Research On Parallel Face Recognition Based On OpenCL Acceleration

With the acceleration of social information process and the rapid popularization of artificial intelligence,face recognition technology is widely used in many areas,such as video surveillance,identity recognition and human-machine interaction,because of its directness,friendliness,convenience and non-invasive.With the increasing resolution of image,people have put forward new demands on the speed of face recognition.Generalized face recognition includes face detection and face recognition.In this paper,I study and design a parallel acceleration method of face recognition process based on OpenCL,In this paper,by studying the calculation method of integral Image,the pixel values between rows and between columns can be calculated at the same time.For this reason,the parallel prefix summation algorithm is used in the column direction firstly.Each OpenCL'5 work item uses a vector structure to calculate the partial prefix summation,then prefix summation of the entire column direction is calculated.The prefix summation of row direction can be calculated in the same way.The improved parallel method is more than 12 times faster than the implementation on the CPU.In the process of face detection,the scanning window between different positions on the same scale is completely independent when detecting human faces,and each weak classifier does not interfere with each other.Aparallel face detection method is designed.The data of classifier are needed to be accessed frequently,so the data are stored in constant memory.The acceleration of the whole detection process achieves more than 12 times.In this paper,a method for solving the eigenvalues and feature vectors of a real symmetric matrix by single side Jacobi rotation transformation is studied.According to the fact that each transformation affects only two columns of the matrix,a parallel single side Jacobi rotation transformation is designed to solve the eigenvalues and feature vectors.In the process of solving Jacobi matrix,it ensures that the inner product of the previous column in the indexing queue is bigger than the latter column,so that the final eigenvalues are arranged in descending order.It is convenient for the selection of the principal element during the PCA dimension reduction process and it achieves about 6 times speedup.The face recognition of process based on PCA is a series of matrix operations including the transposition and multiplication of matrices.The covariance matrix of this paper does not solve the transpose of matrix first.Instead,it is obtained by multiplying the column and column of the matrix.The training and recognition stages of face recognition achieve about 22 and 5 times,faster than implemented on CPU.At the end of this paper,a software system of face recognition is designed by combining the parallel method mentioned above.