is minimized to solve it, that is,

[0013] (L 0 ): Y 0 =argmin||X|| 0 stY=DX (1)

[0014] Where Y represents the sample, D is the constructed hybrid dictionary, and X is the coefficient vector of Y represented by D. However, solving L 0The problem of minimizing the norm is an NP-hard problem. The norm minimization problem can be approximated by an iterative greedy pursuit algorithm. The latest theoretical research on sparse representation and compressed sensing shows that if the solution X is found 0 is sparse enough, then find L 0 The solution to the norm minimum problem can be transformed into finding the following L 1 Solution to the norm minimization problem:

[0015] (L 1 ): Y 0 =argmin||Y|| 1 stY=DX (2)

[0016] Step 6: From the hybrid dictionary matrix D, calculate the inner product of the signal y and each column (atom) in the dictionary matrix, and select the atom with the largest absolute value, which is the atom that best matches the signal y in this iterative operation, see formula (3).

[0017]

[0018] In this way, the signal y is decomposed into two parts: the vertical projection component at the most matching atom and the residual value, namely:

[0019]

[0020] To avoid the situation where the signal (residual) is non-orthogonal when projected vertically on the selected atoms, the result of each iteration is not optimal but suboptimal, and convergence requires many iterations. The residual needs to be regularized to satisfy the requirement that the residual is orthogonal to each previous component. Orthogonalization is performed on all selected atoms at each step of the decomposition, which can increase the convergence speed of the algorithm while maintaining the same accuracy requirements.

[0021]

[0022] In this way, an atom (that is, a column) that best matches the signal y is selected, a sparse approximation is constructed, and the signal residual is calculated. Then, the atom that best matches the signal residual is selected, and the signal y can be represented by the linear sum of these atoms and the final residual value. If the residual value is within a negligible range, the signal y is the linear combination of these atoms.

[0023] Step 7: Calculate the residual according to formula (6), where δ i (x ^ )∈R n The non-zero entries are the vectors of entries in x^ that are associated with category i.

[0024] min r i (y)=||y-Aδi (x^)|| 2 (i=1,2,...,k) (6)

[0025] Step 8. By analyzing the experimental results, correctly classified samples often have the characteristics of concentrated energy, and the error value of the category has a clear valley value compared with the error values ​​of other categories; while the wrongly classified samples have the characteristics of non-concentrated distribution and small variance of error values. This is because the weights are distributed on atoms of different categories without obvious peaks. In order to further improve the recognition accuracy, an additional probability prediction module is added in the classification stage. The association score of the weight vector is calculated by the category of the atom corresponding to the weight. The association score is converted into a predicted probability through the SoftMax function, and the combination of error and predicted probability is used as the basis for classification.

[0026] Step 9: According to step 8, the category with the smallest residual is selected as the recognition result.

[0027] In step 1, the specific method is: use a professional camera to collect a face image under normal lighting conditions to ensure that the face is in the center of the image.

[0028] In step 3, the Gabor filter is used, and its main advantage is in the extraction of object texture features. The Gabor filter is highly robust to changes in image brightness and contrast, as well as changes in image posture, and it expresses the most useful local features for image recognition, so it is widely used in the fields of computer vision and texture analysis.

[0029] The dictionary construction in step 4 must ensure that D N*K Among them, N must meet the requirement of being much smaller than K, that is, relative to the complete orthogonal basis, the basis of the over-complete basis must be redundant, so that D is complete enough to ensure that the space opened by all atoms is large enough to contain all training samples, and the redundant characteristics of the dictionary can better capture the essential characteristics of the samples.

[0030] The present invention has the following innovative features:

[0031] 1. The invention innovatively uses a combination of global features and hybrid features to construct hybrid features. Previous studies have shown that the singular value vector obtained by singular value decomposition of the entire face image cannot obtain sufficient information required for face recognition, and local features lack global color information. At the same time, hybrid features increase the atomic dimension because other features are introduced, and thus cannot meet the constraints of an over-complete dictionary in an over-sparse representation. To avoid this, the PCA dimensionality reduction module is introduced in the feature extraction stage.

[0032] The invention innovatively introduces a probability prediction mechanism in the sparsity representation stage, and calculates the correlation score of the coefficient distribution of each category by calculating the correlation of the distribution in the weight coefficient. The correlation score is used to obtain the probability of the sample in each category through softmax, and the error r is combined with the predicted probability to obtain the final category. The advantages of the invention are: by constructing a hybrid dictionary and introducing a probability prediction mechanism, the performance of face recognition in terms of adaptability and recognition accuracy is further improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Attached Figure 1 It is an implementation flow chart of the method of the present invention. DETAILED DESCRIPTION

[0034] The following is a more detailed description of the embodiments of the present invention. Figure 1 By referring to the attached Figure 1 The described embodiments are only used to illustrate the present invention and are not to be construed as limiting the present invention.

