Characterizing Human Expertise Using Computational Metrics of Feature Diagnosticity in a Pattern Matching Task

Area of Science:

• Forensic Science
• Cognitive Science
• Computer Vision

Background:

• Forensic evidence evaluation, particularly fingerprint comparison, relies on human expertise.
• Assessing the strength of evidence requires considering the rarity of observed features among potential sources.
• Current computer systems lag behind human experts in fingerprint comparison accuracy.

Purpose of the Study:

• To characterize human expertise in fingerprint analysis using computational models.
• To quantify the diagnosticity of fingerprint regions for excluding alternative sources.
• To validate computational models using expert eye gaze data.

Main Methods:

• Applied the Attention via Information Maximization (AIM) model to quantify region rarity and diagnosticity.
• Utilized the CoVar model to capture low-level feature relationships, mimicking early visual processing.
• Collected and analyzed expert eye gaze data to tune and validate the computational models.

Main Results:

• Both AIM and CoVar models achieved 75%-80% performance in classifying expert gaze patterns.
• A validation study confirmed that experts perform better when focusing on diagnosticity-ranked regions identified by the AIM model.
• The models provide a quantitative measure of evidence strength based on feature rarity.

Conclusions:

• Computational models can effectively capture aspects of human expertise in fingerprint analysis.
• Quantifying feature diagnosticity can enhance the reliability of forensic evidence evaluation.
• These metrics may serve as a safeguard against wrongful convictions by providing objective strength-of-evidence assessments.