Criminal and victim identification based on soft biometrics

Abstract

Criminal and victim identification is always important in forensic investigation. However, it can be a very challenging problem for identifying criminals and victims in digital media when only their non-facial body sites are available in evidence images. These criminals and victims can be masked gunmen, paedophiles, and victims in child pornographic and voyeur images.To solve the above problem, several novel alignment and identification approaches are proposed in this thesis. Firstly, lower leg geometry is proposed as a soft biometric trait for criminal and victim identification. This study provides a foundation for further research based on body geometry. Secondly, leg geometry and hair follicles are proposed to align the androgenic hair patterns in consideration of viewpoint and pose variations, which were ignored by a recent paper suggesting androgenic hair patterns for identification.Experiments on 1,138 high and low resolution images from 283 different legs show that the proposed alignment algorithms provide improvements of 5%-10% on different experimental settings.Thirdly, a new approach is developed to improve the identification of androgenic hair patterns significantly. In the past, it was believed that androgenic hair patterns in low resolution images are not a distinctive biometric trait because of the previous result. A new algorithm, which makes use of leg geometry to align lower leg images, large feature sets (about 60,000 features) extracted through multi-directional gridding systems to increase discriminative power and robustness, the partial least squares (PLS) method to handle imbalanced training data and to perform the multi-grid feature fusion, and scheme I thank my fellow labmates in Biometrics and