Adaptive dual teacher incremental learning for defect detection of flexible printed circuit

Abstract

Flexible Printed Circuit (FPC) is a high space utilization and high flexibility circuit board. Due to the flexible production method of FPC, new class defects will constantly appear during the production process. Existing methods for incremental class learning can enable the model to learn new class defects, however less attention is paid to the balance between old and new class defects. In addition, the knowledge of new class defects extracted by old class teacher is inaccurate because the teacher model has not learned new class defects. To address these issues, this paper proposes a model called Adaptive Dual Teacher Incremental Learning with Decoupled Feature Distillation for defect detection (ADT-ILDFD), which employs an adaptive dual-teacher model to adjusts the importance of old and new classes through adaptive modular learning during the training process, and thus better accommodate both old and new classes simultaneously in different incremental environments. In addition, decoupled feature distillation was designed, which decouples the feature knowledge in order to distill the feature knowledge extracted by old class model more accurately. Finally, we constructed the FPC defect dataset named FPCSD2023 and conducted experiments on both FPCSD2023 and NEU-DET to demonstrate the effectiveness of ADT-ILDFD.