Enhancing automated shoeprint comparison via synthetic data generation and deep segmentation

Forensic footwear impressions are routinely used to support shoeprint comparison and investigative decision-making, yet automated analysis remains challenging due to background clutter, partial occlusion, print degradation, and the limited availability of paired data with reliable source labels. We present a framework that couples controllable synthetic data generation with deep segmentation to evaluate and improve algorithmic shoeprint comparison under a range of challenging conditions. We first generate a synthetic dataset by systematically varying background textures, noise, occlusion, and degradation while preserving ground-truth correspondence between impressions. We then train a deep segmentation model to separate shoeprint regions from cluttered backgrounds, improving the quality of extracted impression patterns for downstream similarity scoring. Within the controlled experimental setting, segmentation-based preprocessing was associated with improved verification performance relative to the non-segmented baseline and other evaluated baselines. Overall, this framework offers a reproducible evaluation pipeline and a segmentation-based preprocessing approach for controlled evaluation of algorithmic shoeprint comparison methods.