Likelihood Ratios for Deep Neural Networks in Face Comparison.

Area of Science:

• Forensic Science
• Computer Vision
• Biometrics

Background:

• Forensic facial comparison is crucial for legal proceedings.
• Transparency and open-source methods are vital in forensic science.
• Automated systems offer potential support to human experts.

Purpose of the Study:

• To compare the performance of automated facial comparison systems and human forensic experts.
• To evaluate the utility of machine learning in supporting courtroom evidence.
• To assess likelihood ratio computation for facial comparisons.

Main Methods:

• Utilized three open-source convolutional neural network systems: OpenFace, SeetaFace, and FaceNet.
• Converted system outputs (distance/similarity) to likelihood ratios using Weibull distribution, kernel density estimation, and isotonic regression.
• Compared system performance against forensic investigators using low- and good-quality frontal images.

Main Results:

• Automated systems achieved 100% precision and specificity for non-match detection with low-quality images, surpassing investigators (89% and 86%).
• Forensic experts performed better with good-quality images.
• A rank correlation of approximately 80% was observed between investigators and software performance.

Conclusions:

• Open-source facial comparison software can assist forensic reporting officers with faster, reliable comparisons, particularly for full-frontal images.
• Automated systems demonstrate potential as a supportive tool for human experts in forensic casework.
• The study highlights the complementary roles of automated systems and human expertise in facial identification.