Feature extraction and classifier design are the key problems for pattern classification. Numerous feature extraction methods and classification models have been proposed and applied successfully in the past decades. However, most of them are only suitable for small-category problems and are based on the i.i.d. assumption (independently and identically distributed). Therefore, they cannot fulfill the requirements of real applications such as the handwritten Chinese character recognition (HCCR) problem, which is typical of large category set. For HCCR, the traditional Fisher linear discriminant analysis (FDA) cannot overcome the class separation problem, while the large variability of handwriting styles across individuals breaks the i.i.d. assumption and makes HCCR a challenging problem. To deal with “large category” and “non i.i.d.” problems, from the perspectives of dimensionality reduction, classifier design, and classifier adaptation, this thesis proposed four effective methods summarized as follows. 1. Large category dimensionality reduction based on weighted Fisher criteria (WFC). To solve the class separation problem of the traditional FDA model, using a weighting function to emphasize the close class pairs has been proposed to obtain a better reduced subspace. We evaluate different WFC with five weighting functions and three weighting spaces comprehensively, and further, propose a nonparametric WFC method which can achieve the best performance in handwritten Chinese character recognition. 2. Locally smoothed modified quadratic discriminant function (LSMQDF). To deal with the over-fitting problem of modified quadratic discriminant function (MQDF), we propose the LSMQDF which smoothes the covariance matrix of each class with its neighboring classes. As a regularization to avoid over-fitting and also an extension of the global smoothing method, LSMQDF can improve the generalization performance significantly. 3. Classifier adaptation with style transfer mapping (STM). To deal with the non i.i.d. problem, we propose a classifier adaptation model based on STM, which maps a source point set towards a target point set. The objective function of STM is a convex quadratic programming problem and therefore STM has a closed-form solution. STM can be combined with different types of classifiers for supervised, unsupervised, and semi-supervised adaptation. The experiments on a large scale online handwritten Chinese character recognition problem showed ...
修改评论