Recursive feedback-based feature refinement network for camouflaged object detection

Abstract

Accurate segmentation of camouflaged objects from camouflaged images is a valuable but challenging task. Although some achievements have been made in camouflaged object detection, several challenges remain. Camouflaged objects, occluded by the foreground or embedded in the background, exhibit unclear boundaries and high similarity with their surroundings, especially when multiple objects are involved. To confront these challenges, a novel Recursive Feedback-based Feature Refinement Network (RFF-Net) is proposed for camouflaged object detection. Specifically, a boundary detection network is proposed first, which is used to generate boundary-aware features. Then, the extracted boundary features are integrated into the segmentation network. In particular, we design a segmentation network that utilizes recursive feedback from higher-level features to guide lower-level features, thereby gradually refining the segmentation results for camouflaged objects. Furthermore, a Boundary Feature Aggregation Module (BFAM) is proposed to fuse boundary information with multi-level features, which can enhance the multi-level backbone features to generate finer segmentation results. Further, an Interclass Discrepancy Enhancement Module (IDEM) is proposed to amplify the interclass differences between a camouflaged object and its surroundings, which can make full use of multi-scale and contextual information to highlight the location of the camouflaged object. Quantitative and qualitative experiments on four challenging benchmark datasets prove the outperformance of our RFF-Net compared to various state-of-the-art camouflaged object segmentation models.