1.College of Information Engineering, Liaoning Provincial Key Laboratory of Marine Information Technology,Dalian Ocean University, Dalian 116023, China; 2.Key Laboratory of Environment Controlled Aquaculture(Dalian Ocean University), Ministry of Education, Dalian 116023, China
Abstract: In order to solve the problem of low accuracy of fish detection caused by underwater imaging blur and distortion in actual aquaculture environment, a fish detection method(SK-YOLOv5)combining visual attention mechanism SKNet(selective kernel networks)and YOLOv5(you only look once)is proposed.In this method, UNet(convolutional networks for biomedical image segmentation)is firstly used to preprocess images to obtain clear fish images, and then SKNet is fused to Backbone end of YOLOv5 to form feature extraction network focusing on pixel-level information to strengthen the recognition ability of fuzzy fish.In this study, ablation test and model comparison test were carried out on underwater fuzzy fish swarming image data se to verify the effectiveness of SK-YOLOv5.The results showed that SK-YOLOv5 was effective in fish swarm detection task, and had recognition accuracy of 98.86% and recall rate of 96.64%, 2.14% higher and 2.29% higher compared with YOLOv5, respectively.Compared with XFishHmMp and FERNet with the maximal detection accuracy underwater target detection model, SK-YOLOv5 had the best detection effect, 5.39% higher in recognition accuracy, and 5.66% higher recall rate, and compared with FERNet, the recognition accuracy was improved by 3.59% and recall rate by 3.77%.The findings indicated that the fish detection of fusing SKNet and YOLOv5 can effectively enhence the identification ability of fuzzy fish, and improve the overall effect of fish detection and recognition.