Research On Local Occlusion Face Recognition Based On Gabor Features

Facial recognition technology has received more and more attention in various real-time applications such as law enforcement,security,access control,surveillance,and smart cars.In actual scenes,the face may be occluded by various objects in various ways,such as scarves,sunglasses,and masks,which will seriously affect the performance of the face recognition system.At present,it is a difficult problem in the face recognition application of the face.This article studies the existing face recognition algorithm,and proposes two types of face recognition algorithms to achieve the goal of accurately identifying the face of the face of the human face.The algorithm is as follows:(1)An algorithm of occlusion face recognition based on Gabor feature fusion is proposed.In this algorithm,the image is first divided into blocks,and the image sub-blocks are filtered by Gabor filter to obtain the magnitude subband and angle subband,and then the correlation matrix is calculated respectively for fusion.In order to obtain the discriminative features,Log-Euclidean is used to embed the fusion correlation matrix,to obtain the face features and the learned features are reduced and normalized by WPCA and ZSCORE.Then the processed features are input into the Ada Boost integrated classifier to realize occlusion face recognition.The algorithm was experimentally verified in data sets including scarf occlusion,sunglasses occlusion,random block occlusion and mask occlusion.The experimental results show that the proposed method has certain recognition advantages,and the recognition rates can reach 89.5%,63.83%,85.79% and 93.1% respectively under the above four occlusion conditions.(2)The face recognition algorithm based on partition selection is proposed.This method first divides the image into the sub-block that is not connected,and uses the square root error of the image to determine the blocking area of the face image and abandon it;use the Gabor filter in different scale and direction To obtain the characteristic information of the image in the local frequency and the airspace;then according to the feature information,use the similarity of the string to identify the image and classify;finally,the identification results of the test image partition are fixed to obtain the final classification result.Through experimental verification,the accuracy of algorithms was 97.27%,98.83%,95.26%,and98.28%,respectively.Compared with many traditional algorithms,this algorithm has also improved to a certain extent.In response to the problem that the existing face recognition system is difficult to solve the local occlusion,this paper puts forward two different algorithms from the point of view of whether to abandon the occlusion part.The results show that the two proposed algorithms can improve the influence of occlusion on the existing face recognition system to a certain extent.