关键词: 图神经网络, 足迹识别, 相似度度量, 法医学应用

Abstract: The retrieval and identification of barefoot and sock-wearing footprints in mixed sample databases has consistently been a challenging problem in real-world cases. This paper proposes a footprint recognition algorithm based on graph neural networks for barefoot and sock-wearing feet. First, a conversion framework from footprint images to graph structures is designed. Through grid division, footprints are segmented into multiple local regions with statistical features extracted as node attributes, and connections established based on spatial adjacency relationships, effectively preserving the visual features and spatial topological structure of footprints. Second, a graph neural network model specifically designed for footprint similarity calculation is constructed. This model integrates the advantages of graph attention convolution and graph convolutional networks, realizes the conversion from node-level to graph-level feature representation through attention pooling mechanism, and enhances the accuracy of similarity calculation through tensor network module that integrates multiple similarity measurement strategies. Finally, data processing and optimization strategies adapted to the characteristics of footprint recognition are proposed, including weighted loss function design and adaptive learning rate scheduling mechanism. Experimental results demonstrate that the proposed method achieves 79.53% accuracy and an F1 score of 82.27% on the test set, significantly outperforming traditional deep learning methods such as ResNet-50 and Siamese networks. Ablation experiments further validate the contribution of each component to performance, confirming the rationality and effectiveness of the proposed architecture. This research not only advances footprint recognition technology but also provides new ideas and methods for related fields.

Key words: Graph Neural Network, Footprint Recognition, Similarity Measurement, Forensic Application