Reconhecimento de estilos de escrita em documentos manuscritos: uma abordagem em nível de documentos, linhas e palavras

Abstract

Script recognition is a necessary step in automating the recognition of Optical Character Recognition (OCR) systems. The task of script recognition in handwritten documents is challenging due to the similarity between some of these styles. Our main objective in this paper is to analyze and evaluate the performance of robust techniques in eleven different script recognition. In addition to performance, we analyzed the impact of working with document, line, and word-level recognition. The Local Binary Patterns (LBP), Local Phase Quantization (LPQ), and Speeded Up Robust Features (SURF) texture descriptors were used in conjunction with the Support Vector Machine (SVM) classifier. The experiments were performed in a database with eleven classes, in order to achieve hit rates similar to those presented in the literature. The best hit rate was 97.73% using the document-level SURF texture descriptor. In line and word levels the best performance was 96% and 94.37%, respectively, using the LPQ descriptor. Through our experiments, we find that the best performance applies to the document level, followed by lines and words. We also noted that the performance improvement of handwriting recognition applications is not only related to the amount of samples used in the training process.