Research On Partial Occlusion Algorithm Based On Fusion Of Global And Local Features

With the continuous maturity of deep learning,computer vision and other related technologies,face recognition has become the mainstream of the world in the field of video intelligence applications.Many stations and airports have already used a large number of face recognition clearance inspection systems,making face recognition this technology has received great attention from various fields,its market demand is increasing,and application scenarios are constantly being tapped.In addition to the two major areas of security and finance,face recognition has been widely used in many scenarios such as medical treatment,education,and transportation,and has shown significant application value.For each person,the facial features are highly unique.Although the face recognition research under ideal conditions and controllable settings has achieved good results,the face collected under the uncontrollable real natural environment Image information will be disturbed by many factors such as human hair,masks,hats,sunglasses(eyeglasses)and other commonly worn objects,and other factors such as light and expression,which makes part of the face information missing,which affects the extraction and recognition of features,reducing the correct rate of face recognition.Therefore,the research on partial occlusion face recognition is one of the important subjects of the increasingly widely used face recognition technology.At the same time,it is also very important to conduct more in-depth research on face linear reconstruction and face feature extraction.It is critical to design an algorithm that minimizes the impact of occlusion on face recognition and has good performance.In response to this problem,many scholars and researchers have conducted in-depth research,but the difficult to overcome areas still exist: if the occlusion area exceeds the critical point,such as when the occlusion area of the face is greater than 20%,the face recognition effect will be Significantly decreases,so the robustness of the algorithm is average;if the occlusion of the face is caused by other objects randomly,the current method has not been able to achieve the desired experimental effect.This paper aims at face recognition with the appearance of the local shelter,which appears frequently in practical application,and make relevant research.This paper focuses on the research of face recognition under partial occlusion from the following two aspects:at first,aiming at the shortcomings of the traditional facial featureextraction algorithm in the feature description when partial face occlusion occurs,an improved error image feature extraction algorithm about partial face occlusion is used;Then,based on the research and analysis of traditional face recognition algorithms,a partial occlusion recognition algorithm,which is based on the Haar's diagonal Hog feature and combinating the dual attribute model,is proposed.The effectiveness of the proposed method is verified by the occlusion experiments in AR and Yale face database.The specific research work is as follows:(1)When using the PCA principal component analysis,the global features with partial occlusion can be extracted.Due to the effects of occlusion,the accuracy of face recognition will be too low.Taking into account the above issues,in this paper,the face error image obtained by using the improved LRC algorithm(ILRC)not only effectively reduces the colinearity(also known as autocorrelation)existing between variables by using singular value decomposition,so as to weaken the impact of adverse factors in linear regression problems,but also the feature extraction of the image achieves global compensation to obtain relevant error feature information about partially occluded faces.(2)Global feature information of the face area is extracted by PCA method,and the target face is linearly reconstructed by the improved LRC algorithm(ILRC)in(1)to obtain the error image,then the feature vector information of the face error image is fused with the global feature information of the PCA by means of the dual attribute model to obtain dual attribute attribute information about the face.Then,the diagonal Hog feature based on the Haar characteristic is used to evenly extract local feature information from the target face,then it uses the designed overall classifier to classify the extracted dual-attribute feature information and local feature information through corresponding weight allocation to determine the category of the target face.(3)In the face library AR and Yale,partial occlusion related experiments are performed.One is the performance analysis experiment completed in the AR library,and the other is the random face occlusion experiment performed in the Yale library.The above experiments show that the proposed face recognition algorithm based on the improved LRC(ILRC)and diagonal Hog,using dual attribute model is effective for partial occlusion.