An Improved Underwater Object Detection Algorithm Based on YOLOv5 for Blurry Images

High-precision underwater object detection technology has important research value and broad application prospects in marine biological resources exploration and marine fisheries monitoring. However, blurry images pose great challenges to the detection of underwater objects of different sizes. The detection accuracy of underwater objects of different sizes in blurry images needs to be improved significantly. To this end, we propose a convolutional neural network based on YOLOv5 named YOLOv5-MobileOne-Attention (YOLOv5-MA) in this work. In YOLOv5-MA, we replace the backbone of YOLOv5 with MobileOne to improve the accuracy of the algorithm, and add the convolutional block attention module and the normalization-based attention module to enhance the feature extraction capability for underwater blurry images. Furthermore, we add the normalized gaussian wassernstein distance loss function based on the location loss function to reduce the sensitivity of the anchor frame of small objects to the intersection over union values. In the experiment, we investigate the performance of YOLOv5-MA using a dataset containing blurry images selected from the underwater robot professional contest 2020 dataset and the real-world underwater image enhancement dataset. Experimental results show that the mAP_0.5 value of YOLOv5-MA reaches 54.4%, which is 2.8% higher than that of YOLOv5s. Moreover, YOLOv5-MA can achieve a balance between accuracy and speed compared to YOLOv5, YOLOv8, and RE-DETR. This validates the advantages of YOLOv5-MA in underwater object detection for blurry images. It has great application prospect in future object detection based on underwater robots in complex underwater environments.