Footwear impressions retrieval through textures and local features

Abstract

The proposed artefact applies pre-processing filters, extracts key features, and retrieves the relevant matches from a shoeprint impression repository. Two functions were utilized for matching impressions. One function is texture based and creates an MPEG-1 movie out of two input images and employs the size of the output movie as a similarity measure. The other function is local feature based and uses SURF feature extraction and MSAC for matching. For pre-processing of the prints, a set of well-known techniques were employed. Also, we implemented a technique to facilitate better matching through splitting the input prints into smaller prints and then matching on these. FID 300 is a publicly available dataset of footwear impressions in greyscale. It comes with 1175 reference prints (e.g. sole images from tip to heel), and 300 prints lifted from real crime scenes, the latter being incomplete and with low image quality. The evaluation was done over various options and always against all reference prints in the FID 300. Clearly the evaluation results are affected by the quality of the lifted images. Evaluations were done in three batches (each having different pre-processing): first, all crime scene prints with the texture function got an average accuracy of 61%; second, a sample of 43 lifted prints with the texture function got 65% average accuracy; third, all crime scene prints and the local feature function applied got 50% average accuracy.