A Division Normalization Method for Saliency Detection

Since most existing approaches of saliency detection output low-resolution and biologically implausible saliency maps, in this paper, we propose a division normalization method to solve these problems. Our method is biologically motivated and is capable of producing a full resolution saliency map. Firstly, we decompose the input image from L^*a^*b^* color space into five feature channels: green, red, blue, yellow and luminance. Then, we normalize the channels by use of their respective energy. Next, we integrate five normalized channels in the L^*a^*b^* color space. Finally, we employ the Euclidean norm to compute the saliency map. This procedure simulates the same feature suppression in primary visual cortex of humans. Experimental results show that the method is simple and computationally efficient, and outperforms other conventional approaches of saliency detection both on the psychophysical pattern test and on the eye fixations prediction tasks.