Image set based ear recognition using novel dictionary learning and classification scheme

In this work ear recognition of a moving person with the help of a single fixed-in-position video camera is investigated, a novel problem undertaken to the best of our knowledge and belief. The challenges associated with this work are that during data capturing process each moving subject appears in unconstrained views together with random noise, motion blur and different level of illumination. Secondly, the collection of large number of redundant samples for each subject makes the training database bulky, which eventually increases classification time. So, in this paper, we introduce a novel, reduced time, metaface dictionary learning approach which by employing Frobenius norm based regularization reduces redundancy of training database with much lesser time as compared to other available database learning methods. In general, ear samples with random poses, severe motion blur and different levels of illumination loose significant class specific information and hence inflict severe nonlinearity to the system. The simple and well accounted solution for the above problem is kernel framework which makes samples of different classes linearly separable by elevating them to higher dimensions. In our proposed solution, we have used novel l2l2-norm regularized affine hull based kernel collaborative representation based classification scheme, which represents each query set as an affine hull and then collaboratively represents this hull over the linear span of gallery sets of all classes in the high dimensional space. Finally, the query set is assigned to that particular class which gives least representation or residual error among all the available classes. Results of extensive experimentations carried out over an indigenously developed database in our laboratory (named ERVIDJU) aptly demonstrate that our proposed method is a valid biometric identifier and it can produce superior performance compared to several other contemporary algorithms, developed for similar purposes.