Feature extraction by best anisotropic Haar bases in an OCR system

摘要

In this contribution, we explore the best basis paradigm for in feature extraction. According to this paradigm, a library of bases is built and the best basis is found for a given signal class with respect to some cost measure. We aim at constructing a library of anisotropic bases that are suitable for the class of 2-D binarized character images. We consider two, a dyadic and a non-dyadic generalization scheme of the Haar wavelet packets that lead to anisotropic bases. For the non-dyadic case, generalized Fibonacci p-trees are used to derive the space division structure of the transform. Both schemes allow for an efficient O(Nlog N) best basis search algorithm. The so built extended library of anisotropic Haar bases is used in the problem of optical character recognition. A special case, namely recognition of characters from very low resolution, noisy TV images is investigated. The best Haar basis found is then used in the feature extraction stage of a standard OCR system. We achieve very promising recognition rates for experimental databases of synthetic and real images separated into 59 classes.