Return to Workshop Listing| Author: | Silviu Minut |
| Advisor: | Sridhar Mahadevan |
| Email: | minutsil@cps.msu.edu; http://www.cps.msu.edu/~minutsil |
Face recognition is a very well-studied problem in computer vision. However, the majority of current approaches use a non-biologically plausible imaging process where the face is viewed at constant resolution. The most successful systems are also non-incremental and require subsampling of the image. We present an alternative approach that is incremental, uses a sequential foveal image processing model, and that can scale gracefully to arbitrary image sizes. We simulate foveated vision in software, by transforming a constant resolution image into a variable resolution image in which the resolution is acute in a small patch (the fovea) and decreases exponentially from the fovea towards the periphery. For each individual in a database of faces we train a hidden-Markov model (HMM) classifier. The observation sequences used to learn the HMMs are generated from foveated images, by fixating on different regions of a face. We present experimental results, comparing two foveal HMM classifiers with a more traditional HMM classifier built by subsampling the image. The results show that foveal processing significantly outperforms subsampling for HMM-based recognition on a database of 133 faces, but ultimate performance is still not as good as classical pattern recognition methods (eigenfaces).