- Sparse Related Visual Learning, Self-taught Learning and Feature Extractions:
Though there are many image decomposition methods, it is hard to get both of
the basis and the features to be independent without the normal
distribution assumption. Recent research shows that sparseness and
other constraints will lead to part-based representations results,
which is similar to the receptive fields in V1 cortex in human Brain.
Sparse Coding, Sparse Bayesian Learning and Compressive Sensing have
been proposed for pattern learning, feature extraction, denoising and
compression during the past 10 years.
In vision, self-taught learning means studying the knowledge from free-cost images in our natural environment; it is an active area in machine learning in recent years. The significance of self-taught learning is to revisit the fact that sometimes not only the labeled target data but also the relevant unlabeled data are hard to get, while at the same time the basic patterns can be embedded in the general data although it is unlabeled and with quite different distribution.
a model-based feature extraction approach, which uses micro-structure
modeling to design adaptive micro-patterns. We first model the
micro-structure of the image by Pair-wise Markov Random Field. Then we
give the generalized definition of micro-pattern based on the model.
After that, we define the fitness function and compute the fitness
index to encode the image’s local fitness to micro-patterns.
Papers: ICIP08, AMFG-ICCV05, US Patent
- Face Recognitions and Similarity Measurement:
Papers: ICIP05 and Microsoft Techfest Demo 2005
- Geometric Methods and Applications:
Papers: Electronics Letters 2008 and CISS08