A Real-Time Underwater Detector Based on YOLOv5 Using Forward-Looking Sonar Images

Abstract

Underwater detection plays a crucial role in various applications such as marine exploration, underwater robotics, and environmental monitoring. In recent years, forward-looking sonar images have gained significant attention due to their ability to provide real-time, high-resolution imaging of the underwater environment. However, real-time detection and low detection precision remain significant challenges in underwater detection using forward-looking sonar images. In this paper, to address the limitations of existing methods by significantly improving mean Average Precision (mAP) and achieving a faster detection rate for underwater objects, we first introduced noises into the samples to augment the data in UATD, a publicly available sonar dataset. Then we proposed a novel real-time underwater detector called Fast-YOLOv5, which combines the efficiency of MobileNetV2 and the powerful feature extraction capabilities of the self-attention mechanism SENet based on YOLOv5. The experimental results on UATD demonstrate a 2.65% improvement in mAP, a 162.12% increase in detection speed for the proposed detector, and a 37.7% decrease in model parameters.