Underwater Object Detection from Sonar Images Using Transfer Learning

Underwater systems are always prone to complex environmental conditions and various interferences that is the reason why Sonar images are filled up with contamination, noise, and blurriness. Collection of real-world underwater images data is very costly with respect to human resources, time, equipment and logistics. Due to these challenges, there are very few publicly available datasets of underwater sonar images and only few datasets are available for underwater images which are scarce because of lack of resources. Precise and fast detection of objects in underwater sonar images is very less explored and still an open challenge for underwater systems. This paper evaluates performance of state-of-the-art object detection deep learning network YOLO (You Only Look Once) for underwater sonar images through transfer learning on publicly available benchmark Marine Debris dataset and custom-tailored dataset captured in real sea environment. Experimental results shows that YOLOv8n was able to achieve high mAP score of 95.3% on Marine Debris dataset with inference speed of 1.3ms.