CN104392246B - It is a kind of based between class in class changes in faces dictionary single sample face recognition method - Google Patents

Abstract

The present invention proposes a single-sample face recognition method based on an inter-class and intra-class face change dictionary, which solves the limitation problem of the current single-sample face recognition algorithm. Step 1: Obtain the image expression of the face image in the compressed domain; Step 2: Construct a face image training sample matrix containing k classes; Step 3: Construct the average face matrix and the inter-class face change matrix of the face database; Step 4: Add low-rank and sparse constraints to the inter-class face change matrix; Step 5: Solve the inter-class similarity matrix and inter-class difference matrix; Step 6: Project the average face matrix, inter-class similarity matrix and inter-class difference matrix to low Dimensional space; Step 7: Normalize the average face matrix, inter-class similarity matrix and inter-class difference matrix after dimension reduction, and use the norm optimization algorithm to iteratively solve the training samples based on face images The sparse coefficient vector of the matrix; Step 8: Select the column vector face label in the average face matrix corresponding to the maximum value of the sparse coefficient as the result of the final face recognition.

Description

A Single-Sample Face Recognition Method Based on Between-Class and Within-Class Face Variation Dictionary

technical field

The invention belongs to the technical field of computer vision and pattern recognition, and relates to a single-sample face recognition method based on an inter-class and intra-class face change dictionary.

Background technique

As an important biometric technology, face recognition has the advantages of naturalness and strong privacy. Therefore, it has attracted widespread attention in the past few decades, among which there are simpler but more successfully applied face recognition methods based on statistical features such as Eigenface, Fisherface, Laplacianfaces, etc. In recent years, sparse representation has been applied in the field of face recognition and has achieved great success. The face recognition classification method based on sparse expression (Sparse Representation Classifier, SRC), its idea is to use all training images as a super-complete dictionary, and the test image can be expressed as a linear combination of a few face images in the dictionary. Wright et al. confirmed that the problem can be solved by fast l ₁ norm optimization algorithm, and achieved the best results in face recognition experiments. At present, researchers have successively proposed a variety of improved algorithms for the shortcomings of sparsely expressed face recognition methods. For example, on the basis of SRC, low-rank matrix recovery is added, and the uncorrelation of dictionary representation is used to distinguish the face from the noise part.

However, using the face as a biological feature makes face recognition inherently insurmountable shortcomings. First of all, although the similarity of face structure is beneficial to face detection and positioning, it is inconvenient to distinguish different faces. Secondly, there are many instabilities in the shape of the human face, such as rich facial expressions, complex external environments (lighting, imaging angles, etc.), face occluders (masks, sunglasses, etc.), and age. Facial changes. These instabilities will affect the performance of face recognition algorithms. In the face recognition algorithm, the change between different faces belongs to the inter-class change, while the shape change of the same face belongs to the intra-class change. The variation between classes and within classes makes face recognition considered one of the most difficult research topics in the field of biometrics and even in the field of artificial intelligence. With the rapid popularization of video surveillance, in many applications, especially in large-scale identity verification occasions, such as law enforcement, driver's license and passport card verification, we usually only collect one sample image for each person in the database. In this case, it is necessary to complete the face recognition task under different viewing angles, illumination, occlusion, expression and other changing factors only based on a single image of the face. However, the face recognition algorithm based on sparse expression proposed by Wright et al. needs to provide face images of each person under various changing conditions. That is, the algorithm requires a large training sample database to ensure the accuracy of the algorithm. W. Deng et al proposed to extend the sparse expression classifier to solve the under-sampled face recognition problem, and even the single-sample face recognition problem. Then, based on the extended sparse expression classifier, a single-sample face recognition algorithm with average face plus face variation was introduced.

In recent years, various signal processing methods based on compressed sensing theory have become one of the standard signal processing methods for computer vision and pattern recognition. Therefore, it is a logical idea to apply compressive sensing theory to the field of sparsely expressed face recognition. A. Majumdar et al. demonstrated that random projections are robust to several recently proposed classifiers, namely sparse classifier (SC), group SC and nearest neighbor classifier (NN). Since face recognition needs to process a large number of high-precision face images, using compressive sensing for dimensionality reduction in image domain face recognition has attracted high attention.