[0035] Reference Figure 1 The present invention provides a face recognition technology based on a multi-feature hybrid dictionary and probability prediction, and the steps are as follows:

[0036] Step 1. Use a professional camera to capture facial images under normal lighting conditions, convert the captured facial images into grayscale images, and normalize them to 32×16.

[0037] Step 2. Considering the high correlation between the RGB channels, the three color components are easily affected by factors such as light intensity and have poor stability; the grayscale image reflects an important piece of information, which is light intensity. From the perspective of the overall and local color and brightness level distribution characteristics of the entire image, the grayscale image description is consistent with the color image description, but the grayscale image can be used to obtain edges (gradient information), which is very important for edge detection (contour). Because it is more robust in the face of lighting changes and can better reflect the essential information of color, the RGB components are not used, and the global color features are extracted from the 8-bit grayscale image.

[0038] Gray=(R*299+G*587+B*144+500) / 1000 (7)

[0039] Step 3. Face recognition based on classic pattern recognition methods. Research shows that compared with global features, local features such as HOG, Gabor, and Gist can better overcome global interference problems such as uneven lighting and posture changes in natural environments, and have better recognition effects on faces. The local feature extraction method is performed on grayscale images and does not fully utilize the color information of the image. In practical applications, the image itself contains rich global information, and images belonging to the same category have great similarities in the distribution of brightness values. Therefore, based on the use of the color information of the image and combined with the local features of the image, the study is based on the construction of dictionaries and sparse representations based on a combination of multiple features to improve the recognition effect.

[0040] Step 4. Divide the original grayscale image into four parts and i (i∈1,2,3,4) Each block is subjected to separate singular value decomposition to obtain the singular value eigenvector of each block of the image, and then the singular value eigenvectors of all local images are combined with the singular value eigenvector of the overall image to obtain a singular value eigenvector that combines the local and the overall. The combination method is: the singular value eigenvectors of each local image are arranged in order from top to bottom and from left to right, and then the singular value eigenvector of the overall image is added to the end of the arranged vector, that is, the local and the overall eigenvector is obtained.

[0041] Step 5. Mixing features at the same time will increase the atomic dimension because of the introduction of other features, which cannot meet the constraints of the complete dictionary in the over-sparse representation. To avoid this, SVD is used to reduce the dimension of the mixed features in the feature extraction stage to achieve the purpose of reducing the dimension of the features.

[0042] Step 6. Use the Gabor filter group to convolve the input image with the original grayscale image, divide the filtered result image into equal and non-overlapping l×l image sub-blocks, and take the grayscale average value for each image sub-block. Connect the grayscale average values ​​of all image sub-blocks in series to form a feature vector representing the image. This method takes l=4 when calculating the Gist feature, and uses 32 Gabor filters in 8 directions and 4 scales to participate in the convolution operation. The feature map based on singular value decomposition and the Gist feature are fused to construct a mixed feature column vector. The specific combination method is to first convert the feature map based on singular value decomposition and the Gist feature map into column vectors, and then splice them into a mixed feature column vector in the manner of the feature map based on singular value decomposition on the top and the Gist feature map on the bottom.

[0043]

[0044] Step 7. Use the fusion features to construct an overcomplete dictionary based on mixed features for the training samples. Each column vector in the dictionary represents a training sample, and the samples of the same type are in the same subspace. Apply the overcomplete dictionary to perform multi-feature sparse representation on the training images. By solving L 1 The optimization problem in the sense of norm obtains a sparse solution, so that the coefficients of the training samples other than the category where the training sample belongs are all zero or close to zero. Using the OMP algorithm, in the atom selection stage, a set of indexes that best match the iteration amount is selected, and the selected atoms are orthogonalized to avoid repeated selection and improve the algorithm rate. Finally, the residual value is updated, and after iteration, the sparse decomposition of the original image and the minimized reconstruction error are obtained.

