Class and Data-Incremental Learning Framework for Baggage Threat Segmentation via Knowledge Distillation

With air travel growing rapidly worldwide, airports are busier than ever, making efficient security screening a top priority. To meet this growing demand, various advanced X-ray baggage scanners have been deployed at airports worldwide. Researchers have proposed multiple automated threat detection systems to enhance security screening efficiency; however, automated baggage threat segmentation remains a complex task, especially in the context of multi-class threat detection across diverse datasets. Traditional deep learning models struggle with differentiating between multiple threat types and suffer from catastrophic forgetting when exposed to new classes. To address these limitations, we propose an incremental learning framework that enables the model to progressively learn new threat categories while retaining previously acquired knowledge. Our approach utilizes SegFormer as the backbone and introduces a custom loss function, combining mutual distillation loss, KL divergence, and cross-entropy loss, to enhance knowledge retention and adaptability. The model is trained sequentially on three publicly available datasets, SIXRAY, GDXRAY, and PIDRAY, enabling it to generalize effectively across diverse baggage imagery. Through extensive experiments, we demonstrate that our method outperforms state-of-the-art incremental learning techniques, achieving superior segmentation accuracy and knowledge retention. Furthermore, Grad-CAM visualizations and t-SNE plots provide interpretability, offering insights into the model's learning behavior and class separability. The proposed framework establishes a scalable and adaptable solution for real-world security screening applications, enabling efficient threat detection without requiring model retraining from scratch.