Abstract
Sonar is an efficient and indispensable tool for exploring underwater environments under challenging conditions. The research has a crucial impact on domains such as marine exploration, environmental monitoring, defence, archaeological discoveries, resource exploration, and climate research. Extracting valuable insights from sonar images and detecting objects within obscure sonar images is tedious and challenging for both human experts and conventional machine learning models. This study explores state-of-the-art deep learning methods for enhanced underwater object detection in multibeam forward-looking sonar images, to determine optimal trade-off model between accuracy and complexity. A sequence of preprocessing steps is proposed in this work to mitigate noise and enhance images, improving the accuracy of benchmark object detection performance. This comprehensive framework for forwardlooking sonar images integrates the preprocessing techniques and target detection thereby enhancing target visualization. The model localizes and predicts the target of each class by overcoming the challenges of target detecting in hazy images and imbalanced class distribution. The proposed approach yields a mAP of 96.3% with 3M parameters which implies a significant increase in efficiency in real time processing in comparison to other models. From the analysis, the proposed framework improves visual perception and enhances object localization of the targets in the sonar images.
Keywords: Deep Learning, Forward-Looking Sonar, Object Detection, Preprocessing, Underwater, YOLO.