Dynamic selection fusion network for RGB-D salient object detection

Abstract

RGB-D salient object detection (SOD) has garnered increasing attention due to the introduction of depth cues. While RGB images and depth maps have inherent differences, they possess strong semantic consistency in representing objects. However, existing methods often fail to consider the consistent representation of the two modalities when exploring cross-modal interaction. To this end, we propose a novel dynamic selection fusion network (DSFNet) for RGB-D SOD by incorporating cross-modal consistency enhancement and cascade feature refinement. Specifically, in the encoding phase, a cross-modal consistency enhancement (CMCE) module is introduced to dynamically enhance both interaction consistency and semantic consistency representations through attention guidance. In the decoding phase, a cascade feature refinement (CFR) module is designed to progressively refine segmentation results from coarse to fine by incorporating multi-scale pooling and enhancement operations. Experimental evaluation shows that our DSFNet outperforms 14 current state-of-the-art RGB-D SOD methods.