Saliency Detection: A Divisive Normalization Approach

Abstract

Saliency detection for images has become a valuable tool in applications like object segmentation, adaptive compression, and object recognition. In this paper, we propose a method for saliency detection that outputs full resolution saliency maps of the input images. The key idea is to exploit a computational process of divisive normalization that simulates the similar feature suppression in human primary visual cortex, and thereby is capable of generating visual saliency. The method, which only employs low-level features of color and luminance, is simple and computationally efficient. We compare our method with five state-of-the-art saliency detection algorithms by use of psychophysical patterns and natural images. Experimental results show that our method outperforms these five algorithms both on the psychophysical ground-truth evaluation and on the eye fixations prediction task.