H-CMAF: A Deep Learning-Based Cross-Modal Attention Fusion Framework for Dangerous Goods Detection from X-Ray and Visible Light Imagery

Abstract

Hazardous goods identification still faces challenges such as low accuracy and difficulty identifying items in low-light, densely packed environments. Therefore, this paper combines multimodal fusion information with deep learning models to construct an identification and classification system. This aims to improve the accuracy of hazardous goods identification in the public domain and its robustness in diverse environments and conditions. Based on the attention mechanism, this study fuses feature information from X-rays and visible light to construct an H-CMAF framework model. In challenging scenarios such as low light, partial occlusion, and dense stacking, H-CMAF achieves minimal mAP degradation (6.5%, 5.9%, and 7.1%, respectively), demonstrating exceptional robustness. Furthermore, the model achieves an inference speed of 22.8 FPS on edge devices, with only 34.8M parameters and a memory footprint of 100MB, outperforming most competing models and achieving a good balance between accuracy and efficiency.