基于GS-YOLOv8的轻量化水下生物目标检测算法

A Lightweight Underwater Biological Target Detection Algorithm Based on GS-YOLOv8

摘要:

为解决现有的水下生物目标检测模型参数过多,难以部署到资源有限的移动端的问题,提出了一种基于GS-YOLOv8的轻量化水下生物目标检测模型；该模型基于YOLOv8s模型进行改进,设计一种轻量化的RepHGNetV2网络作为YOLOv8s的主干网络,以降低模型的计算复杂度和参数量；使用轻量化卷积GSConv替换颈部网络中所有的标准卷积,进一步减少模型参数,提高检测性能；引入设计的C2fAK模块,使模型能够更好地适应不同形状和大小的水下生物目标, 从而提高检测精度；实验结果显示,在URPC2020数据集上,与原模型YOLOv8s相比,改进后的GS-YOLOv8网络模型的参数量降低了37.7%,计算量降低了27.8%, mAP@0.5提高了0.9%；此外,与目前较为先进的YOLOv10模型相比,改进后的GS-YOLOv8模型在检测精度和轻量化方面更有优势。

Abstract:

In order to solve the problem that the existing underwater biological target detection model has too many parameters and is difficult to deploy to the mobile terminal with limited resources, a lightweight underwater biological target detection model based on GS-YOLOv8 is proposed. The model is improved based on YOLOv8s model, and a lightweight RepHGNetV2 network is designed as the backbone network of YOLOv8s to reduce the computational complexity and the number of parameters of the model. Lightweight convolutional GSConv is used to replace all standard convolution in the neck network, further reducing model parameters and improving detection performance. The C2fAK module is introduced to make the model better adapt to different shapes and sizes of underwater biological targets, so as to improve the detection accuracy. Experimental results show that in the URPC2020 dataset, compared with the original model YOLOv8s, parameters of the improved GS-YOLOv8 network model is reduced by 37.7%, computation is reduced by 27.8%, and mAP@0.5 is increased by 0.9%. In addition, compared with the current more advanced YOLOv10 model, the improved GS-YOLOv8 model has more advantages in detection accuracy and lightweight.