摘要
为解决太阳能电池板缺陷类型和尺度多样、小目标难以检测的难题,同时平衡各类缺陷的检测精度和速度,提出了一种改进轻量型YOLOv5的太阳能电池板缺陷检测方法。首先将网络模型部分卷积块替换为改进后的MobileOne Block模块,减少了模型参数量,提高模型检测速度;同时将主干网络的最后一层替换为SepViT Block,增强模型对全局信息的提取;然后设计了融合SimAM注意力机制的ASFF自适应特征融合模块,在改进多尺度特征提取的同时减轻模型的重量;最后增加P2检测层,提高小目标的检测效率,给模型带来持续的性能提升。实验结果表明,改进算法与原YOLOv5模型对比,参数量压缩了23.47%,检测速度达到了103 F/S,更好地实现嵌入式使用;检测精度达到了96.2%,比最新的YOLOv7-tiny提高了5.3%,证明了其优势。
In order to solve the problem that the defect types and scales of solar panels are diverse and small targets are difficult to detect,and balance the detection accuracy and speed of various defects,an improved lightweight YOLOv5 defect detection method for solar panels was proposed.Firstly,replace some convolutional blocks of network with the improved MobileOne Block module,which reduces the amount of model parameters and improves the detection speed.At the same time,the last layer of the backbone network is replaced by SepViT Block to enhance the extraction of global information from the network;Then,an ASFF adaptive feature fusion module is designed to integrate SimAM attention mechanism,which can improve multi-scale feature extraction and reduce the weight of the model;Finally,P2 detection layer is added to improve the detection efficiency of small targets and bring continuous performance improvement.The experimental results show that,compared with the original YOLOv5 model,the improved algorithm in this paper has 23.47%parameter compression,and the detection speed has reached 103 F/S,so the embedded application is better realized;The detection accuracy reaches 96.2%,which is 5.3%higher than the latest YOLOv7 tiny,proving its advantage.
作者
李婷
孙渊
LI Ting;SUN Yuan(School of Mechanical,Shanghai Dianji University,Shanghai 201306,China)
出处
《组合机床与自动化加工技术》
北大核心
2023年第11期95-99,106,共6页
Modular Machine Tool & Automatic Manufacturing Technique
基金
上海市多向模锻工程技术研究中心项目(20DZ2253200)
上海市高峰高原学科资助项目(A1-5701-18-007-03)。