[0045]

[0046] Step 8. In L 1 Under the constraint of the paradigm, the coefficient vector α has energy concentration. Correctly classified samples often have the characteristics of energy concentration, and the error value of the category has a clear valley value compared with the error values ​​of other categories. The error value of the wrong category is 5 to 10 times that of the correct category; while the wrongly classified samples have the characteristics of non-concentrated distribution and small variance of error values. This is because the weights are distributed on atoms of different categories without obvious peaks. Therefore, in order to further improve the recognition accuracy, an additional probability prediction module is added to the classification stage. The association score of the weight vector is calculated by the category of the atom corresponding to the weight. The association score is converted into a predicted probability through the softmax function, and the error r is combined with the predicted probability to obtain the final category. Step 9. Solve and calculate the sparsity concentration index of the test image to determine whether it is a target unit. If the sparsity concentration index is greater than the threshold, the test sample is judged to be a recognition target, otherwise it is a non-recognition target. If it is a recognition target, the approximate residual between the test sample y and each type of sample is calculated, and the category corresponding to the minimum residual value is the final classification of the test sample.

Claims

1. A face recognition method based on a multi-feature hybrid dictionary and a probability prediction mechanism includes the following steps: Step 1: Convert the captured image into a grayscale image and normalize it; Step 2, the original grayscale image is divided into four parts, each block is subjected to a separate singular value decomposition to obtain a singular value eigenvector of each image, and then the singular value eigenvectors of all local images are combined with the singular value eigenvector of the overall image to obtain a singular value eigenvector combining the local and the overall; the combination method is: the singular value eigenvectors of each local image are arranged in order from top to bottom and from left to right, and then the singular value eigenvector of the overall image is added to the end of the arranged vector, that is, the eigenvector combining the local and the overall is obtained; Step 3: Use the Gabor filter group to convolve the input image with the original grayscale image, divide the filtered result image into equal non-overlapping image sub-blocks, and take the grayscale average value of each image sub-block, and connect the grayscale average values ​​of all image sub-blocks in series to form a feature vector representing the image; Step 3: The feature map based on singular value decomposition and the Gist feature are merged to construct a mixed feature column vector. Considering that when the feature dimension of the constructed atom increases, the number of atoms will not meet the constraint condition that is much larger than the atomic dimension, the PCA compression operation is performed on the constructed mixed feature here. Step 4: After converting all images of the dataset into mixed feature column vectors according to step 3, divide the dataset into training set, validation set and test set in a ratio of 7:2:1; construct a mixed feature dictionary D using the mixed feature vectors generated by the images of the training set; Step 5: Under the condition that the solution X0 is sufficiently sparse, transform the minimum value problem under the L0 norm into the solution of the following L1 norm minimum value problem: (L1): Y0=argmin ||Y||1 st Y=DX (1) From the hybrid dictionary matrix D, calculate the inner product of the signal y and each column of the atom in the dictionary matrix, and select the atom with the largest absolute value, which is the atom that best matches the signal y in this iteration operation, see formula (2); In this way, the signal y is decomposed into two parts: the vertical projection component at the most matching atom and the residual value, namely: Then, all the selected atoms are orthogonalized at each step of the decomposition, which can increase the convergence speed of the algorithm under the same accuracy requirements; In this way, an atom that best matches the signal y is selected, a sparse approximation is constructed, and the signal residual is calculated. Then, the atom that best matches the signal residual is selected, and the iteration is repeated. If the residual value is within a negligible range, the signal y is the linear combination of these atoms. Finally, the signal y is represented by the linear combination of these atoms plus the final residual value. Step 6: Calculate the residual according to formula (5), where δ i (x^)∈R n The non-zero entry is x ^ The vector of items in that are associated with category i; min. i (y)=||y-Aδ i (x^)||2,i=1,2,...,k (5) Step 7: Obtain the correlation score by analyzing the distribution of the statistical coefficients. According to the SoftMax function, convert the correlation score into the prediction probability, and use the combination of error and prediction probability as the basis for classification. Step 8: According to step 7, the category with the smallest residual is selected as the recognition result.

2. The face recognition method based on a multi-feature hybrid dictionary and a probability prediction mechanism as claimed in claim 1, characterized in that: Step 1 specifically includes: using a professional camera to collect a face image under normal lighting conditions to ensure that the face is in the center of the image.

3. The face recognition method based on a multi-feature hybrid dictionary and a probability prediction mechanism as claimed in claim 1, characterized in that: The advantage of the Gabor filter used in step 3 is reflected in the extraction of object texture features; the Gabor filter is highly robust to changes in image brightness and contrast as well as changes in image posture, and expresses the local features that are most useful for image recognition.

4. The face recognition method based on a multi-feature hybrid dictionary and a probability prediction mechanism as claimed in claim 1, characterized in that: The dictionary construction in step 4 must ensure that D N*K Among them, N must meet the requirement of being much smaller than K, that is, relative to the complete orthogonal basis, the basis of the over-complete basis must be redundant, so that D is complete enough to ensure that the space opened by all atoms is large enough to contain all training samples, and use the redundant characteristics of the dictionary to better capture the essential characteristics of the samples.

Images