Feature Selection Algorithm Based On L₀-SVM And The Application In Face Recognition

Face recognition is an important branch of pattern recognition,face image detection and recognition is also a representative problem in the field of image processing.It is very difficult to identify face images because it is influenced by complex factors such as illumination,expression and occlusion,as well as face orientation and size difference in face images.Therefore,the study of face recognition algorithm which is accurate,fast and stable,not only has important value for the relevant application field,but also can be extended to other applications,for example,target recognition,detection and tracking,and has far-reaching significance.In this paper,we propose a feature selection algorithm and face recognition based on face landmark detection location,Gabor block feature characterization,L₀-SVM approximation function construction and DC optimization theory.Firstly,adopting the landmark detection method with structure tree model,face places and 68 landmarks in the image of natural scene can be obtained,meanwhile,it determine the rotation angle of the face in the image.Secondly,preprocessing operations,such as gray processing,shear,standardized size,can be taken base on the detected face image and landmarks.Then in this paper,a feature of Gabor block histogram,based on the existing Gabor features,was proposed to represent the location of each landmark.A block,or 10*10 pixel whose center is the landmark,can be divided into 5*5 cell in which Gabor feature can be calculated in 8 directions and 5 size,then a feature vector can be get by accumulation.Then,the block feature can be expressed as 4 cell series,and face feature can be expressed as all block feature series.Then,in view of high dimensions,large quantities of Gabor feature vector,this paper proposes a feature selection method based on L₀-SVM to reduce feature's dimension,remove redundant and irrelevant features.The recognition in feature subsets greatly reduces the time and space complexity and improves the classification accuracy.Subsequently,the approximate function model is constructed,and the DC decomposition and iterative solutionare carried out according to the idea of DC programming and DC algorithm,and its validity is proved in theory.Finally,the algorithm is verified by numerical experiments,which proves the advanced nature and effectiveness of the proposed algorithm.