Class-Incremental Photovoltaic Cells Defect Detection via Local Foreground Distillation

The production process of photovoltaic (PV) cells can easily lead to various defects. Defect detection is a necessary method to ensure the quality of PV components. Computer vision-based detectors are widely used in the quality inspection process, which is an important means to ensure the quality of PV cells production. During the quality inspection process, the number of defects that need to be detected may increase gradually. However, traditional object detectors cannot adapt to streaming data. Fine-tuning the detection model directly with new data will result in catastrophic forgetting, and the dataset needs to be rebuilt and retrained whenever a new class needs to be detected. We build an Incremental Object Detection (IOD) method based on Knowledge Distillation (KD) called Local Foreground Distillation (LFD) is proposed for the feature map of the detector. We distill the local region of old classes in feature map by relying on the ability of the trained teacher model to identify old classes, so as to avoid the influence of background noise on the distillation process, and obtain a better stability-plasticity balance. A large number of experiments on PVEL-AD datasets show that our method achieves the most advanced results.