Research on Underwater Object Detection Algorithm Based on Improved YOLOv5

Abstract

To address the complexities of underwater environments, poor imaging conditions, the challenges of underwater object detection, slow detection speeds, underwater target density, and underwater biodiversity, a YOLOv5-based underwater object detection method is proposed. To meet the performance and efficiency requirements of underwater robots, the ShearC3 module is introduced, which improves the YOLOv5 C3 module to reduce model parameters and network layers, thus enhancing detection speed and focusing on the most effective receptive field’s central part. To tackle the issue of low-scale information loss, the PixelShuffle upsampling module is used to integrate channel dimension information for generating high-resolution images. Through multiple dataset experiments, the improved model achieves a detection accuracy of 84.3%, a 1.2% improvement over the original model, while also increasing the detection speed by 10.16%, meeting the operational needs of underwater robots.