基于SRM方法的人脸识别分析-仪器科学与技术专业论文.docxVIP

Abstract Face recognition (FR) is origined from pattern recognition, a hot research topic in image processing area. After several decades of study, FR has achieved a lot goals and rooted in many fields. In most current reasearches one of the important issues which affect FR accuracy is feature extraction, which is mostly done through linear transformation. Optimal feature extraction remains to be a tricky question for researchers. However, based on signal statistics, the Sparse Representation Method (SRM) provides a novel solution. It exploits a linear combination of the training samples to represent test samples and uses NS principle to classify test images, regardless of feature extraction. Accoding to SRM, to build a robust classifier is much more important than the feature extraction in a face recogntion system. SRM has its own advantages and weeknesses. It is an iterative method with heavy computational load for solving an L1 minimum problem. In this paper, we propose several improved algorithms based on SRM, which require lower computational cost. We also evaluate the performance of our algorithms through many FR experiments. The algorithmic difficulty of SRM lies in how to calculate the sparsest coefficients of linear expression of the training samples. This means the classification of test sample will cause a heavy load of calculation. In this paper, based on building simplified classifiers and other practical factors, we propose four new improved algorithms considering the following key points respectively: using average face of samples to build representation, adding average face as a sample, applying a different face recognition algorithm as pretreatment, exploiting virtual training sample. We use well-known public face databases (ORL, YALE, AR) to test our algorithms. Different propotions of training and test examples are used to verify the reliability of our approaches. All four experiments are implemented in MATLAB. From the results of the experiments