1. Introduction
Security X-ray screening is critical for identifying prohibited items in baggage and cargo, but current systems often struggle with occlusions and diverse object appearances. This work addresses the need for more accurate and efficient automated detection to reduce human workload and enhance security. We propose ALDI-ray, an adapted framework building upon the Attribute-Learned Detector and Inferrer (ALDI) model to tackle these challenges specifically within the X-ray domain. The models used in this article include the adapted ALDI framework, incorporating modifications to its backbone and detection head, and potentially comparative baseline models such as Faster R-CNN and YOLO.
2. Related Work
Existing literature on object detection in security X-rays frequently employs deep learning architectures like CNNs, with many focusing on specific contraband types. Traditional ALDI frameworks have shown promise in general object detection tasks due to their attribute-based learning approach. However, applying these general frameworks directly to X-ray data presents unique challenges due to material penetration and pseudo-color variations. This section reviews prior work in X-ray image analysis, emphasizing the gap that ALDI-ray aims to fill by combining efficient detection with specialized X-ray feature extraction.
3. Methodology
The methodology involves several key steps, starting with the preprocessing of a large-scale security X-ray dataset, including data augmentation tailored for X-ray specific characteristics. The core ALDI framework was then adapted, involving architectural modifications to the feature extraction network to better process multi-energy X-ray channels and to improve robustness against varying object densities. Training involved a transfer learning approach, initializing weights from a pre-trained ALDI model on a general dataset and fine-tuning extensively on the X-ray dataset. This process aimed to optimize the model for high recall and precision in identifying small and occluded objects.
4. Experimental Results
Experimental evaluation of ALDI-ray on a diverse X-ray security dataset demonstrates significant performance improvements across key metrics. The system achieved a higher mean Average Precision (mAP) and F1-score compared to several state-of-the-art baselines, indicating its superior ability to accurately detect various prohibited items. Furthermore, the inference speed of ALDI-ray was competitive, making it suitable for real-time security screening applications. The table below summarizes the comparative performance metrics of ALDI-ray against baseline models on the test set.
5. Discussion
The results clearly indicate that adapting the ALDI framework for security X-ray object detection yields substantial benefits, particularly in improving detection accuracy and efficiency. The attribute-learning capability of ALDI, when tailored to X-ray features, proved effective in discerning challenging objects often missed by generic detectors. These findings suggest that ALDI-ray can significantly enhance the effectiveness of automated security screening systems, reducing false alarms and improving the detection rate of critical threats. Future work will explore further optimization of the ALDI-ray architecture and its deployment in real-world operational environments.