Abstract
This paper describes a new method by which images can be recognized robustly in the presence of random perturbations to the images. In general, a computer could recognize images by the following procedure: (1) Calculate a cross-correlation coefficient between an unknown image and each of images memorized as standards. (2) Find the memorized standard image which maximizes a maximum value of the cross-correlation coefficient. (3) Recognize that the input image belongs to a class of the found standard image. Although this conventional method is effective if imaging environments are calibrated, it does not always work well under a natural environment in the presence of random perturbations; therefore, a problem to be solved next is how to construct a method for the robust image recognition under the natural environment.
To solve the problem in case that input images have a texture region constructed by a partial image i.e. a texel, we introduce, as a measure suitable for the robust image recognition, local point symmetry inherent in the cross-correlation coefficient between the input image and the partial image. We show that a disorder measure for the local point symmetry of the cross-correlation coefficient is robust against random perturbations; then, we propose a method for robust image recognition with the disorder measure.
We have made experiments on recognizing input images distorted with random perturbations. Their results have shown that the input images can be recognized by our method more robustly than those done by the conventional method.
