Underwater Detection using Forward-Looking Sonar Images based on Deformable Convolution YOLOv3

Underwater target detection is crucial for the exploration and utilization of marine resources. The utilization of forward-looking sonar has become prevalent as an active sonar device for underwater perception, making target recognition based on forward-looking sonar images a subject of increasing interest. However, the complexity and variability of underwater scenes, coupled with deformations that occur during sonar imaging due to changes in the field of view pose significant challenges. This paper addresses these challenges by proposing the YOLO-DCN network, a fusion of YOLOv3 with Deformable Convolutions. These additions aim to specifically handle the deformations observed in sonar images. To enhance the robustness of underwater target detection, we employ an effective data augmentation method to address the variable nature of underwater scenes. In evaluating the performance improvement of our proposed YOLO-DCN network, we conducted experiments using the publicly available forward-looking sonar image dataset, UATD. The results reveal a notable enhancement in overall performance, with a 1.4% increase in mean Average Precision (mAP) across the ten target classes. Particularly significant is the substantial 10.2% improvement observed in the detection of irregularly shaped targets, such as the human body. This underscores the adaptability and effectiveness of our model in addressing the challenges posed by underwater sonar imaging, especially in scenarios involving complex and non-standard target shapes.