基于改进YOLOv5的光伏组件缺陷检测

Improved YOLOv5-Based Defect Detection in Photovoltaic Modules

电致发光(EL)检测技术作为太阳电池和组件缺陷检测的重要手段被广泛运用,但是EL检查中的缺陷筛查仍然需要持续完善。为了克服以往研究中可识别缺陷的种类少、无法对缺陷进行定位、模型参数多体积大及检测速度慢的局限性,使用改进的YOLOv5网络对电致发光图片中常见的隐裂、断栅、裂片和黑斑4类主要缺陷进行检测和分类。使用Ghost模块代替YOLOv5骨干提取网络中的普通卷积模块,减少网络模型的参数量;为了保证良好的检测性能,在骨干网络尾端加入Squeeze-and-Excitation(SE)注意力模块,提升算法的目标检测能力;在特征融合网络中引入双向特征金字塔网络(BiFPN)结构,进一步加强网络的特征融合能力。结果表明,所提模型成功地识别和定位了4类常见的缺陷,与YOLOv5算法相比,模型体积减小了21%,在没有GPU加速的情况下,单张图片的检测速度提升了17.4%。

Electroluminescence(EL) inspection technology is widely used as an important means for solar cell and module defect detection.However,defect screening in EL inspection is still a major challenge.Herein,to overcome the limitations in previous studies,such as few types of defects to be identified,the inability to locate defects,the large size of model parameters,and slow detection speed,the upgraded YOLOv5 network is used to detect and classify the four types of defects that are commonly found in electroluminescent images,including crack,finger interruption,break,and black zone.This shows that the improved Ghost module extracts ordinary convolutional modules in the network to reduce number of network model parameters compared with the YOLOv5 backbone.Additionally,to ensure good detection performance,the Squeeze-and-Excitation(SE) attention module is added to the tail of the backbone network to improve the algorithm' target detection ability.In the neck part,the bidirectional feature pyramid network(BiFPN) structure is used to further strengthens the feature fusion capability of the network.Experimental results show that the proposed model successfully identifies and locates the four common defects,has a reduced volume by 21% compared with the YOLOv5algorithm and achieves an improved single image detection speed by 17.4% without GPU acceleration.