Enhancing forensic shoeprint analysis: Application of the Shoe-MS algorithm to challenging evidence

• Published in: Science & justice Vol. 65; no. 4; p. 101255
• Main Authors: Jang, Moonsoo, Carriquiry, Alicia, Park, Soyoung
• Format: Journal Article
• Language: English
• Published: England Elsevier B.V 01.07.2025
• Subjects: Algorithms; Databases, Factual; Deep Learning; Forensic Sciences - methods; Humans; Image Processing, Computer-Assisted; Pathology; Pattern evidence; Pattern Recognition, Automated; Shoe-MS algorithm; Shoeprint; Shoes; Deep learning; Shoe-MS algorithm; Pattern evidence; Shoeprint
• ISSN: 1355-0306; 1876-4452; 1876-4452
• DOI: 10.1016/j.scijus.2025.101255

Quantitative assessment of pattern evidence is a challenging task, particularly in the context of forensic investigations where the accurate identification of sources and classification of items in evidence are critical. Emerging deep learning approaches can become useful tools for examiners responsible for pattern recognition and analysis. This paper explores the Shoe-MS algorithm, a deep learning-based framework specifically designed for forensic footwear analysis where the input consists of two paired images, and the output is an estimated similarity score that takes on a value between zero and one. We implement Shoe-MS on two different databases that permit assessing the algorithm’s performance for source identification and for the classification of degraded images. Our experimental results demonstrate that the Shoe-MS algorithm achieves high performance across both tasks, highlighting its potential for forensic footwear analysis. No algorithm can substitute examiners, but Shoe-MS produces reliable similarity scores and can help examiners make probabilistic, reproducible, and repeatable assessments. Initial findings suggest that Shoe-MS can be a valuable tool for examiners evaluating pattern evidence, especially when crime scene images are not of the highest quality. •Shoe-MS is a deep-learning-based shoeprint analysis algorithm using a Siamese net.•Evaluated on comparably clean (DB1) and degraded (DB2) footwear impression.•Shoe-MS outperformed other methods, especially on degraded and noisy condition.•Fine-tuning improved performance on task-specific and challenging conditions.•Future work targets generalizability, partial prints, and broader shoe datasets.