| As a comprehensive subject, face recognition covers the fields of pattern recognition, image processing, artificial intelligence, computer vision and so on. Experts and scholars in the field of face recognition have proposed many algorithms by now, but algorithms in current research have the trends of becoming more and more complex and costing a lot of computation. This paper did some research on classical algorithms and tried to improve the recognition rate with the thoughts of modular and two-level methods. The creative work of this paper includes:(1) Modular Two-dimensional Locality Preserving Discriminant AnalysisWhen the sample images were exposed to light and shade, the recognition rate of the original 2DLPDA algorithm would be badly affected. So we use the modular method to solve this problem. We collect the blocks of the same position to make up a new subset. Then 2DLPDA was carried out on each new subset to get the projection matrix. The query image was divided into blocks, and each block was made the projection on the relative projection matrix. At the end of this, all the local projections are integrated together to be the basis for the identification. Experimental results show the proposed method is effective to improve the recognition rate.(2) A Two-level Face Recognition Method Based on Modular Two-dimensional Locality Preserving DiscriminantThis algorithm is a continuation of the problem in the previous algorithm. Samples were firstly divided into blocks, and then 2DLPDA was carried out on the new subset which made up by the blocks in the same position to seek out the category which the test sample was locked in. It is found that M2DLPDA can be effectively locked the test sample into a very small category of classes. In the second stage, we just identity the image in this category, and get the final class of the test sample was recognized.(3) A Two-level Face Recognition Method Based on Modular Principal Component AnalysisIn the new method, we don’t make the recognition of the query sample in the whole structure, after we obtain the feature vector of each block. Each block of the query sample is just identified in its corresponding training sample blocks of the same position. So each block in the query sample can have a recognition result. Then we make the second stage of algorithm, we just identity the query sample in this category of results, and get the final class of the test sample was recognized. The process of the second stage is made on the whole of the original sample, and we make no longer any feature extracting.(4) A Two-level Reconstruct Face Recognition Method Based on Nearest Neighbors Weighted Collaborative RepresentationIn this algorithm, we introduce the neighbor and reconstruction to the weighted collaborative representation. Firstly, we find out the neighbors in each class of the query sample. And then we use the neighbors of the training sample to make a new subset. Based on this subset, we construct the weighted collaborative representation of the query sample to get the coefficient. With the coefficient we reconstruct the query sample in each class and get a category of reconstruction samples to make up a new reconstruction subset. In the second part, we construct the weighted collaborative representation of the query sample on the new reconstruction subset to get the coefficient. In this part, we design two different methods to make the second reconstruction. One method we use the whole reconstruction subset to construct the weighted collaborative representation. Another one we use the neighbors of the query sample in the reconstruction subset to make the weighted collaborative representation. |
PDF Full Text Request