A new saliency detection method for stereoscopic images using contrast and prior knowledge

Abstract

In this paper, we propose a new visual saliency detection method, which is effective regardless of unreliable disparity information, by using contrast and prior knowledge. Our proposed method consists of two phases. In the first phase, we used region based contrast information to compute the saliency of an input image. We consider not only global but also local contrast in color and disparity information to efficiently extract salient regions in a stereoscopic image. In addition, we introduce a confidence measure to handle unreliable disparity information. In the second phase, we used region based prior knowledge existent in a stereoscopic image. The region based prior knowledge is constructed from low-level features such as color, frequency, location and disparity in the stereoscopic image. Finally, we integrate contrast-based and prior knowledge- based saliency to accurately detect saliency from input stereoscopic image. Experimental results show that our method efficiently detects salient contents in stereoscopic images.