The Research Of Braille Recognition Based On Deep Transfer Learning

Braille recognition is a very critical step in the research of braille information processing.It is not only of great significance to the braille workers,but also helps to promote the communication between the blind and the sighted people in real life,so as to promote the development of China's information accessibility.Because the traditional method of braille recognition is greatly influenced by the environment and equipment,and the labor cost of marking samples is too high to realize the automatic recognition of braille,the development of effective braille recognition method is of great significance for the development of information accessibility in China.At present,deep learning methods have been applied to the study of braille recognition,and automatic braille recognition has been realized.However,at present,the research of deep learning in the field of braille recognition is still based on the dataset made by researchers themselves.Braille pictures are relatively standardized,and there is no public braille public dataset to verify the effectiveness of the algorithm.Limitations when considering the acquisition of braille pictures more and there are certain conditions,so the reality of braille image datasets are usually small and identification is difficult,and need to develop a braille identification module to solve the public production of braille image dataset and realistic scenarios of braille image datasets to identify problems.Deep transfer learning,as the current mainstream of machine learning algorithm,has been successfully applied in many practical scenarios.Therefore,in this paper,on the basis of the existing braille recognition algorithm,image datasets from real scenarios braille identification perspective,using the existing large scale,high standardization degree of braille image datasets and introduce deep transfer learning methods,practical stronger braille image recognition model is established.The main contents of this paper are as follows:First,the existing research methods of braille recognition are analyzed and sorted out.And the deep learning method and the deep transfer learning method are briefly explained,focusing on the two convolutional neural network models used in our work,and the two convolutional neural network models and deep transfer learning are analyzed in principle advantages and some practical applications in the field of image recognition.Then,a deep transfer learning algorithm named Domain Auto-alignment(DAA)was proposed to reduce the domain distribution differences between two braille image datasets in order to address the differences between the two braille image datasets.Firstly,the sample embedding distributions of the source and target domains are preliminarily aligned to obtain the highly batch-normalized embedding vectors of the source and target domains in the intermediate state,and then the Maximum Mean Discrepency(MMD)is used for fine measurement to continue to reduce the distribution difference between the embeddings of the two domains.At the same time,under the deep learning framework Caffe,using the GoogleNet network model as the carrier of the deep transfer network,a variety of transfer network models with different structures are built.By selecting from the classification accuracy of the network model,the complexity of the network structure,the selection of measuring and comparing the network convergence speed and so on,to determine the best transfer network model.The validity of the DAA algorithm was verified on the Office-31 and Office-Caltech datasets.Finally,for the recognition problem of braille image datasets in real scenes,a braille recognition study based on deep transfer learning algorithm was carried out.The large-scale and high-standard braille image dataset A was used to simulate the standard braille image dataset and used as the source domain input of transfer learning,the braille image dataset B with a small scale and poor standardization,simulates the braille image dataset in a real scene and is used as the target domain input for deep transfer learning.Dataset A mainly comes from the "Information Accessibility Research Center" jointly established with the China Blind Association and the China Disabled Persons' Federation.It is obtained by collecting and manually checking each type of braille points;the braille image dataset B is mainly taken by mobile phones and captured by the web page.In the process of making the experimental dataset,we selected as many braille images as possible in the actual environment to improve the robustness of the model.The experimental results show that the deep transfer learning method can effectively help the identification of braille image datasets in real scenes.This paper uses deep transfer learning to deal with the classification problem of braille images in real scenes,which enriches the research of deep transfer learning methods in the field of braille image classification,broadens the thinking of braille recognition research,and at the same time,the method and braille machine are used in subsequent studies.The combination of translation methods provides new ideas for the expansion of existing braille datasets and the production of public braille image datasets.