基于机器视觉的太阳能电池片表面缺陷检测  

Surface Defect Detection Method of Solar Cell Based on Vision

作　　者：张鹏飞 王淑青[1] 黄剑锋 刘逸凡 张子言 ZHANG Peng-fei;WANG Shu-qing;HUANG Jian-feng;LIU Yi-fan;ZHANG Zi-yan(School of Electrical and Electronic Engineering,Hubei University of Technology,Wuhan 430068,China;Wuhan National Research Center for Optoelectronics,Huazhong University of Science and Technology,Wuhan 430074,China)

机构地区：[1]湖北工业大学电气与电子工程学院,武汉430068 [2]华中科技大学武汉光电国家研究中心,武汉430074

出　　处：《组合机床与自动化加工技术》2022年第8期144-147,151,共5页Modular Machine Tool & Automatic Manufacturing Technique

基　　金：国家自然科学基金项目(61873195)。

摘　　要：为提高太阳能电池片表面各类缺陷的检测精度及速度,设计了一种轻量化YOLO深度学习算法的检测模型。首先在YOLOv4模型的基础上,采用反向线性卷积与深度可分离卷积重新设计主干网络;其次,将Neck部分的路径聚合网络及特征金字塔替换为信道增强特征金字塔,引入亚像素连接,并结合亚像素上下文信息完成特征集成映射,提升高层次信道特征的使用率;最后使用通道注意力引导模块,增强缺陷定位的稳定性。试验证明,模型的检测准确率达97.5%,平均检测速度可达23 ms,检测精度高、规模小、耗时低。To improve the detection accuracy and speed of various defects on solar cell surface,a detection model based on lightweight YOLO deep learning algorithm was designed.Based on YOLOv4 model,inverse linear convolution and depth separable convolution are used to redesign the backbone network.Then,the path aggregation network and feature pyramid of Neck part are replaced by channel enhanced feature pyramid.Subpixel connection is introduced to integrate feature mapping with context information to improve the utilization rate of high-level channel features.Finally,the stability of defect location was enhanced by channel attention guidance module.Experiments show that the detection accuracy of the model is up to 97.5%,the average detection speed is 23 ms,which has high detection accuracy,small scale and low consumption time.

关 键 词：太阳能电池片 缺陷检测 轻量化YOLO 反向线性卷积 深度可分离卷积 特征金字塔 

分 类 号：TH165[机械工程—机械制造及自动化] TG66[金属学及工艺—金属切削加工及机床]