Face Recognition Algorithm Research Under No-ideal Conditions

How to eliminate the the impact of non-ideal conditions on face recognition is the research priority in this field, the core content is obtained unchanged factors in the various environments. At the present stage, main methods include: the total variation quotient image model, the Logarithm Total Variation,, the best anti-block facial synthesis algorithm and so on.Based on the analysis and summary about the face recognition algorithm under non-ideal conditions, in order to elimination the impact of unstable illumination, occlusion and posture, we studied in depth. In This paper we proposed three algorithms to eliminate the impact of unstable environment.1. In order to eliminate the effects of unstable illumination environmental on face recognition, a new method is proposed, called partial differential equation second-order cone program (PDE-SOCP) based on logarithm total variation model (LTV). Anisotropy diffusion PDE model can decompose image into multi-scale space,and second-order cone program has the advantages of using existing internal points. Combine them and optimize LTV model algorithm to get the PDE-SOCP algorithm.2. Occlusion has a large impact on face recognition, in order to resolve this problem we proposed alternating minimization algorithm (AMA), mainly to solve the problem of image restoration after de-noising. The algorithm is based on the regularization of total variation model and the sparsity constraints, and then to calculate the accurate results through the iterative algorithm. From the perspective of statistical analysis, we prove the practicability of AMA. So we can obtain the image edges and texture part, in this way, we can restore the image well.3. Based on the principle of scale invariant feature transform algorithm (SIFT), we optimize the matching process to improve similarity measurement criterion. in image matching process, the criteria used linear combination of the absolute distance and chessboard instead to instead of Euclidean distance. In this way, the number of feature points will not change with the posture, so reduce the matching number of the key points during comparison. This optimized algorithm can improve face recognition rate.The feasibility and effectiveness of three methods had been demonstrated through extensive experiments conducted on several face databases.