Study On Face Recognition Based On Regression Theory And Parallel Computing

As a scientific problem, face recognition involves in many fields such as image processing, pattern recognition, artificial intelligence, machine learning, and cognitive science etc. Therefore, the development of face recognition must have benefit on principles and technologies of these areas. As an important biometric technology, face recognition can be applied to many fields such as national public security, information security and financial etc. Therefore, it has broad application prospects.Recognition accuracy and recognition speed are two important performance indexs of a face recognition system. To improve recognition accuracy, an analysis on the limitation of the existing regression methods is first presented, and then, three modified methods based on regression theory are proposed in order to promote the robustness for variations of light, expression, occlusions and pose. To increase the recognition speed, two kinds of parallel implementation schemes are investigated:(1) researches on face recognition algorithms that are easily implemented by hardware circuits to perform parallel and distributed computing;(2) researches on parallel algorithms running on the PC cluster system. The main contribution of this dissertation is as follows:To improve recognition accuracy, an analysis on the limitation of the sparse representation classification (SRC) method is presented at first. Then, a weighted multi-channel Gabor sparse representation classification algorithm is proposed by exploiting the fact that different Gabor features distinctively contribute to the face recognition task. To exploit the local information provided by the spatial locations of Gabor features, an ensemble classifier constructed with local Gabor sparse representation classifiers by fuzzy fusion is further proposed. By analyzing the reason why the SRC method can't deal with large pose variation, a pose-invariant face recognition method based on kernel ridge regression is also proposed by using non-linear kernel transformation technique. The porposed method can estimate the pose-invariant representation coefficients on the3D face space more accurately. In this way, the variation of representation coefficients between the images of the same person with different poses is reduced, and the identity correlation between intra-individual images with different poses is strengthened.Since it is very time-consuming to compute the sparse coding coefficients for the SRC method, one-layer recurrent neural network (RNN) model is introduced for solving the minimum l1norm problem and the minimum l0norm problem, respectively. The RNN model has the distributed parallel processing ability when it is implemented by hardware circuits. To solve the minimum l0norm problem, a group of piecewise inverted Gaussian functions are first designed, and their summation is taken as objective function to approximate the l0norm. Then, a heuristic search strategy is further proposed during the RNN optimization process. The constructed objective function provides a convex condition under which the RNN is globally convergent. Besides, the RNN yields very satisfactory reconstruction results with the heuristic search scheme. The RNN optimization method is also applied to the SRC method to solve sparse representation coefficients, and better recognition accuracy is achieved.To increase the computational speed, three parallel face recognition algorithms are designed, which includes the parallel weighted multi-channel Gabor sparse representation classification algorithm, the parallel face recognition algorithm based on LGBPHS, and the parallel error-correction SVM algorithm. These parallel algorithms are implemented based on PC cluster system, and their parallel performance is also evaluated. Experimental results show that the proposed parallel algorithms can speed up the training and classification tasks efficiently, at the same time, keep the recognition accuracy unchanged.