Regressor driven incremental instance segmentation for contraband items identification

Detecting suspicious objects contained within passenger baggage is one of the most difficult tasks, even for the security experts. To address this problem, many researchers have developed computer-aided screening methods employing deep learning models to detect the presence of contraband items using X-ray imagery. However, all of these models have limitations or bottlenecks when it comes to detecting prohibited objects that are heavily obscured, cluttered, overlapping, and well concealed within the baggage. To overcome these challenges, we present a novel instance segmentation framework that transforms the conventional semantic segmentation models to perform instance-aware segmentation, via incremental learning, to automatically detect suspicious baggage items. This framework leverages knowledge distillation, enabling the model to iteratively learn and retain multiple instances of threat items. By continuously adapting, the model improves its ability to distinguish between different instances, showcasing significant advancements in security and threat detection. To overcome the under-segmentation and over-segmentation problem of the incremental instance segmentation, we introduced the use of a lightweight regressor model that can accurately identify the overlapping instances of the suspicious objects which enables the optimal selection of the segmentation model to perform the required task. Moreover, the proposed framework has been rigorously tested on two publicly available datasets on which it achieved the mask mean average precision scores of 0.54 and 0.51, respectively, at the inference stage. Similarly, the proposed method outperformed the state-of-the-art methods by 5.88%, and 8.51% in terms of mask mean average precision scores across both datasets, respectively. In addition to this, the proposed framework provides an optimal trade-off between performance and efficiency as compared to its competitors.