Hierarchical fusion of multi-scale features with transformer for crime scene trace identification

To address the technical challenge of low identification accuracy for crime scene traces in complex environments, this paper proposes a multi-branch feature fusion model based on transformers.Traditional methods rely on manually selected features, while existing deep learning models struggle to balance trace-level details with global structural patterns.To clarify this point, these global structural patterns indeed often encompass critical spatial distribution characteristics of trace features.The relative positions, densities, and geometric relationships of features are fundamental to achieving accurate and robust trace matching.This study constructs four specialised branches -fingerprints, tool marks, footprints, and biological traces -to extract distinctive features from each trace type.It further designs cross-branch interaction and adaptive fusion mechanisms to achieve effective integration of multi-source trace information.Experiments on public datasets demonstrate that our method achieves an average identification accuracy of 95.7%, surpassing mainstream models by over 6%.This significantly enhances trace recognition capabilities in complex scenarios involving blurred or fragmented evidence.