An underwater organisms recognition method based on machine vision in complex marine environment

To meet the requirements for parallelity and accuracy in marine biosensing, this article proposes an improved YOLOv5 algorithm based on lightweight enhanced networks. Before training improved models, UWCNN algorithms enhanced underwater target images to solve problems such as color deviation, image noise and image vagueness.In improving the YOLOv5 algorithm, this paper introduced, first, the Swin-Transformer main core module to improve the model’s generalization capabilities; secondly, the use of the EMA structure in the head prediction section and the introduction of the GAM attention mechanism in the main core to enhance the robustness of the model; and finally, the introducation of Focal-EIOU Loss for precision boundary frame regression efficient losses.The results showed that the detection speed improved by 2 points compared to the original YOLOv5 algorithm, with the AP of sea cucumber, sea urchins, scallops, and starfish increased by 14 % , 1 % ,5 % and 5 % respectively, and the mAP increased 6.26 % . Furthermore, the fps value has increased to 38.86 phases per second.The method is directly targeted at detection of shallow-sea organisms and can provide useful reference to the intelligent equipment of underwater robots.