Reconhecimento de caracteres manuscritos off-line utilizando Support Vector Machine (SVM)

Abstract

This work investigates the use of Support Vector Machine (SVM's) for recognition of uppercase handwritten characters of the Latin alphabet. The experiment made use of offline data from IRONOFF database. The data had already been gone through thresholding and bounding box preprocessing techniques. With regard to extraction of characteristics, concavity and convexity were obtained? detected? isolated? determined? observed? by labeling of the background pixel. Subsequently, the perceptual zoning mechanism was applied by dividing the characters into Z parts (z = 0, z = 4, z = 5 horizontal, z = 5 vertical and z = 7). The data was divided into training and testing sets to create generalist and expert SVMs. The experiments were performed through the use of the WEKA tool. Kernel configurations (linear, radial and sigmoid) were applied to SVMs thus creating a total of 15 generalist and 390 specialist SVMs. SVMs with linear kernel configuration with z = 5h and z = 5v zoning achieved better performance with 94.4% and 94.7% hit averages respectively. The results were compared with the Neural Networks proposed by Aires in 2005, where all SMV's results were superior to those of the NNs. The biggest difference was in the? zoning z = 5h, where, the Neural Networks had an average of 82.4% accuracy and the SVM of 94.4%, while the smallest difference was in the? zoning z = 7 with an average of accuracy? of 88.9% and 94,1% for NNs and SVMs respectively.