Occluded Face Recognition Algorithm Based On Partitioned Sparse Representation

Occluded face recognition is widely used in video surveillance and humancomputer interaction and other application scenarios.On the basis of sparse representation-based classification,i.e.SRC,algorithm,this master dissertation focuses on the occluded face recognition algorithm based on partitioned sparse representation.Aiming at the problem of a major feature of face divided into several sub-blocks,vote allocation according to error of similarity among occluded sub-blocks and so on for the existing partitioned SRC occluded face recognition algorithm,a "?-weighted voting feature block occluded face recognition algorithm",namely "algorithm E",is proposed,feature block of which ensures that a major feature and its occluded region of face fall into a sub-block alone.The ?-weighted voting method is used to allocate votes for a sub-block to increase difference of vote allocation of occluded sub-blocks.In addition,"algorithm E" reduces misclassification of occluded sub-blocks with the help of that occluded face images are added to the training sample set to increase its proportion of occluded images.The experimental results show that the recognition rate of "algorithm E" for occluded face images is gradually improved with the increase of proportion of occluded images in a training sample set,recognition rate of which is 93.17% under occlusion of sunglasses,and its recognition rate is 95.33% under occlusion of scarf,which is higher than one of other algorithms.The experiment verifies effectiveness of the proposed "algorithm E" for occluded face images under conditions of adding occluded images to a training sample set.Aiming at the problem of poor recognition rate of "algorithm E”,the training sample set of which does not include any occluded face image,an "improved ?-weighted voting occluded face recognition algorithm" is proposed by introducing weight matrix.On the basis of feature block of image,the weight of the occluded subblock is increased,while the weight of the nose sub-block is kept unchanged,and the weight of the rest of sub-blocks is reduced.The experimental results show that the recognition rate of "improved ?-weighted voting occluded face recognition algorithm" is 87.83% and 93.50% in the case of sunglasses and scarf occlusion respectively,which is 20.16% and 5.00% better than one of "algorithm E" respectively.In order to further improve recognition rate of occluded face images and eliminate influence of error information brought by the occluded region to participate in voting classification in occluded face images,a "SCI and voting weighted occluded face recognition algorithm" is proposed by using of sparsity concentration index,i.e.SCI,in which threshold is set,and sub block with SCI below the threshold is culled,while the remaining sub blocks are weighted.The experimental results show that the recognition rate of the "SCI and voting weighted occluded face recognition algorithm" is 94.17% under occlusion of sunglasses,and is 93.67% under occlusion of scarf,which is higher than one of other algorithms respectively.The experiment verifies effectiveness of the "SCI and voting weighted occluded face recognition algorithm" for occluded face recognition.In conclusion,the above-mentioned algorithms show high recognition rate and robustness to the face image under occlusion of sunglasses and a scarf,and they can be applied in the field of image recognition engineering.