UNet-Based YOLO for Underwater Object Detection

Vision-guided autonomous underwater vehicles (AUVs) are becoming crucial tools for deep-sea exploration, resource assessment, and mariculture, among other underwater tasks. However, the challenging underwater environment often leads to image distortion and quality degradation, which negatively impacts the accuracy of vision systems and hinders effective target detection. Recently, the advancement of deep learning, particularly through Transformer-based architectures, has opened new avenues for improving underwater image enhancement. However, applying Transformer models directly for this task faces challenges, including the difficulty of capturing global information at the pixel level while preserving local details. In this paper, we introduce an innovative approach that combines the U-Net network with YOLOv5 for enhanced underwater target detection. Firstly, we construct an improved U-Net model, which can better correct color distortion in global and local features, and overcome the limitations of Transformer in extracting details and textures through convolutional neural networks. Next, YOLOv5 is used to detect the target of the enhanced image, and the quality of information extraction at high and low resolutions is improved through multi-level feature fusion, thus strengthening the attention and correction of severely degraded areas in the image. The experimental results show that our method has obvious advantages in subjective visual perception and comprehensive performance evaluation compared with the existing technology, and shows excellent learning and generalization ability.