Automatically classifying crime scene images using machine learning methodologies

Over the last several decades, forensic science has undergone a digital transformation whereby laboratories have become heavily reliant on information technology. Although computer-based tools have helped automate processes in forensic science, the digital transformation is not without serious technical challenges. Ever-increasing volumes, diversity, and complexity of data has overwhelmed forensic practitioners despite technological advances. In particular, it is increasingly difficult for forensic practitioners to successfully conduct the manual content labelling and cataloguing of crime scene image databases. In this paper, we investigate the application of two different machine learning classifier models on a rich collection of real-world forensic casework images of drug-related offences sourced from the Australian Federal Police illicit drug database, as proof-of-concepts for the automatic classification of crime scene images. The casework database includes 97,287 illicit drug-related casework images that have been labelled into well-defined categories by AFP personnel. The first model is based on a Support Vector Machine (SVM) classifier combined with Bag-of-Visual-Words (BoVW) visual dictionaries derived from various local image feature descriptors. The second model is based on a hierarchical Deep Convolution Neural Network (DCNN) method called Tree-CNN. Experimental results based on a subset of 60,520 images from the illicit drug database reveal average True Positive classification rates for the BoVW-based and Tree-CNN models of 66.48% and 89.17%, respectively. Furthermore, a closer assessment of the classification results suggests that the Tree-CNN model has the greatest potential for further development and use in real-world applications.