一种基于电致发光图像的光伏电池稀有缺陷定位新少样本检测器

Abstract Accurate defect detection in Photovoltaic (PV) cells is vital for ensuring the long-term stable operation of PV power generation systems . Existing deep learning based methods are widely utilized in PV cell defect detection using Electroluminescence images. However, these methods heavily depend on a substantial amount of annotated data, and thus their effectiveness is limited when dealing with rare defects that have less available data. To address this issue, a novel method called OURS_RARE_PV is proposed for rare PV cell defect localization. In particular, OURS_RARE_PV utilizes a transfer learning based few-shot object detection framework. Firstly, contrastive learning loss is employed to amplify the distinctions among different classes and strengthen the similarities within a single class, leading to more discriminative feature representations that make samples from various classes more easily distinguishable. Next, we design the knowledge compensation module to address base defect class knowledge forgetting, caused by parameter updates during fine-tuning. It decouples training and testing by introducing an additional branch with frozen parameters to preserve base defect class knowledge during training. Furthermore, we establish a new dataset named few-shot PVEL_AD. Finally, OURS_RARE_PV was evaluated on the few-shot PVEL_AD dataset, using different sample settings across two different scenarios. Compared to other methods, OURS_RARE_PV achieved the best performance, demonstrating its superiority. Code is available at: https://github.com/Researcher-TJER/OURS_RARE_PV .