LS-DETR: Lightweight Transformer for Object Detection in Forward-Looking Sonar Images

Abstract

A transformer-based end-to-end detection method lightweight sonar detection transformer (LS-DETR) is proposed, which is specifically tailored for enhancing detection accuracy in forward-looking sonar images while significantly reducing the computational load. Despite the challenges posed by the complexity of underwater environments that have led to suboptimal detection performance and the lack of lightweight optimization for underwater devices, LS-DETR addresses these issues effectively. In LS-DETR, the backbone employs a newly proposed lightweight-gated attention block (LGABlock), which reduces computational redundancy through low-complexity convolutions and gated attention. A lightweight hybrid encoder (LHE) is designed to facilitate scale-internal feature interaction and optimize the feature fusion approach. Furthermore, wise complete IoU (WCIoU)-aware query selection is proposed and integrated with NWDLoss in the decoder, enabling the scores to integrate classification and positional information while focusing on the small targets. Results demonstrate that on the multibeam forward-looking sonar dataset UATD, LS-DETR achieved a 2.8% increase in accuracy and a 31.5% reduction in parameter count, proving the effectiveness and superiority.