Solar Cell Anomaly Detection by ResNet Work with SE Attention Module

Aiming at the limitations of low detection accuracy and low detection efficiency of current anomaly detection methods for solar cells, especially the insufficient detection of subtle defects such as cracks and finger failures, an anomaly detection model SE-PatchCore is proposed for anomaly detection of solar cells in electroluminescence (EL) images by adding attention module in ResNet. Firstly, the SE (Squeeze and Excitation) attention mechanism is introduced into ResNet, and the network is applied to the advanced local anomaly detection model PatchCore. Finally, the Fair k-Centers method is used for coreset subsampling. Through testing on ELPV dataset, the score of SE-PatchCore in image-level AUROC is up to 98.7%, and the score of pixel-level AUROC is up to 98.5%, which has an improvement of 0.2% and 0.6% compared with PatchCore. And it has higher improvement over other methods such as SPADE and PaDiM. A higher AUROC score means the model can more accurately distinguish defective solar cells from normal ones and reduce false positives and false negatives. This helps to cut down on unnecessary rework, lower production costs, and ensure that more high-quality solar cells reach the market, meeting the increasing demand for reliable renewable energy products. The above results show that the introduction of SE attention module and the Fair k-Centers method can effectively improve the accuracy of anomaly detection of solar cells in EL images, which provides strong technical support for solving problems in this field.