SGER: Grid-to-grid neural networks for innovative pose invariant face recognition

University Of Memphis, Memphis TN

Investigators

Abstract

SGER: Grid-to-grid neural networks for innovative pose invariant face recognition This project will develop and demonstrate the initial building blocks for a whole new paradigm for pattern recognition, applicable to pose invariant face recognition. INTELLECTUAL MERIT: Conventional pattern recognition systems develop or train fixed pathways of computation, from the raw input to the final classification or description. These pathways usually start out from complex, cleverly programmed and/or hand crafted but fixed preprocessors or feature extractors. In the new paradigm, the transforms are performed by a new class of practical tools now available which learn mappings from data defined over a 2D image grid to outputs defined over a grid or to global summary variables. In the proposed paradigm, more powerful mapping of image data will be adapted so as to maximize performance in the pose invariant face recognition task. This project will include exploitation of newer structures such as the cellular simultaneous recurrent net (CSRN), which are hoped to provide greater transformational capability. AT&T developed something similar years ago, for ZIP code digit recognition, which was the best system available for that task -- but it was only able to do feed forward analysis, which made it unsuitable for handling more complex images such as faces. This project will go as far as possible to address a series of benchmark challenges, from the maze traversing problem to pose invariant face recognition in 2D image grid. BROADER IMPACTS: Face recognition is an area of great importance to homeland security, among other applications. The PI has been investigating affine transformation (such as scale, translation, rotation, clutter etc.) invariant recognition in objects and faces for last several years. There has been intense research in different aspects of transformation invariant face recognition. However, current systems are typically very inaccurate in recognizing the same face after many months -- an issue of great practical importance. Another critical shortcoming of existing face recognition techniques is that if we have seen faces only from one viewing angle, in general, it is difficult to recognize the faces from disparate angles. The proposed novel CSRN paradigm results from analysis of the pose invariant face recognition problem and of how to correct the problem. It is anticipated that a quantum improvement in performs a will result. It is also hoped that this will shed light on the question of how the human brain achieves such capabilities, which is important in turn to a deeper understanding of learning and intelligence in the brain.