基于RGHS图像增强和改进YOLOv5网络的黄颡鱼目标提取

Object extraction of yellow catfish based on RGHS image enhancement and improved YOLOv5 network

针对水下能见度不佳,黄颡鱼目标提取精度低、速度慢等问题,提出了基于相对全局直方图拉伸(RGHS)算法和改进YOLOv5的黄颡鱼目标提取模型。首先,为解决光照不均、噪声大等因素带来的图像质量问题,采用RGHS算法对黄颡鱼图像进行亮度增强。然后,在YOLOv5主干网络中引入C3ghost模块和坐标注意力(CA)机制,在颈部网络中用gnConv替换普通卷积,建立改进YOLOv5模型,提升黄颡鱼目标提取精度。结果表明,改进模型的AP值、准确率、召回率比YOLOv5模型分别提升了2.76%、3.16%、3.1%,F1值提升了0.03,所占内存减少了2.3 MB,单张图片推理时间减少了0.001 s。同时,在与YOLOv4、SSD、Faster-RCNN、YOLOx模型的对比实验中,改进模型的AP值分别提升了3.27%、8.63%、2.48%、2.52%。基于RGHS图像增强的改进YOLOv5模型在保证较快速度的情况下,显著提高了黄颡鱼目标提取精度,可为鱼类状态监测方法的研究提供有益参考。

Aiming at the problems of poor underwater visibility,low accuracy and slow speed of object extraction,a yellow catfish object extraction model based on RGHS algorithm and improved YOLOv5 was proposed.Firstly,in order to solve the image quality problems caused by uneven illumination and high noise,RGHS algorithm was used to enhance the brightness of yellow catfish image.Then,C3ghost and CA attention mechanisms were introduced into the YOLOv5 backbone network,and gnConv was used to replace the common convolution in the neck part,so as to establish an improved YOLOv5 model and improve the target extraction accuracy of yellow catfish.The results show that compared with YOLOv5,the AP value,accuracy rate and recall rate of the improved model are increased by 2.76%,3.16% and 3.1 % respectively,the F1 value is increased by 0.03,the memory occupied by the improved model is reduced by 2.3 MB,and the reasoning time of a single image is reduced by 0.001 s.Meanwhile,compared with the YOLOv4,SSD,Faster-RCNN and YOLOx models,the AP values of the improved models are increased by 3.27%,8.63%,2.48% and 2.52% respectively.The improved YOLOv5 model based on RGHS image enhancement can significantly improve the target extraction accuracy of yellow catfish while maintaining a fast speed,which can provide useful reference for the study of fish status monitoring methods.