A finely tuned deep transfer learning algorithm to compare outsole images

Abstract

In forensic practice, evaluating shoeprint evidence is challenging because the differences between images of two different outsoles can be subtle. In this paper, we propose a deep transfer learning-based matching algorithm called the Shoe-MS algorithm that quantifies the similarity between two outsole images. The Shoe-MS algorithm consists of a Siamese neural network for two input images followed by a transfer learning component to extract features from outsole impression images. The added layers are finely tuned using images of shoe soles. To test the performance of the method we propose, we use a study dataset that is both realistic and challenging. The pairs of images for which we know ground truth include (1) close non-matches and (2) mock-crime scene pairs. The Shoe-MS algorithm performedwell in terms of prediction accuracy andwas able to determine the source of pairs of outsole images, even when comparisonswere challenging. When using a score-based likelihood ratio, the algorithm made the correct decision with high probability in a test of the hypothesis that images had a common source. An important advantage of the proposed approach is that pairs of images can be compared without alignment. In initial tests, Shoe-MS exhibited better-discriminating power than existing methods.