DBD-Net: Dual-Branch Decoder Network with a Multiscale Cascaded Residual Module for Ship Segmentation

Abstract

The segmentation of visible ship images is an important part of intelligent ship monitoring systems. However, this task is faced with many difficulties in practical applications, such as complex background environments, variations in illumination, and target scale changes. In view of these situations, we present a dual-branch decoder network with a multiscale cascaded residual module for ship segmentation. Specifically, in the encoding stage, we introduce a multiscale cascaded residual module as a replacement for traditional convolution layers. By leveraging its multiscale architecture, the module effectively captures both the global context and fine-grained details. In the decoding phase, our framework incorporates two parallel branches, both of which utilize the cascading residual module to enhance feature extraction and representation. Additionally, one of the branches is equipped with spatial attention and channel attention mechanisms. Finally, comprehensive tests were conducted on the publicly available ship datasets MariBoatsSubclass and SeaShipsSeg. Our proposed network achieved impressive results, with Dice, Recall, Mcc, and Jaccard scores of 0.9003, 0.9105, 0.8706, and 0.8197 on the MariBoatsSubclass dataset. Similarly, it demonstrated outstanding performance on the SeaShipsSeg dataset, attaining Dice, Recall, Mcc, and Jaccard scores of 0.9538, 0.9501, 0.9519, and 0.9129, respectively. These results highlight the superior accuracy and robustness of DBD-Net in segmenting and detecting ships across diverse scenarios and datasets.