Multi-scale feature decoupling and similarity distillation for class-incremental defect detection of photovoltaic cells

Existing vision-based photovoltaic cell defect detection methods usually update models with all defect data of both old and new categories to adapt to new classes emerging in the dynamic data stream from realistic production lines. This model updating strategy wastes resources and sometimes is infeasible due to the confidentiality of historical data. In this paper, we propose a novel distillation-based model updating method, i.e., multiscale Feature Decoupling and Similarity Distillation (mFDSD), which updates the model with only new data from the dynamic data stream but can effectively identify both old and new categories. In mFDSD, we design a maskbased feature decoupling distillation module and a similarity-regulated feature distillation module to adaptively regulate distillation losses assigned to important defective areas, less-important background areas, and discarded areas of feature maps. Experimental results demonstrate that our mFDSD outperforms the current state-of-the-art distillation-based model updating methods for the class-incremental defect detection of photovoltaic cells.