A Robust Semantic Segmentation Framework for Baggage Threat Detection

The identification and segmentation of concealed threats in baggage X-ray images are among the specific works contributing to heightening security within screening systems. This study prompts a resilient semantic segmentation framework which might be based on SegFormer transformers. It is empowered by its superior feature extraction engine, making identification accurate to identify threats. The framework was evaluated against three benchmark datasets, namely SIXray, GDXray, and PIDray, which all have mixed samples of guns and knives to measure their performance. They finally presented the experiment results showing the efficacy of the model, with an overall mean Intersection over Union (IoU) of 0.822 for SIXray, 0.648 for GDXray and 0.749 for PIDray. The mean class-wise accuracy was 88.1%, 83.2%, and 80.3%; overall accuracy across the datasets made it up to 98.9%, 95.7%, and 98.9%. The Grad-CAM visualizations highlight focused regions in the model, making the decision-making process interpretable. This work emphasizes the promise of transformer-based models such as SegFormer in solving challenging segmentation tasks within security applications and provides evidence for increased accuracy and interpretability.