A comparative study of one-class classification methods for bloodstain detection in hyperspectral forensic imaging

Blood is one of the most important and common types of physical evidence left at crime scenes and used by forensic investigators to determine the dynamics of violent crimes. Therefore, it is important to be able to detect blood stains at the crime scene but without contaminating it. For this reason, recently, classifying hyperspectral images has gained popularity in the procedures to detect blood at the crime scene with respect to traditional approaches based on chemical tests. However, defining a classifier that distinguishes between blood spectra and all the other possible ones requires to train machine learning models with all possible classes of substances left at a crime scene. Since this is no a realistic task, the current approaches perform a classification between blood spectra and a limited number of other substances. In order to overcome this limited vision, this paper proposes, for the first time, to detect blood at the crime scene using one-class classification methods that have the feature of being trained only with samples of the target class (in our case, blood class). In detail, the main goal of this paper is to compare several one-class classification methods and analyse their performance. As shown in the experimental session involving a publicly available hyperspectral-based bloodstain dataset, one-class classification methods result in accurate techniques to detect blood at the crime scene.