Salient Object Detection Using Reciprocal Learning

Abstract

Typically, Objects with the same semantics may not always stand out in images with diverse backgrounds. Therefore, accurate salient object detection depends on considering both the foreground and background. Motivated by this observation, we proposed a novel reciprocal mechanism that considers the mutual relationships between background and foreground for salient object detection. First, we design a patulous U-shape framework comprising a shared encoder branch and two parallel decoder branches for extracting the foreground and background responses, respectively. Second, we propose a novel reciprocal feature interaction (RFI) module for the two decoder branches, allowing them to learn necessary information from each other adaptively. The RFI module primarily consists of a reciprocal transformer (RT) block that utilizes modulated window-based multi-head cross-attention (MW-MCA) to capture mutual dependencies between elements of the foreground and background features within the current two windows. Through the RFI module, the two decoder branches can mutually benefit each other and generate more discriminative foreground and background features. Additionally, we introduce a cooperative loss (CL) to guide the learning of foreground and background branches, which encourages our network to obtain more accurate predictions with clear boundaries and less uncertain areas. Finally, a simple but effective fusion strategy is utilized to produce the final saliency map. Extensive experiments on five benchmark datasets demonstrate the significant superiority of our method over the state-of-the-art approaches.