Content-based image retrieval (CBIR) system in digital forensics

Abstract

Contrary to popular perception, footprints and footwear impressions are more commonly found at crime scenes than fingerprints and DNA. Systems for the analysis and authentication of the latter, have been thoroughly researched throughout the years and are almost perfected. However, the same progress has not been made for footprints. Possible explanations are that footprints are underestimated as evidence, because no shoe sole is unique even though each wears distinctively. Due to the lack of research in this area, a forensic analyst must compare the images manually which is time consuming. This dissertation aims to provide an automatic footprint extraction and correlation system. The system proposed is able to apply pre-processing, extract key features and retrieve the relevant matches from a footwear impression repository. The repository used for this study contains both prints from real crime scenes and prints from newly manufactured shoes known as reference prints. In order to compare images, a distance function was utilised. This function creates a MPEG movie out of the images and employs the size of the movie in order to calculate the similarity. For pre-processing of the prints, apart from common techniques, an original concept of tessellations was applied. The main process of the system takes a crime scene print and compares it to all the reference prints in the repository. Once the distances are calculated by the distance function, an ordered list of prints is generated in ascending order, from most similar to least. The results obtained from the development of this project have shown that the accuracy achieved depends on the quality of the images that are being used. Comparisons are done in two batches: first all crime scene prints are compared with all the reference prints, then the procedure is repeated with pre-processing being applied to the prints.