TRAM: Transformer-Based Mask R-CNN Framework for Underwater Object Detection in Side-Scan Sonar Data

Accurate detection and segmentation of underwater objects in side-scan sonar (SSS) imagery remain challenging due to noise, cluttered backgrounds, and low-contrast conditions. In this paper, a novel TRAnsformer-based Mask r-cnn (TRAM) framework is proposed for robust underwater object detection and segmentation. TRAM consists of three progressively enhanced models-TRAM V1, V2, and V3-built on a modified Swin Transformer backbone with multi-scale feature extraction via Feature Pyramid Networks. TRAM V2 integrates a CBAM, Convolutional Block Attention Module, for refined feature learning, while TRAM V3 introduces CLAHE, Contrast Limited Adaptive Histogram Equalization, preprocessing to enhance contrast and improve training stability. TRAM is benchmarked against YOLOv5, YOLOv9, and Mask R-CNN with ViT and Swin Transformer backbones. Results show that TRAM V3 achieves the highest performance, with a mAP@0.5 of 0.8896 for detection and 0.7626 for segmentation, outperforming all baseline models.