APLICAÇÃO DE REDES NEURAIS CONVOLUCIONAIS NO RECONHECIMENTO DE CARACTERES EM PLACAS INFORMATIVAS JAPONESAS

Abstract

Unlike most Western countries, which have a Latin-derived base alphabet, Japan has two syllabic alphabets called Hiragana and Katakana, and a Chinese alphabet, called Kanji. The vast differences in the writing of these Eastern alphabets to Western alphabets, Western alphabet-based OCR algorithms tend not to efficiently detect Japanese characters. This work contributes to a methodology applying digital image processing techniques, such as color range-based segmentation, edge detection and mathematical morphology techniques, to detect Japanese traffic informationalplates correctly the perspective and segment the characters contained in it. A convolutional neural network wasused to perform the classification of Hiragana characters contained in the segmented plates, withaccuracyof 94.37%.