| This thesis work mainly concerns the problem of design, implementation and integration of software modules for the automated structural analysis of difficult face images, such as those with eyeglasses, mustaches/beards, very dark or very bright complexions, and faint feature contrast. Perhaps of greater importance is a unified analysis approach that treats both side-view and frontal-view face images within a common framework. Previous systems, each designed to deal with either frontal- or profile-view images, but not both, have often failed when encountered the aforementioned complexities of face images.; The first part of the thesis, Chapters 2 through 5, describes the system core: an integrated image analysis system with a broad variety of new image operators and algorithms: fast and robust grayscale segmentation, grouping, basic shape analysis, and robust feature detection that can cope with strong glares on eyeglasses yet also be sensitive enough to detect very faint features. The system core basically locates the face roughly, detects feature blobs, and preclassifies them into candidates for the mouth, eyes, and eyeglasses.; The second part of the thesis, Chapters 6 and 7, describes recognition modules based on a new rule induction (RISCC) and a novel neural networks (BBP) algorithms. These recognition modules enable the system to: (i) identify the mouth among the candidates; (ii) identify the view-class (one out of possible five, from profile-left to profile-right) using DNF rules induced by RISCC; and (iii) using trained BBP neural networks, recognize the eyeglasses pieces, and decide if an eyeglass pair exists. The RISCC algorithm (better than C4.5), and especially the BBP neural network algorithm, are novel algorithms with high performance. On the 2-bit XOR problem, a BBP neural network can learn at 20-40+ times as fast as a back-propagation neural network.; The view-classification accuracy achieved is about 93% (42/45); and that for the eye-glasses recognition is conservatively estimated at about 95% (18-19/19), with one false-positive (9.1%) out of eleven cases without eyeglasses. Many profiles and half-profiles, eyeglasses with glares or dark spectacles, beards and/or mustaches, and very faint feature contrast due to very dark complexions or weak lighting, etc., have been analyzed well. |
