| Title |
Token-based Interpretable Framework for Shoeprint-outsole Matching |
| Authors |
김준규(Jun Kyu Kim) ; 송형섭(Hyeong Seop Song) ; 권현섭(Hyun Seob Kwon) ; 장경훈(Kyung Hoon Jang) ; 김동주(Dong Ju Kim) ; 서영주(Young Joo Suh) ; 이창엽(Chang Yeop Lee) |
| DOI |
https://doi.org/10.5573/ieie.2026.63.2.112 |
| Keywords |
Image retrieval; Deep learning; Computer vision; Cross domain; Transformer |
| Abstract |
This study proposes a Token-based Cross-domain matching framework to simultaneously improve the accuracy and interpretability of shoeprint-outsole matching in crime scene investigations. Conventional shoeprint analysis relies on experts’ visual comparison or manual work, which are prone to subjective errors and have limitations in effectively handling the cross-domain gap between low-quality shoeprint images and high-quality outsole images. To address these issues, the proposed framework consists of three core modules. First, image grouping using embedding-based iterative label refinement is performed to alleviate label inconsistency caused by manual labeling and to ensure data consistency. Second, Multi-scale Re-ranking is applied to Re-evaluate candidate pairs at various resolutions, thereby improving retrieval accuracy by addressing for differences in pattern scale. Third, attention map-based similarity visualization is employed to visually compare the common patterns and attended regions between shoeprints and outsoles, thereby enhancing the reliability and interpretability of cross-domain matching results. Experimental results show that the proposed method achieves superior retrieval accuracy compared to existing methods, and demonstrates through visual evidence that it constitutes a reliable and interpretable automated shoeprint-outsole matching framework. |