A Quantitative Approach for Forensic Footwear Quality Assessment using Machine and Deep Learning

Forensic footwear impressions play a crucial role in criminal investigations, assisting in possible suspect identification. The quality of an impression collected from a crime scene directly impacts the forensic information that can be garnered from any future comparison, which in turn impacts the performance of matching algorithms, and an examiner’s opinion of source association. However, accurately assessing the quality of footwear impressions remains a challenging task; at present, there is no standard definition or methodology to assess the quality of this domain-specific imagery. In this paper, we propose a quantitative approach to predict impression quality utilizing Machine Learning (ML) and Deep Learning (DL) algorithms. A publicly available footwear impression dataset was used to crowd-source quality labels using a five-point rating scale. Subsequently, each image was decomposed into a series of features, extracted using traditional image processing techniques, and through transfer learning using a pre-trained VGG16 model. Random Forest (RF) and Multinomial Logistic Regression (MLR) classifiers were trained on these extracted features to predict quality (i.e., very poor, poor, moderate, good, and excellent) using crowd-sourced opinions as ground truth. Results indicate success ranging from 80% to 100% depending on the approach used, and when accuracy is defined by no more than one quality-level difference between prediction and ground truth (e.g., good versus excellent, or very poor versus poor). The highest accuracy was associated with transfer learning, and the results lay the foundation for a reference-free and standardized quality assessment model for forensic footwear applications.