3D face recognition using low resolution imagery

Abstract

Computer vision, which is one of the most researched areas in the field of computer science, has raised the interest of a wide variety of people, ranging from professionals in neuroscience to experts in psychology. Together with the idea of making computers comprehend and reason about actions occurring in a video sequence, such interest is also attributed to the successful implementation of such systems in a wide variety of fields, including automatic surveillance, and time and attendance systems. Research in the field of automated face recognition started in the 1960s, leading to the development of the first semi-automated system that required human intervention for the localisation of features before distances and ratios to a common reference point could be calculated for comparison to reference data. In this thesis, we investigate the various research carried out in the field of automated face recognition, whilst focusing on techniques of face detection, the extraction of features from each of the detected faces, and the fitting of a reference model to the extracted feature points, the result of which is compared to reference data. Research suggests that while face recognition in 3D space is capable of achieving better results than its 20 counterpart, systems making use of both technologies proved to be more robust than any of the other two system categories. Consequently, in order to increase the robustness of our system, the presented artefact was developed in such a way as to exploit both technologies. When presented with a video frame, the system first performs image pre-processing through the application of the Gaussian filter so as to remove any background noise which might be present in the frame, after which Histogram Equalization is applied so as increase the overall contrast of the frame by evenly distributing the brightness across the entire image, resulting in an increase in the level of detail in the frame. A number of face detection algorithms were implemented and compared so as to determine the one capable of obtaining the best results. In particular, a face detection technique based on skin colour detection was implemented and compared with two other techniques based on Viola and Jones's approach, one of which performs some filtering so as to further decrease the false detection rate, whilst keeping the false dismissal rate to a minimum. Feature points extraction was carried out through the application of the Active Appearance Model approach, whereby a Point Distribution Model {PDM) was constructed from 240 manually labelled face images so as to be subsequently fit to novel face images, the result of which is a set of 2D points representing the locations of the corresponding feature points as indicated by the trained PDM. Such points, together with the corresponding 3D points on a reference 3D face model are then employed for the recovery of the head pose in the form of a rotation matrix and translation vector, which are in turn used for mapping the principal components of a 3D morphable model from 3D to 2D space, so as to be able to compute the shape coefficients of the personalized 3D model. Different model fitting techniques were investigated and compared, and the combination of algorithms capable of obtaining the best overall results in acceptable time spans, was outlined.