Methodology for the design of NN-based month-word recognizers written on Brazilian bank checks

The study of handwritten words is tied to the development of recognition methods to be used in real-world applications involving handwritten words, such as bank checks, postal envelopes, and handwritten texts, among others. In this work, the focus is handwritten words in the context of Brazilian bank checks, specifically the months of the year, and no restrictions are placed on the types or styles of writing or the number of writers. A global feature set and two architectures of artificial neural networks (ANN) are evaluated for classification of the words. The objectives are to evaluate the performance of conventional and class-modular multiple-layer perceptron (MLP) architectures, to develop a rejection mechanism based on multiple thresholds, and to analyze the behavior of the feature set proposed in the two architectures. The experimental results demonstrate the superiority of the class-modular architecture over the conventional MLP architecture. A rejection mechanism with multiple thresholds demonstrates favorable performance in both architectures. The feature set analysis shows the importance of the structural primitives such as concavities and convexities, and perceptual primitives such as ascenders and descenders. The experimental results reveal a recognition rate of 81.75% without the rejection mechanism, and a reliability rate 91.52% with a rejection rate of 25.33%.