融合多尺度特征与注意力的太阳能电池表面缺陷检测

Solar cell defect detection network combining multiscale feature and attention

为提高对太阳能电池电致发光(EL)成像各类表面缺陷的检测精度并降低漏检率,提出融合多尺度特征与注意力机制的太阳能电池表面缺陷检测算法CMFAnet。首先,针对太阳能电池表面缺陷尺度跨度大的特点,设计了增强型多尺度特征融合方法,其基本单元由特征对齐模块和特征融合模块串联组成,对于不同语义级别的特征信息,特征对齐模块通过调整它们的尺度,使这些特征更容易融合在一起;其次,针对太阳能电池表面缺陷特征与背景特征相似程度高、几何形状多变的特点,设计了可形变幽灵卷积模块,其基本单元由可形变卷积、多路坐标注意力机制和幽灵卷积(Ghost conv)组成,多路坐标注意力机制优化了可形变卷积中offset的生成,幽灵卷积机制的引入则有效降低了网络模型的计算复杂度。实验结果表明,在光伏电池缺陷异常检测数据集PVEL-AD上,本文方法的平均检测精度(mAP)达91.4%,相较其他主流目标检测网络均有不同程度的提升。

In order to improve the detection accuracy of various types of defects in the electroluminescence imaging of solar cells,a solar cell surface defect detection algorithm CMFAnet was proposed by fusing multiscale features and attention mechanism.Firstly,for the characteristics of solar cell surface defects with large scale span,an enhanced multi-scale feature fusion method was designed,whose basic unit con⁃sists of a feature alignment module and a feature fusion module connected in series,and for the feature in⁃formation with different semantic levels,the feature alignment module adjusts their scales,so that these features can be fused together more easily;secondly,for the characteristics of solar cell surface defects with high level and variable geometry,a deformable ghost convolution module is designed.Secondly,for the characteristics of high degree of similarity between defective features and background features on the so⁃lar cell surface and variable geometry,a deformable ghost convolution module was designed,whose basic unit consists of feasible variant convolution,multiplexed coordinate attention mechanism,and ghost convo⁃lution;the multiplexed coordinate attention mechanism optimizes the generation of offset in the feasible variant convolution,and the introduction of ghost convolution mechanism effectively reduces the FLOPs of the network model.The experimental results show that the mAP of this paper's method reaches 91.4%on the photovoltaic cell defect anomaly detection dataset PVEL-AD,which is improved to different de⁃grees compared to other mainstream target detection networks.