Unsupervised statistical methods for processing of image sequences
Michael S. Gray, Javier R. Movellan, Terrence J. Sejnowski · 1998
The performance of pattern recognition systems depends on preprocessing. Unsupervised learning techniques can be used to construct statistically efficient preprocessing filters from an ensemble of images. The efficiency of local and global principal component and independent component image representations was examined on the task of visually recognizing the first four digits spoken in English using a hidden Markov model (HMM) for the recognition system. Local representations consistently outperformed global representations in generalizing to new speakers. In addition, the use of a novel regression-based variable selection technique substantially boosted performance. Global representations, on the other hand, performed better than local ones for speaker identification tasks. These results imply that local structure in an image contains more information than global structure for lipreading, a task that requires the extraction of speaker invariant information from sequences of images.