Improving underwater camouflage object segmentation with dual-decoder attention network

Underwater Camouflaged Object Segmentation is to segment the camouflaged object from the background in underwater images. Compared to generic objects, camouflaged objects are difficult to segment due to their high similarity with the surrounding environment. This paper presents a new convolutional neural network called Dual-Decoder Attention Network (DDANet) to address this problem. Inspired by the visual process of identifying camouflaged objects, DDANet adopts the design of dual-decoder, introduces non-local attention mechanisms, and optimizes structures such as global and edge guidance bypass to achieve better segmentation performance. The experiments demonstrate that DDANet outperforms select relevant methods in multiple metrics on different camouflaged object datasets. The paper also introduces a new composite underwater camouflage object dataset called Aqua-Test and reports DDANet’s competitive performance on this dataset. These results demonstrate the effectiveness of DDANet in identifying underwater camouflage objects.