Research On Face Recognition Method Based On Compressed Sensing Theory

Face Recognition is an important technique in the field of pattern recognition, itsmain task is to obtain the effective information from face images, the face samples aredivided into corresponding categories of model identification. Here feature extractionis very important. A good facial feature extraction method has the followingadvantages: simplifies classifier designing in the face recognition; improves themethod performance.Compressed Sensing is a new signal sampling and compression theory. Manyexperts and scholars have applied it to the problem of face recognition, and achievedfruitful results. One of the most classic method is the sparse representation of facerecognition algorithms (SRC). The face recognition algorithm of this paper has beenimproved on the basis of SRC algorithm, and achieved good results. The work andresults obtained are as follows：1. We introduce the feature extraction and classification algorithms, the facialfeature extraction methods can be divided into two types：feature extraction methodsbased on geometric features and based on statistical learning; in addition, we introducethe theory of compressed sensing and the most classical face recognition algorithm.2. We propose a face recognition algorithm based on principal component analysisand compressed sensing (PSL0). The algorithm uses principal component analysis toreduce the dimension of the image data, and then use that smooth algorithm based onfast sparse representationl0norm to solve thel0norm minimization，so as to obtaina set of optimal coefficients of image reconstruction, to compute the residuals betweentest and train images for face recognition. Experimental results show that this methodnot only has a high recognition rate in a lower dimension, but also reduces thecomputational complexity.3. We propose ISL0algorithm for compressed sensing signal reconstruction. Theideal is to use a smooth function to approximate thel0norm minimization problem incompressed sensing algorithm and use the modified Newton algorithm in the subspaceoptimization problem. Meanwhile, the search range is constrained by using iterativesubspace method, to improve the efficiency of the algorithm.