DIFFERENCE THEORETIC TEXTURE FEATURES ALONG THREE ORTHOGONAL PLANES FOR FACE RECOGNITION FROM VIDEOS

Abstract

The efficiency of a human face recognition system depends on the capability of the system to recognize faces accurately in the presence of different changes in the appearance of face. The appearance of a person may change with varying lighting conditions, facial expressions, occlusions or facial features, such as beard, mustache and glasses. Several face recognition techniques have been introduced by researchers which are being used for many practical applications such as identity verification at highly secured locations, checking of criminal database records, identifying a person from surveillance cameras at a public place, etc. While recognition using static images are common, face recognition from videos has become an active topic in the field of object recognition and computer vision. Video-based face recognition techniques have gained widespread interest primarily based on the idea expressed by psychologists from their studies that humans recognize faces through motion especially in most cases when the spatial image quality is low. It has always been the aim of researchers to make the computer behave like a human being and therefore, to make the computer recognize faces like humans do has been one of its goals. In order to do this, a robust face recognition algorithm is required and in this project a new feature set called difference theoretic texture feature set along three orthogonal planes (DTTF-TOP) is proposed which will help extract the facial features of a person from videos and thus further use it for recognition. The challenge of change in face appearance due to the different face expressions showing emotions such as anger, sad, happy, surprise, etc is taken up in this project. So, a dataset called the Extended Cohn-Kanade (CK+) database containing videos of person showing different emotions is chosen and the one nearest neighbor and SVM classifiers are used for the classification purpose in the experiments. Results from the newly proposed feature set are compared with the highly efficient local binary pattern along three orthogonal planes (LBP-TOP). Further comparisons are made with a template based cross correlation (TBCC) method and a face feature weighted fusion method based on fuzzy membership degree for video based face recognition. Face recognition done using the proposed DTTF-TOP is found to give better performance than the other techniques used for comparison in the experiments.