University of Technology Sydney: Enlisting AI to improve crime scene image classification

Ever-increasing volumes, diversity and complexity of digital data from crime scenes, such as CCTV footage, digital photos, emails and documents, create significant challenges for forensic investigators who need to catalogue, label and search this information.

In Australia, the Australian Federal Police (AFP) has well over 300,000 images relating to various illicit drug operations alone. These images capture a huge amount of information, from the physical appearance of the drugs to names and addresses on parcels.

As part of an ARC Linkage project on drug intelligence, researchers from UTS and the AFP have collaborated to explore the use of two different machine learning methods to automatically classify and catalogue images of drug-related offences, to improve the speed and accuracy of the process, and enhance workflows.

“Like society in general, forensic science has undergone a digital transformation and laboratories have become heavily reliant on information technology,” says Distinguished Professor Claude Roux, Director of the Centre for Forensic Science at UTS.

“Digital forensic science is a rapidly growing field, and digital traces now feature in most criminal cases. However, classifying and cataloguing images, as well as other electronic evidence, in a searchable database is a time-consuming task,” he says.

The researchers used a subset of 60,520 real-world forensic casework images from the AFP’s illicit drug database, which had been labelled into well-defined categories by AFP personnel, to assess the potential of the machine learning methods.

They found that a hierarchical Deep Convolution Neural Network method called Tree-CNN was the most effective for cataloguing images, offering greater flexibility than more traditional machine learning models.

Once images are classified, specialised algorithms can be designed to extract specific information such as the physical characteristics of tablets – size, colour and logos – to help establish profiles of the illicit drugs.

The researchers hope these new methods will be taken up by the AFP for further development and use, not only in illicit drug operations but across all crime scene image databases, particularly where cross-referencing or complex queries are required.