Research Of Occluded Face Recognition Based On Image Blocking Method

Face recognition is a frontier research topic that combines multidisciplinary knowledge methods,which is of great theoretical and practical significance.In this paper,the related works of occluded face recognition are analyzed in detail,and then we propose two image blocking methods which are used to locate occlusion.These methods reach higher recognition rate than other methods.The main works are the following parts:Firstly,the background and present situation of face recognition are briefly described.The development of face recognition method based on sparse representation and deep learning is introduced in detail.The principle of recognition algorithms and convolution neural network are described in detail;Secondly,we propose two methods of subdivision named deep-block and overlapping-block respectively.In view of discontinuous and continuous occluded image respectively,these methods give full consideration to the continuity of occlusion and iteratively select out the occluded areas by comparing the differences of the SCIs between adjacent sub-blocks.The two methods can adaptively judge that whether a image is suitable for blocking recognition,which avoids the mechanical partitioning of all images;Moreover,these methods iteratively select out those occluded sub-blocks according to the difference between adjacent sub-blocks,which avoids the limitation of determining whether a block is occluded or not by its SCI value.While Deep-block and overlapping-block are used to locate occlusion in view of discontinuous and continuous occluded image respectively,based on these methods,collaborative representation classification(CRC)is used to identify those remaining sub-blocks.A large number of experiments have shown that these methods'recognition rates are higher than the traditional method--CRC,CRC voting,and CRC voting combing with SCI in AR,GT,ORL face databases.Finally,in the situation of limited sample,we propose a method of convolution neural network(CNN)recognition based on image partitioning to identify severe occluded face images.According to the SCI of each sub-block,we firstly filter out those occluded sub-blocks,and then CNNs are used to identify the other un-occluded sub-blocks respectively.The mean square error of each CNN judges when majority vote generates multiple voting results.In this situation,the result voted by CNNs with smaller mean square error is considered as the final recognition result.On the one hand,this method reduces the number of training parameters of CNN,and therefore reduces the training time cost;On the other hand,the mean square error of each sub-block is used as a criterion when generates multiple vote results,which avoids the disadvantage of taking the first vote as result.In this paper,we combine the image partitioning and CNN effectively,and propose a method of CNN recognition based on image partitioning.This method filters out those occluded blocks according to the SCI of each sub-blocks;then identifies those remaining sub-blocks by CNN;and finally votes for the recognition results combing with each training mean square error of each sub-block.A large number of experiments have shown that,in the AR?GT?ORL face databases,the non-occlusion test samples are continuously occluded and discontinuous occlusion,and the occluded-blocks are removed after the blockage.The experimental results show that this method's recognition rates are higher than CNN,CRC,CRC voting,and CRC voting combing with SCI.