| In this thesis, a new approach to the problem of face (object) recognition is proposed. We introduce, a novel methodology that exploits the advantages offered by active vision architectures, and utilizes highly compressed feature representations for person identification. Specifically, we describe a unified model of low-level visual attention that combines purely data-driven processes with primitive object recognition mechanisms and model-based reasoning, and show that such process can form the foundations of a high performance face recognition system. The proposed architecture employs a number of independent, parallel visual routines responsible for object localization, identification, and scene interpretation, corresponding to the "where" and "what" channels of visual perception. The model is biologically plausible and is motivated by processing strategies in the human visual system.; To test the validity of the proposed face recognition architecture, a number of experiments were carried out on a large and varied face database (FERET). The active vision components were used to detect faces by locating their individual facial landmarks, and to derive a compact face code invariant to changes in viewing geometry and imaging conditions. Simulation results on 100 subjects (216 images) demonstrated that both the "where" and "what" channels perform with high accuracy, and their combined performance reached 100% in detecting all relevant facial landmarks. The identification experiments achieved 89.6% accuracy for the match task and reached 100% for surveillance. These results indicate that the proposed mechanisms are capable of efficiently locating and encoding information relevant for all aspects of face recognition.; The major contributions of this work are as follows: (1) A parallel framework of visual attention capable of integrating bottom-up (data-driven) and top-down (knowledge-driven) attentional processing, corresponding to the "where" and "what" pathways in the human visual system. (2) A novel, low-level image feature extraction technique for measuring local conspicuity, and an integration mechanism to derive saliency information as per need of the recognition process (on demand). (3) The development of spatially non-uniform iconic object representations that encode local image characteristics in an illumination independent manner. (4) The design and construction of neural-network-based visual filters for object recognition, and an appropriate iterative training strategy capable of constructing optimal learning sets. The non-orthogonal, topologically organized feature descriptors developed by these algorithms provide efficient mechanisms to encode real world objects--and faces in particular--as needed for identification.; Although the discussion in this thesis is primarily centered around problems related to face recognition, the methodologies developed are of general nature and can be efficiently used to address many similar problems in the fields of computational vision, pattern recognition and robotics. |
