Saliency detection using parallel non-linear integration of color and gradient using covariances

Visual saliency estimation is an emerging field with various applications specially in computer vision. This paper presents a visual saliency model based on parallel non-linear integration of low level features. Color, gradient and spatial information are the basic features used for saliency computation. At first color and gradient feature matrices are derived from an input image followed by conversion into patches. Secondly these color and gradient feature patches are nonlinearly integrated using covariance. Local average patch dissimilarity is then used to compute saliency and the whole process is repeated with multi size patches. Our approach introduce simple and basic modifications in an existing model and achieve much improved human eye fixation predictions. The main contribution of this paper is the presentation of a parallel approach to integrate color and gradient features for saliency computation. Two of the most popular datasets are used to evaluate the proposed scheme. Our simulations show that the proposed approach performs better than 13 state of the art models on both datasets.