Research On Improved Design Of Correlation Filters In Correlation Pattern Recognition

As an important field of machine learning and artificial intelligence,pattern recognition has been widely used in many fields such as biometrics and image retrieval.Correlation pattern recognition is an important sub-area of pattern recognition,using correlation filters for classification.Although correlation filters have some excellent properties and many variants,there are still problems that the correlation peak is not sharp enough,the required training time is long,and the existence of aliasing.In view of the above problems,this paper mainly studies the improvement of quadratic correlation filter algorithm and zero-aliasing correlation filter algorithm.First,review the general design principles of correlation filters and introduce the excellent characteristics of correlation filters through experiments.Then,quadratic correlation filter is improved using the idea of prewhitening the images in constrained correlation filter and unconstrained correlation filter,so that sharp correlation peak is generated and the recognition performance is improved.For the problem that the large matrix operation is needed for the training of quadratic correlation filter,the fact that quadratic correlation filter only needs the eigenvectors corresponding to the non-zero eigenvalues is used to simplify matrix operation and speed up the training process.The experimental results on the AT&T face database and the Yale face database verify the effectiveness of the proposed method.Finally,zero-aliasing correlation filter is trained using gradient descent method,in order to overcome the problem that the gradient descent method may have zigzag path and slow convergence speed,the nonlinear conjugate gradient method is used instead of gradient descent method to improve training process.The experimental results on the AT&T face database and the Yale face database show that the improved training method can effectively accelerate the convergence speed while maintaining the recognition performance of the zero-aliasing correlation filter.