A theory of classifier combination: the neural network approach

There is a trend in recent OCR development to improve system performance by combining recognition results of several complementary algorithms. This thesis examines the classi er combination problem under strict separation of the classi er and combinator design. None other than the fact that every classi er has the same input and output speci cation is assumed about the training, design or implementation of the classi ers. A general theory of combination should possess the following properties. It must be able to combine any type of classi ers regardless of the level of information contents in the outputs. In addition, a general combinator must be able to combine any mixture of classi er types and utilize all information available. Since classi er independence is di cult to achieve and to detect, it is essential for a combinator to handle correlated classi ers robustly. Although the performance of a robust (against correlation) combinator can be improved by adding classi ers indiscriminantly, it is generally of interest to achieve comparable performance with the minimum number of classi ers. Therefore, the combinator should have the ability to eliminate redundant classi ers. Furthermore, it is desirable to have a complexity control mechanism for the combinator. In the past, simpli cations come from assumptions and constraints imposed by the system designers. In the general theory, there should be a mechanism to reduce solution complexity by exercising non-classi er-speci c constraints. Finally, a combinator should capture classi er/image dependencies. Nearly all combination methods have ignored the fact that classi er performances (and outputs) depend on various image characteristics, and this dependency is manifested in classi er output patterns in relation to input images. Capturing the dependency improves the theoretical error bound of the combinator. This thesis de nes a framework to separate the combinator design from classi er speci c details. Then we present a combination theory based on the neural network approach that possesses all the properties mentioned above. Moreover, in facing these issues, we discover several interesting ndings involving the concept of classi er bootstrapping, the de nition for classi er independence and dynamic classi er selection. Experimental results on handwritten digits recognition verify our theory and ndings.