Photovoltaic cell-like incremental defect detection method based on knowledge distillation

Abstract

The application is a photovoltaic cell incremental defect detection method based on knowledge distillation. First, a photovoltaic cell base category defect dataset is established, an original defect detection model is constructed, and the model is trained. Then, a photovoltaic cell new category defect dataset is established. The trained original defect detection model is used as a teacher model, the student model has the same architecture as the teacher model, and the parameters of the teacher model are used to initialize the student model. Finally, the photovoltaic cell new category defect dataset is input into the teacher model and the initialized student model, the student model is incrementally trained based on knowledge distillation, the trained student model is used as the final defect detection model for photovoltaic cell defect detection. When the defect category to be detected increases, the photovoltaic cell new category defect dataset is repeatedly established and the student model is retrained. When the defect category to be detected increases, the student model is incrementally trained, so that the model has the ability of continuous learning, and the time cost is reduced without sacrificing the model detection performance.

Images