| Summary: | 碩士 === 東海大學 === 統計學系 === 94 === Following the advancement of epoch, the recognizing technology has attained excellent result already. Some results are even more formidable than human recognition ability. In the field of statistics, the recognizing question is a kind of classification problem actually. In this research, the main purpose is to have recognizing ability up to certain level and to decrease the time spending on recognition at the same time. For character recognition problems, we consult several methods to proceed in this thesis. One of those methods that we follow closely is “ Tangent Distance (TD)“ proposed by Simard et al. (1993). The key idea is to estimate the minimum distance between two patterns by using different direction of tangent vectors. These tangent vectors include x-translation, y-translation, rotation, scaling, parallel hyperbolic transformation, diagonal hyperbolic transformation and thickening. Another method we considered is “Discriminant Adaptive Nearest-Neighbor “ (DANN) proposed by Hastie and Tibshirani (1996). They consider the variation in every clusters of observations near the target point to find out
direction of variation. Combining the idea of TD and DANN, we use also different linkage clustering methods to reduce prototypes and retain representatives from training data set. We combine these concepts and propose some new methods.
Here, we use ZIP data set which is one of typical handwritten digit recognition data set to demonstrate our methods. The data set comes from handwritten zip codes that appeared on some envelopes of U.S. mail passing through the Buffalo, NY post office. The digits were written by many different people with a great variety of writing styles and instruments. Each digit is converted into a 16 by 16 pixel image after some preprocessing. There are 7291 training data and 2007 testing data in the data set. By our new methods that we proposed, the best prediction error is 0.0259. The second best prediction error is 0.0393 and the classification time spends 14.25 minutes for classifying 2007 digits. That is 0.43 second for each digit. For another faster processing method we proposed, it only took 32 seconds to classify 2007 digits. That is 0.0159 second for each digit. However, the error rate increases a little bit to 0.0418.
|
|---|
