Stereoscopic Image Saliency Detection Optimization: A Multi-Cue-Driven Approach

Abstract

Numerous stereoscopic image saliency detection algorithms have been presented to detect the salient objects in a stereoscopic image. However, they typically fail to uniformly highlight all the objects when the image contains multiple objects or complex backgrounds. In this paper, we propose a multi-cue-driven optimization (MCDO) for stereoscopic image saliency detection. MCDO leverages multiple cues, including depth, color, and spatial position to optimize the saliency maps generated by existing saliency detection algorithms. Fully connected conditional random field is used to integrate the depth, color, and spatial cues from the input stereoscopic image to ensure that pixels with similar depth, color, and/or spatial position have similar saliency values. Compared with original saliency maps, the optimized saliency maps have more uniformly highlighted salient objects, whose boundaries are more precise, and fewer incorrectly detected background regions. The experimental results on three datasets demonstrate that the proposed MCDO method can effectively improve the performance of stereoscopic and 2-D image saliency detection algorithms.

Keywords

• Conditional random field
• depth information
• multi-cue
• saliency detection
• stereoscopic images