Comparing footwear impressions that are close non-matches using correlation-based approaches

Area of Science:

• Forensic Science
• Computer Vision
• Pattern Recognition

Background:

• Forensic footwear analysis traditionally relies on subjective visual comparisons.
• Quantitative methods are increasingly needed for objective assessment of footwear evidence.
• Distinguishing between impressions from identical shoes and those from similar, non-matching shoes is crucial for evidential value.

Purpose of the Study:

• To evaluate the effectiveness of various similarity metrics in discriminating between footwear impressions.
• To compare correlation-based metrics, including those derived from convolutional neural networks, for footwear analysis.
• To determine the best performing metric for assessing the evidential value of footwear impressions.

Main Methods:

• The study examined correlation-based similarity metrics: normalized cross-correlation, phase-only correlation, AvNCC, and AvPOC.
• AvNCC and AvPOC utilized features extracted from a convolutional neural network.
• Experiments were conducted on three distinct footwear impression datasets: Everspry, FBI boot, and West Virginia University collections.

Main Results:

• Phase-only correlation demonstrated performance comparable to or exceeding other examined metrics across all datasets.
• The study confirmed the necessity of considering wear patterns and individual characteristics for accurate discrimination.
• Quantitative metrics show promise in enhancing the objectivity of footwear impression comparisons.

Conclusions:

• Phase-only correlation is a robust metric for forensic footwear analysis.
• Quantitative similarity metrics can significantly improve the assessment of evidential value in footwear comparisons.
• Further development of computational approaches is recommended for forensic pattern analysis.