Research On Face Recognition Algorithm Under Local Occlusion

Face recognition technology has been widely used in all aspects of life,such as unlocking,punching,and payment.With the raging of the new coronavirus,masks have become a necessity for people’s lives,and improving the recognition rate of local occluded face recognition technology is a top priority for researchers in the field of computer vision.There are many related algorithms proposed by researchers,many of which have been used in engineering practice,but the recognition rate cannot reach the current level of face recognition,and there is still room for improvement.The sparse representation classification algorithm(SRC)and the collaborative representation classification algorithm(CRC)have the advantage of high robustness to occlusion noise,and become the research objects that have to be considered when studying the face recognition problem of local occlusion.Robust Principal Component Analysis(RPCA)has the ability to separate occlusions and original images from contaminated images.Based on these three algorithms,this paper obtains two algorithms through improvement: one is RPCA-based sub-module collaborative representation classification algorithm(RBCRC),and the other is RPCA-based weighted sparse representation classification algorithm(RWSRC).The specific work includes the following two aspects:1.The paper improves the sub-module collaborative representation classification algorithm,and obtains the sub-module collaborative representation classification algorithm(RBCRC)based on RPCA.Firstly,the RPCA algorithm is introduced to obtain the occlusion information of the sparse error image,and the occlusion degree of each sub-block is quantified according to the proportion of high-gray-scale pixels in each sub-block of the sparse error image.Then substitute the quantized value into the S membership function,and use the output value to correct the Borda votes corresponding to each sub-block.The final tally of the voting results is the final recognition result.The quantification method of the sub-module occlusion degree proposed in this paper can separate the occlusion area,and the obtained occlusion degree is closer to the intuitive feeling of the human eye,which is the key to improving the accuracy of the algorithm.2.The paper introduces the weighted sparse representation classification algorithm based on the RPCA algorithm,thereby improving and proposing a weighted sparse representation classification algorithm based on RPCA(RWSRC).There are two improvements: First,the σ parameter expression in the Gaussian distance calculation formula is improved,so that the numerator and denominator of the exponential part of the Gaussian distance expression are of the same order of magnitude,so as to avoid drastic changes in the weights when calculating the weight matrix.Second,the sparse error image is obtained by the RPCA algorithm,and then the sparse error image is obtained through two steps to obtain the occlusion support image.Taking the occlusion support map as a reference,remove the influence of occlusion when calculating the Gaussian distance to avoid the Gaussian distance being too large.In this way,a more accurate weight matrix is obtained to correct the non-occlusion dictionary in the Robust SRC algorithm.This paper designs a series of experiments to verify the performance of the algorithm and determine the optimal values of some parameters.Experiments show that on the AR face set,the recognition rate of the RBCRC algorithm is about 15% higher than that of the sub-module collaborative representation classification algorithm(BCRC)without occlusion quantization;the recognition rate of the RWSRC algorithm on the AR face set is higher than that of the weighted sparse representation classification algorithm(WSRC)is improved by around 5%.