Contents of the invention

The purpose of the present invention is in order to overcome the defective of prior art, solve the problem of the limitation of current single-sample face recognition algorithm, propose a kind of single-sample face recognition method based on face variation dictionary within class between classes.

The inventive method is realized by the following technical solutions:

A single-sample face recognition method based on an inter-class intra-class facial variation dictionary, comprising the following steps:

Step 1: The face image in the database is projected and mapped using a random speculative matrix to obtain the image expression of the face image in the compressed domain;

Step 2: Construct a face image training sample matrix containing k classes;

Step 3: Construct the average face matrix and the inter-class face change matrix of the face database, and each column in the average face matrix represents the average face of the ith face training image.

Step 4: Add low-rank and sparse constraints to the inter-class face change matrix;

Step 5: Use the augmented Lagrangian dictionary training algorithm to further decompose the inter-class face change matrix, and solve the inter-class similarity matrix and inter-class difference matrix;

Step 6: Project the average face matrix, inter-class similarity matrix, and inter-class difference matrix to a low-dimensional space using the PCA dimensionality reduction algorithm;

Step 7: Normalize the average face matrix, inter-class similarity matrix, and inter-class difference matrix after dimension reduction, and use the norm optimization algorithm to iteratively solve the sparseness of the training sample matrix based on face images. coefficient vector;

Step 8: Select the column vector face label in the average face matrix corresponding to the maximum value of the sparse coefficient as the result of the final face recognition.

Beneficial effects of the present invention:

The present invention can construct a noise dictionary for noises such as illumination, occlusion, and posture changes that affect the single-sample face recognition algorithm, and divide the noise dictionary into inter-class noise and intra-class noise, thereby effectively improving the performance of face single-shot based on sparse expression. Robustness of sample identification algorithms.

detailed description

The present invention will be described in further detail below in conjunction with specific implementation examples. Here, the exemplary embodiments and descriptions of the present invention are used to explain the present invention, but not to limit the present invention.

The present invention proposes a single-sample face recognition method based on an inter-class and intra-class face change dictionary, which is based on random projection and sparse representation theory, and is an improved algorithm for single-sample face recognition. The key to solving single-sample face recognition in the present invention is to establish a robust and universal noise model that has nothing to do with face images. The noise model is mainly aimed at modeling facial expression changes, environmental lighting changes, facial posture changes, and facial occluders (masks, sunglasses, etc.) to form a dictionary. In this way, these noises are separated from face information, thereby improving the accuracy of single-sample face recognition. Its specific steps include:

Step 1. The face image in the database is projected and mapped using a random speculative matrix to obtain the image expression of the face image in the compressed domain, which can be expressed mathematically as y=Φx. Where Φ is the projection matrix, x is the face image, and y is the face image in the compressed domain.

Among them, for the training images and test images in the face database, the same random projection matrix Φ is used for dimensionality reduction.

Step 2. Construct a dictionary of face image training samples containing k classes Where n=n1+n2...+nk, n _i is the number of training samples for each type of face. Each column in the dictionary D matrix is a column vector representation of face images in the compressed domain. Any query or test sample y can be expressed as a linear combination of this dictionary, which can be expressed mathematically as y=Dα+e, where e is random noise, and α is a coefficient based on the dictionary D

Wherein, in face recognition, n samples of face individuals of k types are included. These n samples are formed into a dictionary D, and each column of the dictionary D represents a face image in the compressed domain.

Step 3. Construct the average face dictionary of the face database and between-class face change dictionaries Each column Pi in the average face dictionary P represents the average face of the _ith face training image.

Among them, each column in the average face dictionary represents the average value of all pixels of all training images of a type of face, and each column of the inter-class face change dictionary V represents the difference between the training sample image and the average face of this type of face.

Step 4: Add low-rank and sparse constraints to the inter-class face variation dictionary V.

Among them, low-rank optimization is to extract positive and clean faces from human faces, and sparse constraints are to constrain various facial changes.

Step 5. According to the formula (1), use the augmented Lagrange dictionary training algorithm to further decompose the inter-class face change dictionary V, and solve the inter-class similarity matrix E and inter-class difference matrix G

Among them, because in the face change dictionary, the face contours of each type of face are highly correlated, so different face changes can be extracted through inter-class similarity and inter-class difference.

