Research On Face Recognition Method Under Local Occlusion

With the development of information technology and Internet technology,identity authentication has gradually become the research focus of the Internet era,and face recognition technology has a unique advantage in identity authentication.At present,face recognition has been successfully applied in security monitoring,face payment,company punch card and other fields,and obtains good recognition results under restricted conditions,but it is still unsatisfactory under unrestricted conditions.Especially regarding the occlusion problem,so far it has not achieved good results because of its complexity.In this thesis,the methods of face recognition under local occlusion are studied.There are the following work and innovations:1.Based on SRC classifier and CRC classifier,from the perspective of error probability distribution and collaborative representation,a classifier based on error weighting and collaborative representation is proposed.Consider occlusion as the error of the original face image without occlusion,and then assign corresponding weights to each pixel to reduce the impact of occlusion.The reconstruction coefficient ? may "overexpress" the non-recognition class because of the similarity of local features,so the 2l-norm is used to "suppress" it.so that more similar samples to express test samples,and then improve the recognition rate.Experiments show that the EW-CR classifier has good robustness under illumination,expression,and occlusion,and the recognition rates of 99.33%,96.00%,and 78.64% on the AR dataset,Extend Yale B dataset,and FLW dataset,respectively.Compared with the latest algorithm of the same type,the recognition rate has been improved by 0.93%,2.80%,2.24%.2.In order to further improve the performance of the EW-CR classifier,a multi-feature classifier is proposed by synthesizing the LBP?HOG?Gabor feature extraction method.New classification residuals have been redefined according to the classification residuals and recognition results acquired by inputting the original pixel features,LBP features,HOG features and Gabor features to the EW-CR classifier,respectively.Then several classification residuals are weighted and the weight vectors are trained by minimizing the sum of classification residuals.Several features are extracted from the test image and input into the EW-CR classifier.Then the classification residual value of the test image is weighted fused by the new weight vector to gain the final classification residual value.The category corresponding to the minimum classification residual value is the final recognition result.Different occlusion ratio experiments,occlusion experiments in uncontrolled enviroment,fixed occlusion type experiments,strong light changes experiments are designed,and good results are procured.98.7% and 83.24% recognition rates are obtained on Extend Yale B dataset and FLW dataset,which are improved by 2.7% and 4.6% respectively compared with the original algorithm,which verifies the effectiveness of the proposed algorithm to deal with local occlusion problem.