Defect detection with generative adversarial networks for electroluminescence images of solar cells

Solar cells are the core module of photovoltaic (PV) modules. Defects will decrease the power efficiency of solar cells and reduce the stability of PV power systems. Electroluminescence (EL) imaging is able to image solar modules with higher resolution so that defects can be better detected. The current manual detection of EL images is slow and requires relevant expertise, so methods based on computer vision for automatic detection in EL images are appearing. However, due to the heterogeneously background of the EL images and the lack of defect samples, automatic detection of defects has been a challenging task. We design a model based on generative adversarial networks (GAN) and auto-encoder (AE) to perform defect detection for EL images of solar cells. It only requires normal images in the training process and detect defects by measuring the residuals between the test image and the constructed image generated by the generator. To reduce the effects of image distortion, we combine structural similarity index (SSIM) with feature residuals to train the encoder, which can get better results than the model using typical mean square error (MSE). During the detection phase, SSIM and MSE are combined as the anomaly score. Our method has higher recognition of defective EL images and achieves detection accuracy of 90.0% on the test set. Compared with the method using only MSE, the F1 score is increased obviously.