Abstract
Degraded printed character recognition is a hard and ever-present problem in optical character recognition. Previous work has explored serial combination of multilayer perceptron (MLP) and autoassociators networks for printed character recognition. The MLP is used as a first step classifier for its discrimination capability and is not very well suited for rejection. On the other hand, the autoassociator used as a second step classifier is more appropriate when a very small error and rejection are required. Unfortunately, this better behavior with respect to rejection is either paid in terms of rejection error or in terms of computational complexity, particularly when the number of classes is high. In this paper, we propose a serial combination of the Hopfield and MLP networks in order to achieve accurate recognition of degraded printed characters. We introduce a relative distance and use it as a quality measurement of the degraded character, which makes the Hopfield-based classifier very powerful and very well-suited for rejection. This relative distance is compared to a rejection threshold in order to accept or reject the incoming degraded character by the Hopfield model used as a first classifier. Due to its discrimination capability, the MLP network is used as a second classifier to avoid rejection error and to diminish computational complexity. The proposed method is devoted to solving the problem of recognition of single font characters collected from poor quality bank checks. We report experimental results from a comparison of three neural architectures: the Hopfield network, the MLP-based classifier, and the proposed combined architecture. The proposed method is also compared to five other recognition systems. It is shown that the proposed architecture exhibits the best performance, with no significant increase in the computational burden. In this paper, we propose also a bank check processing procedure for account check number (ACN) detection, localization, and character retrieval...
