基于改进YOLOv5s小目标检测算法

Improved YOLOv5s small object detection algorithm

目的:针对现有目标检测算法进行小目标检测时检测效果不理想、漏检率高的问题,提出一种改进的YOLOv5s检测算法,提升小目标检测效果。方法:在原有模型基础上,引入BottleneckCSP模块并增加大尺度特征融合结构,提升模型小目标特征捕捉能力;同时在网络结构中融合SE注意力机制,使得网络自主学习更关注小目标特征通道,增强网络模型对小目标的检测效果。结果:在同一自制小目标检测数据集上进行训练验证,与已有算法比较,能够有效提升YOLOv5s目标检测算法的mAP值和训练收敛速度,拓展小目标检测范围(由原有算法的0.002 5~0.010 0缩小至0.000 8~0.001 4),提高小目标检测性能(平均检测率提升46%)。结论:改进算法能够有效提升小目标的检测能力。

Objective:To solve the problem of unsatisfactory detection effects and high missed detection rate when the existing object detection algorithm performing small object detection,an improved YOLOv5s detection algorithm was proposed to improve the effect of small object detection.Methods:On the basis of the original model,the BottleneckCSP module was introduced and the large-scale feature fusion structure was added to improve the model's ability to capture small object features,and the SE attention mechanism was integrated into the network structure,so that the network self-learning paid more attention to the small object feature channel and enhanced the detection effect of the network model on small objects.Results:Comparing with the existing algorithms,the training verification on the same self-made small object detection dataset could effectively improve the mAP value and training convergence speed of the YOLOv5s object detection framework,expanded the small object detection range(from 0.0025—0.0100 to 0.0008—0.0014 of the original algorithm),and improved the small object detection performance(the average detection rate was increased by 46%).Conclusion:The improved algorithm could effectively improve the ability of small object detection.