Advanced facial recognition for digital forensics

Forensic facial recognition has become an essential requirement in criminal investigations due to the advent of electronic devices such as CCTV, digital cameras, mobile phones, and computers and the huge volume of content that exists. Forensic facial recognition goes beyond facial recognition in that it deals with facial images under unconstraint and non-ideal conditions, such as poor image resolution, facial orientation, illumination, expression, and the presence of accessories. These conditions have a huge impact on the recognition performance. A wide variety of facial recognition algorithms exist, each more or less susceptible to various environmental conditions. This paper proposes a multi-algorithmic fusion approach by utilising multiple commercial facial recognition systems to overcome particular weaknesses in singular approaches to obtain the best facial identification accuracy. The advantage of focusing upon commercial systems is because it releases the forensic team from developing and managing their own solutions and subsequently also benefits from state of the art updates in underlying recognition performance. A set of experiments were conducted to evaluate three commercial facial recognition systems (Neurotechnology, Microsoft, and Amazon Rekognition) to determine their individual performance using facial images with varied conditions. The second experiment sought to determine the benefits of fusion. Two challenging facial datasets were identified for the evaluation; the first was a publically available dataset known as ‘CAS-PEAL-R1’. The second dataset represents a more challenging yet realistic set of digital forensics scenarios collected from publically available celebrity photographs. The experimental results have proved that using the developed fusion approach achieves better identification rate as the best tested commercial system has scored 67.23%, while the multi-algorithmic fusion system scored an accuracy of 71.6%.