Selective function learning neural network which unifies conflicting results of multiple methods for distorted handprinted Kanji pattern recognition

We present a new integrated character recognition system involving two unification neural networks which unifies the disparate recognition results of multiple methods. The unification neural networks process the discriminants of each category to accurately select the correct candidate. The training allows the unification networks to automatically form various conflicting relationships between the discriminants of each method. The new learning scheme shares tasks among the two unification neural networks whether the multiple recognition methods fail to agree on the same candidate or not. The system achieves a higher recognition rate than any individual method-an ordinary method using a linear combination of the discriminants, or a multilayer perceptron. The unification neural networks form a mechanism that derives the correct category from conflicting results, and is useful for promoting recognition applications that demand high reliability.