SQnet: An Enhanced Multi-Objective Detection Algorithm in Subaquatic Environments

Abstract

With the development of smart aquaculture, the demand for accuracy for underwater target detection has increased. However, traditional target detection methods have proven to be inefficient and imprecise due to the complexity of underwater environments and the obfuscation of biological features against the underwater environmental background. To address these issues, we proposed a novel algorithm for underwater multi-target detection based on the YOLOv8 architecture, named SQnet. A Dynamic Snake Convolution Network (DSConvNet) module was introduced for tackling the overlap between target organisms and the underwater environmental background. To reduce computational complexity and parameter overhead while maintaining precision, we employed a lightweight context-guided semantic segmentation network (CGNet) model. Furthermore, the information loss and degradation issues arising from indirect interactions between non-adjacent layers were handled by integrating an Asymptotic Feature Pyramid Network (AFPN) model. Experimental results demonstrate that SQnet achieves an mAP@0.5 of 83.3% and 98.9% on the public datasets URPC2020, Aquarium, and the self-compiled dataset ZytLn, respectively. Additionally, its mAP@0.5–0.95 reaches 49.1%, 85.4%, and 84.6%, respectively, surpassing other classical algorithms such as YOLOv7-tiny, YOLOv5s, and YOLOv3-tiny. Compared to the original YOLOv8 model, SQnet boasts a PARM of 2.25 M and consistent GFLOPs of 6.4 G. This article presents a novel approach for the real-time monitoring of fish using mobile devices, paving the way for the further development of intelligent aquaculture in the domain of fisheries.