A unified deep learning application framework for forensic shoeprint analysis under data-limited conditions

This study presents a data-efficient framework for applying deep learning to forensic shoeprint matching under realistic casework constraints. In forensic practice, deployment is often limited by scarce labeled data, distributional differences between reference collections and casework evidence, and a lack of practical guidance on model selection, augmentation, and adaptation. Rather than proposing a new architecture, this study focuses on how existing deep learning components can be systematically adapted to improve reliability under such conditions. We examine three key aspects commonly encountered in practice: the selection of pretrained representations, the design of augmentation strategies, and adaptation to new domains using limited target-domain data. Experiments on two complementary shoeprint datasets demonstrate that generic pretrained models without adaptation exhibit unstable behavior, whereas domain-aligned backbones provide more consistent performance. In terms of data augmentation, standard geometric transformations can lead to non-monotonic and sometimes degraded performance, while morphology-preserving generative augmentation based on a convolutional autoencoder yielded more stable improvements. Under cross-domain dataset shift, few-shot fine-tuning with only 1-3% of labeled target-domain data recovers a substantial portion of the performance achieved with full-data training. In addition, combining generative and standard augmentation strategies produces competitive performance, suggesting that different augmentation mechanisms may introduce complementary variability, although systematic exploration of such combinations remains an open problem. Overall, this study provides practical, evidence-based guidance for applying deep learning in forensic shoeprint analysis under data-limited and domain-shifted conditions, with implications for other domains where similar constraints arise.