Researches Of Face Recognition Based On PCA Feature Dimension Reduction And Sparse Representation

Face recognition, as one significant research branches of pattern recognition, possesses great research value and wide engineering foreground, whose existing related studies have made considerable progress. Sparse representation, an emerging field of signal processing, has attracted tremendous attentions due to its advantages such as simple representation, high accuracy and strong robustness. Theoretic analysis and simulation results have proved that face recognition systems perform well based on sparse representation. However, there are still some issues to affect or restrict its development, including image occlusion and corruption, weak real-time performance, etc. Aiming at its deficiencies, we further study the method for solving underdetermined equations, feature dimension reduction, together with classifier design, for the sake of better accuracy and robustness. The main contents and contributions of our works are as follows:1. Three pursuit algorithms to find the sparse solution are elaborated and compared. And more iterative convergence conditions are added to OMP based on the particularity of face recognition, in order to reduce the computational time and increase the sparse level of solution vectors. Emulation experiments have shown that the performance of our algorithm is better than other greedy methods.2.Another ways to solve the sparse solution--- iterative shrinkage algorithms have been discussed in detail. We adjust the weight matrix in the algorithm and then plug it into an iterative process. Finally, the improved OMP algorithm and improved IRLS algorithm are fused for better effectiveness and recognition rates at the end of chapter 3. Simulations have proved that the fusion algorithm outperforms others both in operation time and recognition results.3. Several feature dimensionality reduction methods are introduced, including down-sampling, random sampling and Principal Component Analysis (PCA). Also, two feature dimension reduction algorithms are proposed for contiguous occluded or disguised samples. The robustness tests have shown the efficiency of the algorithms.4. We have introduced wavelet decomposition in face recognition algorithm based on sparse representation, in order to get better recognition effects. Training dictionary of SRC in spatial domain is replaced by 4 different dictionaries in frequency domains and the corresponding sparse solutions are solved respectively. The final identification results depend on voting and minimum residuals. It has been proved that this algorithm possesses higher identification rates than SRC, especially when the training samples are not sufficient.