Detection of surface defects on solar cells by fusing Multi-channel convolution neural networks

• Published in: Infrared physics & technology 2020-08, Vol.108, p.103334, Article 103334
• Main Authors: Zhang, Xiong, Hao, Yawen, Shangguan, Hong, Zhang, Pengcheng, Wang, Anhong
• Format: Article
• Language: English
• Subjects: Deep learning; Defects detection; Faster R-CNN; R-FCN; Solar cell
• ISSN: 1350-4495; 1879-0275
• DOI: 10.1016/j.infrared.2020.103334

•This work proposes a detection algorithm of surface defects on solar cells by fusing multi-channel convolution neural networks.•This work designs a new fusion mechanism for the candidate bounding boxes from different channel, thus improving detection precision and position accuracy of defect targets.•The multi-scale and multi-aspect regions of anchor points are set to reduce the detection error rate.•The hard negative sample mining strategy is used to solve the problem of low detection precision. Manufacturing process defects or artificial operation mistakes may lead to solar cells having surface cracks, over welding, black edges, unsoldered areas, and other minor defects on their surfaces. These defects will reduce the efficiency of solar cells or even make them completely useless. In this paper, a detection algorithm of surface defects on solar cells is proposed by fusing multi-channel convolution neural networks. The detection results from two different convolution neural networks, i.e., Faster R-CNN and R-FCN, are combined to improve detection precision and position accuracy. In addition, according to the inherent characteristics of the surface defects in solar cells, two other strategies are used to further improve the detection performance. First, the anchor points of the region proposal network (RPN) are set by adding multi-scale and multi-aspect regions to overcome the problem of high false negative rate caused by the limitation of anchor points. Second, in view of the subtle and concealed defects of solar cells, the hard negative sample mining strategy is used to solve the problem of low detection precision caused by the negative sample space being too large. The experimental results showed that the proposed method effectively reduced the false negative rate and the false positive rate of a single network, and it greatly improved the accuracy of the locations of defects while improving the object recall rate.