Feature Set Selection and Weighting for Legal Amount Recognition on Brazilian Bank Checks

This work presents a study about feature selection and weighting for improving the recognition of handwritten words coming from Brazilian bank check lexicon. For this purpose, two global optimization methods are used: Tabu Search(TS) and Simulated Annealing(SA). These methods were combined with k-NN composing two hybrid approaches for features selection and weighting: SA/k-NN and TS/k-NN. The results show that feature sets optimized by the studied models are very efficient when compared with k-NN. Both, accuracy classification and number of features in the resultant set are considered in the conclusions. Furthermore, some new structural features extracted from image upper and lower profiles are proposed.