Evaluation of Character Recognition Systems

A self-organizing neural network model for mechanism of pattern recognition unaected by shift in position. of Japanese Kanji using principal component analysis as a preprocessor to an articial neural etwork. 10 necessary, only one initial guess was used. These results show that for character classication accuracy NN methods and statistical methods have comparable accuracies conrming the COCR results. 4 Conclusions Examination of the results of 11 OCR systems using a wide variety of recognition algorithms has shown that in accuracy and writer independence NN systems have not demonstrated a clear cut superiority over statistical methods. Some neural systems have higher accuracy than statistical methods; others have lower accuracy. The performance of statistical methods is more closely grouped and is approximately the same as the performance of an average NN system considered here. One area where NN's may have an advantage is in speed of implementation and recognition. Examination of Table 3 show that on OCR classication the ranking of the methods is similar. The neighbor-based methods are the most accurate with PNN being the best of these. The comparison of MLP and RBF methods shows that RBF is usually the better method. When MLP and RBF methods are compared to multicluster EMD and QMD methods the NN methods are more straightforward to implement but do not show a clear accuracy advantage. All of the experiments presented here also suggest that the training set sizes used, although large, are not sucient to fully saturate most of the machine learning methods studied here.