Abstract
Aquaculture is of great significance to economic development. It is assessed by manual periodic sampling traditionally, consumes workforce and material resources, and quickly leads to inadequate supervision, which results in substantial property losses. Fish target detection technology can effectively solve the issue of manual monitoring. However, a majority of current studies are based on ideal underwater environments and are inapplicable to complex underwater aquaculture scenarios. Therefore, the YOLOv8n-DDSW fish target detection algorithm was proposed in this article to resolve the detection difficulties resulting from fish occlusion, deformation and detail loss in complex intensive aquaculture scenarios. (1) The C2f-deformable convolutional network (DCN) module is proposed to take the place of the C2f module in the YOLOv8n backbone to raise the detection accuracy of irregular fish targets. (2) The dual-pooling squeeze-and-excitation (DPSE) attention mechanism is put forward and integrated into the YOLOv8n neck network to reinforce the features of the visible parts of the occluded fish target. (3) Small detection is introduced to make the network more capable of sensing small targets and improving recall. (4) Wise intersection over union (IOU) rather than the original loss function is used for improving the bounding box regression performance of the network. Training and testing are based on the publicly available Kaggle dataset. According to the experimental results, the mAP50, precision (P), recall (R) and mAP50-95 values of the improved algorithm are 3.9%, 3.7%, 6.1%, and 7.7% higher than those of the original YOLOv8n algorithm, respectively. Thus, the algorithm is effective in solving low detection accuracy in intensive aquaculture scenarios and theoretically supports the intelligent and modern development of fisheries.