DEFSR-Net: A joint learning network of super-resolution enhanced dual-branch edge features for X-ray contraband security detection in energy dispersive spectrometer environment

Abstract

The expansion of the transportation industry has not only facilitated the flow of people and goods, but also brought security and regulatory challenges. To this end, we propose DEFSR-Net, a novel X-ray contraband detection network with a dual backbone and detection transformer structure. The network integrates image processing and object detection through multi-task joint learning. To enhance the visibility of key features in X-ray images, we adopt a dual-branch structure for collaborative learning. First, we use the Real-ESRGAN super-resolution enhancement dataset and apply a decolorization algorithm to highlight color information. Then, an improved edge enhancement module is used to emphasize edge features and a branched backbone is combined to capture various feature types. Then, we introduce an edge-guided feature fusion module to merge features from different stages of the dual backbone, thereby effectively enhancing multi-scale feature representation and edge receptive field. To address the class imbalance problem, we use Unified-IoU for weight distribution and an annealing strategy to balance training. Extensive experiments on the EDS and CLCXray dataset confirm that DEFSR-Net is suitable for real-time deployment and has high accuracy.