Research On Automatic Recognition Of Braille Based On Deep Learning

Braille,as the tool of text communication for visually impaired people,it guarantees the right to write and read for millions of blind people,and the automatic identification of braille means a lot to the blind school teachers,the braille ancient books managem ent and the blind guardians.In this paper,several parts of braille automatic recognition were researched by combining the method of gray-scale projection and deep learning.The main contents of this paper are as follows:Firstly,based on the same row or the same blind square geometry,a method of braille correction based on gray-scale projection integral graph was proposed in this paper.According to the law of the Angle between the braille gray scale projection and the braille deflection,the braille de flection Angle was determined and corrected.The experiment shows that this method can achieve the deflection correction of braille.Secondly,all blind spots in braille are not connected to each other and general word segmentation couldn't be used.In this paper,the blind spot was determined according to the gray projection integral graph,the middle line between two blind squares was found according to the fixed rule of the gray scale,all the braille characters were separated by those cross-vertical dividing lines.The experimental result shows that this method is very effective for the segmentation of large braille characters.Finally,the recognition rate of traditional braille recognition is not very good,and the feature points need to be extracte d manually,aimed at these shortages,combined with deep learning methods,braille recognition based on deep neural network was studied.Many convolutional neural networks of different structures were constructed with caffe,which is a deep learning framework.The PRe LU activation function which has stronger sparse representation was used instead of the Re LU activation function to avoid the problem of neuronal death and gradient disappearance during network training;Dropconnect was used as a substitute for general Dropout,it could avoid overfitting and enhances the feature learning ability of the network.Comparison experiments of feature extraction layer,full connection layer and data set image size were set up,the optimal scheme was determined accordin g to the recognition accuracy and operation speed of these network models.In the production of experimental data sets,the actual use of braille identification is fully considered,the braille images were collected in a variety of environments and a lot o f noises were added.With this,the application of braille identification system is improved.The experimental result shows that the accuracy of braille recognition is above 99%,the accuracy and practicability of braille recognition system were greatly improved.This paper has some significance for the research and development of related instruments.