A GPU Based PCA Algorithm

Face recognition technology is a method to identify people based on human face features. Nowadays, face recognition technology has become an indispensable part of the society. Principle Component Analysis (PCA) algorithm is one of the major mechanism used in face recognition technologies. With thousands of human faces to recognize and pixels to read in the procedure, the matrix operations become very computationally expensive which makes PCA algorithm bring unsatisfactory performance. The paper brings out a GPU-based parallel PCA face recognition algorithm. It first deeply analyzes both training procedure and recognition procedure of the algorithm, and confirms that only to improve the training procedure due to the unavoidable data copy and the time complexity. Then grabs certain steps from the training procedure which can be parallel improved, namely matrix row calculating, matrix multiplying and Jacobi iterate method. Among them, Jacobi iterating is the most complex, the paper uses a single cycle algorithm to traverse the non-diagonal elements, so as to parallel improve Jacobi iterate method. Next, it’s been found out that GPU improved PCA algorithm is found sensitive to the capability of training set, the training time increases up to200%when the capability of the training set raises from600images to840images. The paper uses2DPCA algorithm instead of PCA because of the reason upon. Finally, the paper uses ORL face database and FERET face database as a test, and proves that the GPU-based2DPCA face recognition algorithm, when compared to serial PCA algorithm both using one computer, can keep the recognition rate, raise recognition speed, and adapt large scale of training set.