Defective Electroluminescence Image Generation for Data Imbalance in Solar Cell Defect Inspection

Abstract

The utilization of high-resolution electroluminescence (EL) images for defect inspection in photovoltaic modules has gained significant popularity. However, there are limited works on the imbalance problem in the EL images of the photovoltaic modules, specifically that the number of defective images is substantially less than the number of normal images. To address the above problem, a fast defective EL image generation method is proposed in this article. To accurately extract the defective region, a normal image, which is the most similar to the defective image, needs to be identified. First, an image type classification network is proposed to recognize the normal images with the same type (monocrystalline or polycrystalline) as the defective image. Then, the cosine similarity is further employed to identify the normal image that is most similar to the defective image. After that, the defective template is acquired by comparing the defective image with the identified similar normal image. To quickly generate diverse and rich defective EL images, effective data augmentation methods for EL images are exploited and applied to the defective template. Specifically, small-scale rotation and Gaussian blurring are first introduced to augment the EL images. Finally, the augmented defective templates are merged with any different normal images to produce a large amount of new defective images. The experimental comparison of the proposed defective image generation method with oversampling and data augmentation, which are commonly used for data imbalance, demonstrates that our proposed method can provide richer information and thus outperform other methods with a big gap.