DEVELOPMENT OF ROBUST DEEP LEARNING FRAME WORKS FOR PROHIBITED ITEM DETECTION IN X-RAY SECURITY IMAGERY

Abstract

The detection of prohibited items in X-ray baggage imagery plays a vital role in ensur ing public safety at transportation hubs, airports, railway stations, metro stations, and other high-security environments. X-ray image analysis provides critical visual infor mation about concealed and overlapping objects inside baggage, enabling security per sonnel to identify potential threats more effectively. However, manual inspection of X ray scans is time-consuming, labor-intensive, and prone to human error, especially in challenging scenarios involving cluttered backgrounds, severe occlusion, object over lap, and complex object orientations. Therefore, AI-based prohibited object detection systems have gained significant attention in recent years because of their potential to improve screening efficiency, detection accuracy, and operational reliability. This thesis investigates key challenges associated with prohibited item detection in X-ray baggage imagery. Although deep learning-based object detection methods have achieved substantial progress over the past decade, challenges related to occlusion han dling, feature degradation, multi-scale object representation, and robust detection in cluttered X-ray imagery remain insufficiently addressed. To overcome these limita tions, this thesis proposes deep learning frameworks that effectively capture discrimi native spatial and frequency-domain features from complex X-ray scans and accurately detect prohibited items under challenging conditions. Specifically, two complementary detection frameworks are proposed. First, an Edge-Aware Deformable Feature Pyramid Network (EDFPN) is introduced to improve feature representation by enhancing object boundary information and spatial adapt ability. The framework incorporates edge-aware feature enhancement operators and deformable convolution mechanisms to better capture irregular object structures and improve localization performance in cluttered environments. An empirical analysis is also conducted to evaluate the effectiveness of deformable feature learning and edge guided representations for prohibited item detection. Second, a Frequency-Spatial Guided Deformable Fusion Network is proposed for robust prohibited object detection in highly occluded and overlapping X-ray imagery. The proposed framework integrates frequency-domain feature extraction using Ga vi bor filtering with spatially adaptive deformable learning to capture both textural and structural information effectively. Furthermore, channel modulation and multi-path feature interaction mechanisms are incorporated to strengthen feature fusion and im prove discriminative learning across multiple scales. These components collectively enhance the network’s capability to detect concealed prohibited objects under chal lenging imaging conditions. Both proposed approaches are designed, implemented, and extensively evaluated on benchmark X-ray security datasets, including PIDray, OPIXray, and HiXray. Ex perimental results demonstrate significant improvements in detection accuracy, robust ness against occlusion and clutter, and real-time inference performance compared with existing state-of-the-art methods. The proposed systems provide efficient, reliable, and intelligent solutions for automated security screening applications, contributing toward safer and more effective public security infrastructure.