Download This PaperAdvanced Underwater Object Detection in Sonar Images: An Exploration Utilizing Innovative Detector and Transfer Learning
14 Pages Posted: 13 Nov 2024
Abstract
The escalating human exploitation and utilization of oceans have propelled underwater detection technology into the forefront of academic and industrial research. Among these, sonar imagery for underwater target detection has emerged as a crucial field, offering unparalleled benefits in low-visibility conditions compared to optical imagery. However, accurate underwater target detection faces challenges due to the scarcity of satisfactory sonar image samples and the inherent low pixel ratio in foreground target areas. To address these constraints, this paper proposes an innovative transfer learning strategy employing a customized You Only Look Once (YOLO) model, underwater sonar (UWS)-YOLO. The approach mitigates overfitting in small datasets by transferring features from underwater optical images to sonar images, resulting in a substantial boost in detection accuracy. UWS-YOLO is designed to tackle blurred small targets within low-resolution sonar imagery, incorporating an orthogonal channel attention mechanism to augment attention to fine details and a dynamic snake convolution module to enhance recognition of tubular structures. Empirical evaluations demonstrate the efficacy of the proposed approach, with UWS-YOLO outperforming contemporary state-of-the-art object detection frameworks by a remarkable 3.5% enhancement in mean Average Precisio (mAP)@0.5 while maintaining modest computational complexity and enabling real-time performance. These advancements hold significant potential to revolutionize applications such as maritime search and rescue operations and underwater detection systems.
Keywords: Underwater Object Detection, Sonar imagery, Transfer Learning, Orthogonal channel attention, Dynamic snake convolution.
Suggested Citation: Suggested Citation