DBN-YOLO: A Dynamic Sparse Deformable Instance Segmentation Method for Sonar Images

Underwater sonar image instance segmentation is a crucial component for tasks such as marine resource exploration, underwater communication and navigation, and marine environmental monitoring. Existing sonar image segmentation models primarily rely on traditional machine learning methods, which face challenges in adapting to the complex characteristics of underwater sonar images. Underwater sonar image targets often exhibit complex scale variations and geometric deformations, which can pose challenges for segmentation models to accurately capture the details of all targets, especially for small objects. Additionally, these targets may possess low contrast and blurry edges, leading to foreground and background mixing during the training process, resulting in the loss of target features. Moreover, there is a class imbalance issue between the target and background categories, which can lead to the generation of blurry or inaccurate segmentation results by the segmentation models. To address the aforementioned issues, this paper proposes DBN-YOLO: a dynamic sparse attention-based deformable underwater sonar image instance segmentation model, built upon the existing YOLOv5-seg instance segmentation model. Firstly, the Deformable Convolution Network (DCN) is introduced. By incorporating learnable offsets, DCN can adaptively adjust the sampling positions of convolution kernels to better accommodate deformations and geometric variations of the targets. This helps improve the segmentation model's ability to accurately capture object boundaries, thereby enhancing the segmentation effectiveness. Next, the BiFormer attention module is introduced. It employs a dual-layer routing attention module to compute key regions at a coarse granularity and perform attention interaction at a fine granularity. This enhances the foreground weights and suppresses the background weights, enabling the model to focus more on the targets and improve segmentation accuracy and effectiveness. Lastly, the Normalized Wasserstein Distance (NWD) loss is introduced to alleviate the sensitivity of the original Intersection over Union (IOU) metric to positional deviations of small targets and mitigate the impact of imbalanced positive and negative samples on target segmentation. The experimental results illustrate that the proposed DBN-YOLO mode improves the recall rate by 3.9% and the mAP@0.5 by 3.1% on the SCTD-seg dataset.