Improvement Of Sparse Coding And Its Applications In Face Recognition

Sparse coding has gained considerable attention in recent years, and it has been widely used in lots of fields including blind speech signals separation, feature extraction, data classification, image denoising, image fusion and restoration, pattern recognition, etc. The test sample is represented as a sparse linear combination of the predefined dictionary with the sparse coefficients. We use the coefficients corresponding to the i-th class to reconstruct the test image. Then, we identify the most similar class of test image according to the smallest residual error between the test sample and reconstructed sample. It begs an urgent question that how to build effective sparse coding models especially in pattern recogniton? Based on the large amount of research work, this paper focuses on three new methods for sparse coding in face recognition.The main contributions and innovations in this thesis are as follows:Contourlet transform is a real multi-directional, multi-resolution, multi-scale local geometry image representation method. It can re-preprocess the2-D image directly. On image re-preprocessing stage, the original images were filtered by the Contourlet wavelet transform. Thus, we got low frequency and high frequency characteristics of the original images. Then, we combined these two kinds of characteristics into one-dimensional vector. The vector was put into the sparse coding algorithm as image characteristics to complete the recognition. We proposed the above process in order to get image feature extraction rapidly, remove noise and redundancy, retain local characteristics including image edges, and reduce the dimensionality of data. Our sparse coding algorithm could get faster speed and better recognition performance eventually.On the reconstruction stage in the sparse coding for recognition, a novel framework for the image reconstruction was proposed. In some cases, the messy coding coefficients may be not appropriate to yield a right classification. So it filtered the redundancy coding coefficients by selecting a number of largest coding coefficients called LCE to generate the new coding residual. And the novel coding residual was used to reconstruct the test image instead of the standard residual. This larger coefficient emphasis framework, which improves SRC and RSC, is evaluated on face databases and the experiment results show its practical advantages compared with that of SRC and RSC in the face recognition. On image recognition stage with the coding coefficients, the other classifiers have obvious advantages in the nonlinear and high dimension aspects. So we combined the sparse coding with SVM and BP neural network classifiers. The algorithm is called the recognition classifiers with the sparse coding coefficient characteristics. The coefficients were put into the above two classifiers as characteristics. The improved algorithm has better adaptability in face recognition.