Research On Face Recognition Algorithm Based Local Feature Extracting And Sparse Coding

In recent years,the sparse representation based classificaiton(SRC) has been proposed for face recognition(FR) and has achived a good result. But in the case of outliers such as sunlight,occlusions and expression variation are existed in the test samples.The performance of the methods drops dramatically.To address this problem.The thesis conducts in-depth study on sparse representation classification(SRC) for face recognition with illumination,occlusion and small samples.Firstly,According to the advantages of Gabor feature extracting and robust sparse coding encoding the residuals accurately.A new method of robust sparse coding based Gabor feature extracting is proposed for classification.Firstly we use Gabor wavelet to extract the multi-scale and direction information of the image.And then make the model of robust sparse coding.Finally test sambles features are classified by comparing the coding residuals., The experiments show that this proposed method can get a higher recognition rate.Secondly,Since Local binary pattern(LBP) can effectively describe the texture structor information of an image. Besides the kernel technology is applied to pattern recognition and performs well. A new algorithm of robust kernel coding of local features fusion.Firstly we extract the local binary and ternary pattern features of the image and make the features fused.And then the kernel technology is introduced to project the fused features to higher dimension.In the higher dimension it is convenient to classify test samples by linear classifier.Finally we build the the model of robust sparse coding to classify the projected samples. Experiments show that the proposed algorithm increases the recogniton rate and speed during classification.Finally, On the basis of auxiliary dictionary can encode the abnormal pixel values including occlusion and interruptions in the test samples robustly.A new algorithm of robust sparse coding based on auxiliary dictionary learning is proposed. Firstly we learned a auxiliary dictionary changed with training samples by auxiliary samples.And then make the model of robust sparse coding combing training samples dictionary and auxiliary dictionary.We solve the coding models by iteratively reweighted technology. Finally test samples are classified by searching best dictinary element to represent the test samples. Experiments show that this method can achive higher recognition rate is in the case of occlusion and small samples.