Boundary-Aware Distracted Attention Network for Camouflaged Object Detection

Abstract

Camouflaged object detection (COD) involves precisely segmenting objects, which are seamlessly blended into the surroundings, due to the similarities between the surroundings and them in color, texture, and other camouflaging techniques. To address the challenges that such similarities pose, distraction mining approaches have been proposed. However, they normally struggle with boundary delineation and extraction of relevant features. To get rid of this dilemma, we introduce a boundary-aware distracted attention network (BADANet), which exploits both the boundary cue and the distraction mining strategy. Specifically, we first use a pretrained network as the encoder for feature extraction. Then the features extracted are sent to a boundary shrinking module (BSM) that we design. After the output is processed by a multiple dense atrous spatial pyramid pooling (MDASPP) module, a boundary map is produced. Given an encoder block, the features extracted are also fed into a multibranch bidirectional fusion block (MBBFB), which performs the bidirectional fusion at the channel dimension, followed by the multiple fusion conducted at the spatial dimension via an individual MDASPP module. We further propose a boundary-aware distracted attention (BADA) block, which receives both the features fused and the boundary map. With regard to the encoder blocks, a series of BADA blocks are comprised of a distracted attention decoder. Finally, the detection mask is generated by the last BADA block. We have conducted a series of experiments on four popular COD datasets. Experimental results demonstrate that the proposed BADANet normally outperforms 18 baseline methods. These promising results should be due to the BADA mechanism that we design.