Automation-Supported Curation of Large Forensic Image Databases

The Forensic Anthropology Center has multiple photographs of each donor (approximately 100 donors annually since 2012) taken daily throughout the entire decomposition process. To date, the collection contains over one million photographs from more than 500 donors. To make the collection accessible, the ICPUTRD: Image Cloud Platform for Use in Tagging and Research on Decomposition was created and used to tag over 1000 images by an expert in human decomposition using a nomenclature of standard terms commonly used in the medicolegal community. Tags are area of an image containing a forensic feature, e.g., scavenging, larvae, discoloration, or purge. ICPUTRD can potentially transform significant aspects of forensic work by enabling re-use of existing digital data in forensic investigations or research. While the ICPUTRD platform has begun to organize, standardize and create a valuable research tool in forensic science, several obstacles must be removed to realize its full potential. First, how would an investigator find a complete set of images that document the phenomena under investigation? Second, how could the information in the relevant images be quantified to answer, for example, plausible timelines (determining the time of death)? Since manually tagging the entire collection is not feasible, we propose to a) train deep learning models on the set of existing tags and use these trained models to tag the remaining million images; b) implement new capabilities in ICPUTRD that make it possible for experts in human decomposition to evaluate and improve the accuracy of these model-generated tags with minimal effort; c) organize the detailed multivariate-temporal data representing the incidence of the forensic features hundreds of donors and integrate it with suitable statistical analysis techniques; and d) evaluate the effectiveness of the system by conducting a study with participants involving forensic researchers and professionals. Initial models in the objective a) and the validation infrastructure for the objective b) will be implemented in 2019, with additional models and manual validation effort continuing until the end of 2020. The objective c) will be implemented in 2020 and the objective d) towards the end of 2020. The proposed research will cross-reference massive image corpus with standard human decomposition-relevant concepts with a validated degree of accuracy. This annotated corpus would serve as a basis to produce highly accurate models of human decomposition thus enabling correspondingly more accurate forensic determinations for criminal justice. The results of this work will be submitted for publication in forensic and machine learning journals. This project contains a research and/or development component, as defined in applicable law, and complies with Part 200 Uniform Requirements - 2 CFR 200.210(a)(14). CA/NCF