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

Area of Science:

• Forensic Science
• Computer Science
• Machine Learning

Background:

• Deep learning (DL) application in forensic science is hindered by data scarcity and domain shift.
• Practical guidance for selecting, augmenting, and adapting DL models in casework is lacking.

Purpose of the Study:

• To develop a data-efficient framework for applying DL to forensic shoeprint matching.
• To systematically adapt existing DL components for reliability under casework constraints.
• To provide evidence-based guidance for DL in data-limited forensic scenarios.

Main Methods:

• Examined pretrained representation selection, augmentation strategies, and few-shot domain adaptation.
• Utilized two shoeprint datasets for experimentation.
• Employed morphology-preserving generative augmentation via convolutional autoencoders.

Main Results:

• Domain-aligned backbones showed more consistent performance than generic pretrained models.
• Generative augmentation yielded more stable improvements than standard geometric transformations.
• Few-shot fine-tuning (1-3% data) recovered substantial performance under domain shift.

Conclusions:

• Systematic adaptation of DL components is crucial for forensic shoeprint analysis.
• Morphology-preserving generative augmentation and domain alignment enhance DL reliability.
• The framework offers practical guidance for DL deployment in data-limited forensic casework.