Step 6: Using the PCA dimensionality reduction algorithm to project the average face matrix, the inter-class similarity matrix E and the inter-class difference matrix G to a low-dimensional space.

Step 7: Normalize the average face matrix P, inter-class similarity matrix E, and inter-class difference matrix G after dimensionality reduction using the L ₂ normalization method, and use the l ₁ -l ₂ norm according to formula (2) The number optimization algorithm iteratively solves the sparse coefficient vector c=[α β γ] based on the face image training sample dictionary.

Step 8: Select the column vector face label in the P matrix corresponding to the maximum value of the sparse coefficient α as the final face recognition result.

Among them, the sparse coefficient α obtained in the previous step indicates the degree of correlation between the test image and the training image, and the largest coefficient is found through enumeration, and the corresponding face is the face category of the test image obtained by the algorithm.

Since then, the one-shot face recognition problem has been completed/implemented.

The specific description above further elaborates the purpose, technical solution and beneficial effect of the invention. It should be understood that the above description is only a specific embodiment of the present invention and is not used to limit the protection of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims (2)

1. a single-sample face recognition method based on a dictionary of facial changes within a class between classes, characterized in that, comprising the following steps: Step 1: The face image in the database is projected and mapped using a random speculative matrix to obtain the image expression of the face image in the compressed domain; Step 2: Construct a face image training sample matrix containing k classes; Step 3: Construct the average face matrix and the inter-class face change matrix of the face database, and each column in the average face matrix represents the average face of the ith face training image; Among them, the inter-class face change matrix V is specifically: <mrow><mi>V</mi><mo>=</mo><mo>&amp;lsqb;</mo><msubsup><mi>d</mi><mrow><mi>n</mi><mn>1</mn></mrow><mn>1</mn></msubsup><mo>-</mo><msub><mi>P</mi><mn>1</mn></msub><mo>,</mo><msubsup><mi>d</mi><mrow><mi>n</mi><mn>1</mn></mrow><mn>2</mn></msubsup><mo>-</mo><msub><mi>P</mi><mn>1</mn></msub><mo>,</mo><mn>....</mn><mo>,</mo><msubsup><mi>d</mi><mrow><mi>n</mi><mi>i</mi></mrow><mn>1</mn></msubsup><mo>-</mo><msub><mi>P</mi><mi>i</mi></msub><mo>,</mo><msubsup><mi>d</mi><mrow><mi>n</mi><mi>i</mi></mrow><mn>2</mn></msubsup><mo>-</mo><msub><mi>P</mi><mi>i</mi></msub><mo>,</mo><mo>...</mo><mo>,</mo><msubsup><mi>d</mi><mrow><mi>n</mi><mi>k</mi></mrow><mn>1</mn></msubsup><mo>-</mo><msub><mi>P</mi><mi>k</mi></msub><mo>,</mo><msubsup><mi>d</mi><mrow><mi>n</mi><mi>k</mi></mrow><mn>2</mn></msubsup><mo>-</mo><msub><mi>P</mi><mi>k</mi></msub><mo>,</mo><mo>...</mo><mo>&amp;rsqb;</mo></mrow> in, is the column vector representation of the face image in the compressed domain, ni is the number of training samples for each type of face, P _i is each column in the average face matrix, j=1,2, i=1,2,... ,k; Step 4: Add low-rank and sparse constraints to the inter-class face change matrix; Step 5: Use the augmented Lagrangian dictionary training algorithm to further decompose the inter-class face change matrix, and solve the inter-class similarity matrix and inter-class difference matrix; Step 6: Project the average face matrix, inter-class similarity matrix, and inter-class difference matrix to a low-dimensional space using the PCA dimensionality reduction algorithm; Step 7: Normalize the average face matrix P, inter-class similarity matrix E, and inter-class difference matrix G after dimension reduction, and use the norm optimization algorithm to iteratively solve the training samples based on face images The sparse coefficient vector c of the matrix = [α β γ]; Step 8: Select the column vector face label in the average face matrix P corresponding to the maximum value of the sparse coefficient α as the final face recognition result. 2. a kind of single-sample face recognition method based on the facial change dictionary within a class between classes as claimed in claim 1, is characterized in that, carry out more accurate modeling for various face face changes, these faces Changes are separated from frontal faces, thereby addressing the impact of face changes on recognition accuracy in one-shot face recognition problems.