Feedback-Based Algorithm for Handwritten Character Recognition
Zhu Qiang Xiao · Chinese Journal of Computers · 2002
Many state of art OCR systems with acceptable accuracy have been developed during passed years, but further improvement of the performance seems difficult. One of the difficulties lies in the trade off between the recognition rate and the rejection rate. In another words, rejection methods are applied to increase the system's accuracy, but decrease the system's overall recognition rate and performance at the meantime. In this paper, a new algorithm of handwritten character recognition is proposed to increase the system's accuracy without decreasing the recognition rate, which is based on the combination of the feedback concept in the cybernetic field with the BP algorithm in the ANN architecture. The proposed system is designed as an effective neural network by adding confidence back propagation and input modification (feedback) to the ANN model and the learning algorithm, so that preprocessing and recognition parts are integrated closely. The feedback signal is acquired through the difference of supposed output (predefined) and real output, and is back propagated to modify the input image signal in the feature space using the gradient descent algorithm. The modification can be considered as a step of the preprocessing phase in the whole system architecture, which can get rid of many kinds of noise showed in the experiment part of this paper. In order to get rid of the interference of the models, specified feedback network is trained for various models. During the recognition phase, only the candidate models in the predefined similarity set are processed by the feedback network to decrease the computational complexity. The final result is given by the model with minimum feedback cost.The proof of the convergence of the feedback algorithm is presented in this paper, and detailed experiments are made on sample sets with different characteristics, which showed that the error rate in such result feedback neural network architecture can be greatly reduced and the robustness to environmental noise increased.