Regression method for inverse correlation filters design for objects recognition

Abstract

The problem of inverse correlation filters design to recognize a set of objects is considered as the problem of regression parameters estimation on the base of input data arrays and desirable response. The data and response should be processes with zero mean to consider this problem as evaluation of regression parameters. The problem is solved using the least squares method with regularization. The regularization is optimized to achieve high resolution of the filters in conjunction with capture` broad band of objects given by a set of templates. The least squares method is using in the terms of singular value decomposition that made it possible to linearize the nonlinear ridge regression optimization problem. The methods to false recognitions elimination are considered, It was shown that the regression approach gives additional condition to recognize classes of objects. This allows to have more high accuracy in recognition of desired objects on a foreign background in comparison with other correlation filters types.