Investigation of generated near-infrared crack defect dataset for solar cells

The generation of near infrared electroluminescence (EL) crack defect images of polycrystalline silicon (poly-Si) solar cells is investigated. For complex environments (such as background object interference) and weak target signals (such as low pixel ratio), object detection is generally difficult, as is the case for the direct generation of such images. Atypical example is the crack defect detection on poly-Si solar cells. The irregular complex environment may cover the defects and cause the generation model to learn irrelevant background and noise information, thus affecting the generation of defect images. Based on StyleGAN model and image processing methods, a two-stage “generation + fusion” simulation method is proposed to generate the EL crack defect images for poly-Si solar cells. The similarity between simulated and real images and the effectiveness of the simulated dataset for small sample learning are studied. The method can also be used for other similar industrial defect detection and data enhancement research.