Anomaly Detection for Photovoltaic Modules Based on Image Inpainting

The method based on deep learning shows excellent performance in the field of photovoltaic modules defect detection. However, the defect samples of photovoltaic modules in industrial production are sparse and the characteristics are very different, which makes the detection method requiring a large number of defect samples training difficult to apply. To solve this problem, a semi-supervised anomaly detection model based on image inpainting (SIAD) is proposed for the detection of photovoltaic modules defects. SIAD network structure follows the design of DRAEM model. In the training stage, a large number of normal samples are used. At the same time, based on the defect information of a small amount of abnormal samples, the hybrid mask strategy is used to simulate the defective area to participate in training. In this way, the model can achieve better performance with little annotation workload. Meanwhile, context attention module (CAM) is used to extract long distance dependence in the bottleneck of autoencoder to improve the quality of image inpainting, and improve the ability of defect location. Compared with other anomaly detection methods, this method achieves the best detection performance on the EL image dataset of photovoltaic modules.