Forensic Comparison of Soil Particles Using Gaussian Mixture Models and Likelihood Ratio Test

• Published in: 2023 15th International Congress on Ultra Modern Telecommunications and Control Systems and Workshops (ICUMT) pp. 188 - 192
• Main Authors: Jonak, Martin, Dorazil, Jan, Kolarik, Martin, Jezek, Stepan, Burget, Radim, Kotrly, Marek
• Format: Conference Proceeding
• Language: English
• Published: IEEE 30-10-2023
• Subjects: deep learning; Error analysis; Forensics; Gaussian mixture model; Image segmentation; Minerals; Mixture models; Scanning electron microscopy; segmentation; Soil; soil analysis
• ISSN: 2157-023X
• DOI: 10.1109/ICUMT61075.2023.10333101

Forensic analysis of soil traces can be highly valuable in criminal investigation as it can provide evidence which links a person or a contaminated object with one specific location. Given two soil samples, the task of a forensic expert is typically to decide whether they both originate from the same location or not. To confidently answer this question it is necessary to perform a complex analysis focused on examination of the sample's organic, anthropogenic, and naturally occurring components. In this paper, we focus on one element of the analysis which studies small mineral particles within the sample. In particular, we propose a novel method for automatic comparison of soil particles, using scanning electron microscope images acquired in the secondary electron (SE) or backscattered electron mode. The method involves segmentation of particles, identification of their contours, and extraction of local descriptors from the images, which are then used to train a sample specific and non-specific Gaussian mixture model (GMM). Finally, a likelihood ratio, based on the GMMs, is calculated to assess the odds that two samples originate from the same location. The proposed method, utilizing Root SIFT descriptors extracted from the SE images along the particle contours, achieved an equal error rate of 13.1 % and an area under the curve of 95.2 %, surpassing our baseline method derived from particle size analysis.