Three-Stage Distortion-Driven Enhancement Network for Forward-Looking Sonar Image Segmentation

Abstract

Forward-looking sonar (FLS) image segmentation aims to accurately locate underwater objects, providing essential support for marine engineering. Unlike natural optical images, FLS images have weak semantic content and complex background, posing segmentation challenging for existing models. To address this issue, we propose a novel FLS image segmentation model, the three-stage distortion-driven enhancement network (TDENet), equipped with an extended IoU loss. TDENet employs a three-stage distortion-driven feature processing strategy. Specifically, we propose the three-stage distortion-driven module (TDM), which consists of three stages (i.e., feature distortion, enhancement, and fusion). First, feature distortion introduces dynamic inputs, compelling the model to learn patterns in a distorted feature space, thereby enhancing robustness. In the enhancement stage, global and local interactions between distorted and undistorted features improve the model’s semantic comprehension and detail retention. Finally, feature fusion ensures a comprehensive representation. To refine segmentation maps, we propose the centroid deviation intersection over union loss (CD-IoU loss), incorporating a CD term to the vanilla IoU loss. This term measures the distance and directional discrepancies between the centroids of predicted and ground truth object regions, quantifying their detail differences. Equipped with the CD-IoU loss, our TDENet can better capture details. Extensive experiments on the public Marine Debris Dataset demonstrate that our TDENet outperforms 16 state-of-the-art models.