Reducing manual labour in forensic microtrace recognition with deep learning.

Area of Science:

• Forensic Science
• Computer Vision
• Machine Learning

Background:

• Microscopic analysis of forensic microtraces is time-consuming and labor-intensive.
• Automated methods are needed to assist forensic experts in trace identification and classification.

Purpose of the Study:

• To develop and evaluate deep learning models for automated microtrace recognition in forensic investigations.
• To investigate pretraining strategies to minimize the required data annotation workload for these models.

Main Methods:

• Utilized deep learning for pixel-wise classification of tape-lift samples to localize and classify microtraces (fibers, hairs, skin, glass, sand).
• Compared various pretraining strategies: ImageNet pretraining, self-supervised learning pretraining, and a sequential combination of both.
• Evaluated model performance based on prediction accuracy and reduction in manual annotation effort.

Main Results:

• Pretrained models reduced the need for annotated data twofold compared to training from scratch, while maintaining prediction accuracy.
• Combining ImageNet and self-supervised pretraining yielded the highest accuracy, achieving a fourfold reduction in manual annotation.
• Mean intersection over union improved from 0.34 to 0.56 due to pretraining, demonstrating enhanced recognition and localization.

Conclusions:

• Deep learning models, particularly when utilizing combined pretraining strategies, offer a robust foundation for the automated analysis of forensic microtrace scans.
• These automated methods can significantly improve efficiency and accuracy in forensic investigations by reducing manual annotation workload.