SD-IDD: Selective Distillation for Incremental Defect Detection

Surface defects in industrial production are complex and diverse. Therefore, deep learning-based defect detection models must consistently adapt to newly emerging defect categories. The trained models generally suffer from catastrophic forgetting as they learn new defect categories. To address this issue, we propose a selective distillation for incremental defect detection (SD-IDD) model based on GFLv1. Specifically, three selective distillation strategies are proposed, including high-confidence classification distillation, dual-stage cascaded regression distillation, and Intersection over Union (IoU)-driven difficulty-aware feature distillation. The high-confidence classification distillation aims to preserve critical discriminative knowledge of old categories within semantic confusion regions of the classification head, reducing interference from low-value regions. Dual-stage cascaded regression distillation focuses on high-quality anchors through geometric prior coarse filtering and statistical fine filtering, utilizing IoU-weighted KL divergence distillation loss to accurately transfer localization knowledge. IoU-driven difficulty-aware feature distillation adaptively allocates distillation resources, prioritizing features of high-difficulty targets. These selective distillation strategies significantly mitigate catastrophic forgetting while enhancing the detection accuracy of new classes, without requiring access to old training samples. Experimental results demonstrate that SD-IDD achieves superior performance, with mAP_old of 58.2% and 99.3%, mAP_new of 69.0% and 97.3%, and mAP_all of 63.6% and 98.3% on the NEU-DET and DeepPCB datasets, respectively, surpassing existing incremental detection methods.