Few-Shot Class Incremental Object Detection for Packaging Defects in Domain Transfer Scenerios

Few-shot class incremental learning presents a promising approach for the adaptation of detect models to evolving scenarios where the diversity of cigarette defect classes dynamically expands. While in practical production scenarios, the diversity of cigarette boxes may also alter due to production plan adjustments, with new brands potentially replacing old ones. This results in a domain shift within the training data, posing challenges for defect detection models. To circumvent the need for redundant training and reduce costs, we can transfer the detection model from old to new brand. We frame this challenge as a Domain-Transfer Few-Shot Class Incremental Learning (DT-FSCIL) problem, where the key difficulties stem from overfitting and catastrophic forgetting when learning from only a few samples of new defect categories within the newly introduced domain. To tackle these challenges, we devise a transfer strategy based on knowledge distillation to harness expertise from the source domain and propose a parameter update method guided by class sensitivity, which decouple the learning of classification and detection to avoid performance degradation of old class. We have verified the superiority of the proposed method through experiments and achieved good performance on the cigarette defect dataset.