Region based partial face recognition

If faces are under occlusions like sunglasses, scarfs etc., classical face recognition algorithms could not succeed in recognition task. For this purpose, there are a lot of studies in the literature which firstly determine occlusions on faces and then perform recognition task on unoccluded parts. In some of these studies, satisfactory success rates haven?t been achieved; in some of others, colors have been used for determining occlusion. Studies which are using color information are not acceptable enough for real life applications, because occlusions can be in any form with various colors. Furthermore, the step of determining occlusions boosts computation cost. Recently, successful results are achieved by applying linear convergence techniques on occlusion problems. In this study, an approach with a two-step architecture is proposed; in the first step kNN (k-Nearest neighbor) classifier is used and in the second step LR (Linear regression) classifier is used. In this approach, both local and holistic features are utilized. Feature extraction is applied by using Euclidean map where variation information on the image is utilized. In Euclidean map, pixel?s euclidean distance to its neighboring pixels are calculated for each point. By using such a technique, unnecessary variations can be thrown out and variation range that is used in recognition task can be specified. In this study, AR database was used in experiments for face recognition. In experiments, two kinds of training set are used; the first one includes single image per subject and the second one includes multiple images per subject. For testing, images that include sunglasses and scarf of the same subjects are used. For multiple training samples per subject case, 100% face recognition success rate in sunglasses occlusion experiments and 98% face recognition success rate in scarf occlusion experiments were achieved. In single training sample per subject case, 98% face recognition success rate in sunglasses occlusion experiments and 96% face recognition success rate in scarf occlusion experiments were achieved.