Object detection in X-ray security inspection images is one of the most representative applications of computer vision in the field of public safety. Unlike natural images, X-ray imaging is characterized by penetration and superposition, color encoding (organic materials in orange, inorganic materials in blue, metals in deep black), and highly variable object morphology. These characteristics dictate that detection algorithms for security inspection must follow a technical route that both intersects with and diverges from generic object detection.
I. The Era of Traditional Methods: Feature Engineering and Classifiers
Early approaches relied on handcrafted features: descriptors such as HOG, LBP, and SIFT extracted texture and edge information from X-ray images, which were then fed into classifiers like SVM and AdaBoost for prohibited item discrimination. The limitations of these methods lie in the restricted expressive power of the features, insufficient robustness against complex conditions such as occlusion, stacking, and rotation, and the need for per-category tuning that drives up engineering costs.
II. The Deep Learning Era: From Two-Stage to One-Stage Detection
Two-stage detectors represented by Faster R-CNN decouple region proposal generation from classification and regression, achieving superior accuracy but at the cost of slower inference. The YOLO family reframes detection as a single-shot regression problem, unifying real-time inference with end-to-end training, and has become the mainstream choice for deployment on edge devices in security inspection. In recent years, Transformer-based DETR and its variants model object detection as a set prediction problem, eliminating the post-processing dependence on anchor boxes and NMS, and continuously setting new records in accuracy. However, their computational demands and convergence characteristics still require further validation in security inspection scenarios.
III. Unique Constraints of Security Inspection Scenarios
Security inspection scenarios impose three types of unique constraints on algorithms. First, class imbalance is extremely severe—normal items outnumber prohibited items by orders of magnitude, making models prone to bias toward the majority class. Second, the cost of false alarms is asymmetric—missed detections (false negatives) directly threaten public safety, while false positives lead to wasteful manual bag-check procedures, requiring the identification of an optimal operating point under asymmetric costs. Third, multi-device imaging domain shift—X-ray machines of different brands and models produce varying imaging styles, demanding that models possess cross-domain generalization capability.
IV. Engineering Responses in Practice
To address the above constraints, Daoxin Dechen has adopted a combined strategy in practice: training a base detection model on large-scale annotated data (including transfer from public datasets); enhancing robustness through data augmentation tailored to X-ray imaging characteristics (perspective transformation, blended superposition, brightness perturbation); mitigating imbalance via hard example mining and class reweighting; and finally, in the deployment phase, implementing a human-in-the-loop mechanism of AI preliminary screening plus manual review, which explicitly delegates model uncertainty to the review stage for resolution. This technology stack has achieved an engineering-acceptable balance between detection rate and false alarm rate in long-term railway operations.