Abstract:
Partial face occlusions such as scarfs, masks and sunglasses compromise face recognition accuracy. This thesis presents a face recognition approach robust to partial occlusions. The approach adopted for this study is based on the assumption that the human visual system ignores occlusion and solely focuses on the non-occluded sections for recognition. Four sections derived from a whole/ un-occluded image and the whole face are used to train a classifier for recognition. For testing, an occluded face image is also divided into the four sections above from which, the non-occluded or the least occluded section is selected for recognition. Two strategies were used for occlusion detection; skin detection and the use of haar cascade classifiers. This thesis mitigated weaknesses from literature review such as use of datasets that simulate real world occlusion scenarios, use of less data in training and not requiring any type of occlusion variation in training data. Additionally, the classifier performed relatively well in the classification task with an accuracy of 92% on the Webface-OCC, 96% on the Pubfig, 92% on the FaceScrub, 96% on the Yale B and 92% on the LFW datasets.
