Multivariate Filter for Saliency

Abstract

A method for analytical computing of finding objects of interest in images has been developed with multivariate normal distribution to improve the visual capability. Key requirement is the ability of significantly shifting attention to image region that is texture based in general case of real images. The visual attention evaluation is mainly involved for initial task of most visual applications including segmentation, gaze tracking and image re-targeting. To enhance the accuracy of saliency detection, we have to analyze the salient distinction of textured region by combining several techniques. As an initial step, the multivariate filters are designed for estimating local texture feature that is rotation invariant. Significant distinction of patches is then calculated to describe the possible interest regions. The final morphological operations bring fixation of objects of interest. On a test set which consists of ten thousands of images in several themes, the method provides a precision of 92%, recall of 83% and F-measure of 86%.