基于轻量级改进的YOLOv8水下目标检测模型
Lightweight improved YOLOv8 based underwater target detection model
周志耀 1马常霞 1杨丽莎 1仲兆满 1胡文彬1
作者信息
- 1. 江苏海洋大学计算机工程学院 连云港 222006
- 折叠
摘要
在恶劣和多变的水下环境中工作的设备是进行水下研究和开发的基本保障.现阶段的水下目标检测模型参数量和计算量过大,在资源有限的水下设备上部署受限.为解决水下检测模型参数量和计算量过大问题,提出一种轻量级的水下目标检测模型RCE-YOLO.首先,利用RFAConv的空间注意力权重来改进CBS处理接受域信息的能力和提升C2f对空间特征信息融合的能力,增强模型对小密集目标的检出能力.其次,融合CCFM与Dysample模块,该融合模块能够更有效的利用不同尺度信息并通过内部的点采样方法减少原先采样产生的模糊和失真.最后,在SPPF前向传播过程中融合高效多尺度注意力机制,该机制使得模型重点关注水下目标关键信息,降低误检率和错检率.实验结果表明,改进的轻量级模型在数据集DUO上进行验证,mAP50、mAP50:90值分别达到83.6%、64.2%,相较于YOLOv8基准模型 mAP50、mAP50:90值分别提升了1.4%、1.2%,参数量和计算量分别下降了32.3%、0.9 G.相较于其他目标检测模型满足了恶劣多变环境下的水下目标检测需求,为水下设备轻量级部署奠定基础.
Abstract
Equipment that operates in harsh and variable underwater environments is essential for conducting underwater research and development.The current underwater target detection models are too large in parameter count and computation,which limits the deployment of underwater equipment with limited resources.In order to solve the problem of excessive parameter count and computational volume of underwater detection models,a lightweight underwater target detection model RCE-YOLO is proposed.Firstly,the spatial attention weights of RFAConv are utilized to improve the ability of CBS to process the information in the receptive domain and to enhance the ability of C2f to fuse spatial feature information,so as to enhance the model's ability of detecting small and dense targets.Second,the CCFM is fused with the Dysample module,which is able to utilize the different scale information more effectively and reduce the blurring and distortion produced by the original sampling through the internal point sampling method.Finally,the Efficient multi-scale attention mechanism is fused in the SPPF forward propagation process,which makes the model focus on the key information of the underwater target and reduces the false detection rate and misdetection rate.The experimental results show that the improved lightweight model is validated on the dataset DUO,and the mAP50 and mAP50:90 values reach 83.6%and 64.2%,respectively,which are 1.4%and 1.2%higher compared to the mAP50 and mAP50:90 values of the benchmark model of YOLOv8,and the number of parameters and the amount of computation drop by 32.3%and 0.9 G,respectively.compared to other The target detection model meets the needs of underwater target detection in harsh and variable environments,and lays the foundation for lightweight deployment of underwater equipment.
关键词
水下目标检测/轻量化/RFA卷积模块/CCFM模块/注意力机制Key words
underwater target detection/lightweighting/RFAConv module/CCFM module/attention mechanism引用本文复制引用
出版年
2024
NETL
NSTL
万方数据