Application Research Of Handwritten Text Recognition Technology Based On Deep Learning

Optical character recognition is one of the most important research fields in pattern recognition,and it has significant application prospects in mail sorting,traffic management and bank ticket identification.The study of minority languages optical character recognition contributes to the construction of infrastructure in minority areas,and it not only promotes the development of their economy,culture and education but also helps to purify the network environment.What's more,the recognition of ethnic characters are not fully excavated except Uygur,Mongolian,and Tibetan.According to statistics,Yi users are more than one million,so the research of Yi recognition is necessary.In this paper,we used deep learning method to study the recognition of handwritten Yi character.Aiming at the lack of dataset,we studied the application of clustering algorithm in the data labeling,and established the handwritten Yi dataset as well as explored the parameter tuning of deep learning network,Specific work is as follows:(1)In this paper,we improved a density-based clustering algorithm with information entropy and applied it to the construction of handwritten Yi dataset.The results showed that this algorithm not only improved the efficiency of data labeling but also obtained better results than the original algorithm in the UCI clustering dataset.(2)In this paper,the parameter tuning of convolution neural network are explored in the design process of convolutional neural network.We compared and discussed the effect of the mainly four parameters on the accuracy of convolutional neural network,and they are:the number of convolution layers,the number of convolution filters,batch-size and learning rate.Then we gave some experience on parameter tuning and used these experience designed the convolution neural network.The experimental results showed that this network get an accuracy of 99.65%on the dataset which contains 100 classes.(3)In this paper,we constructed and opened a handwritten Yi dataset to promote the development of Yi recognition research.The APP of Yi dictionary designed in this paper brings convenience to the researchers and Yi interested scholars.In summary,the proposed clustering algorithm which based on information entropy and density as well as the experience about the parameter tuning of convolution neural network are useful for other optical character recognition issues and can be used in other similar pattern recognition problems.In addition,by modifying the distance measurement method,the clustering algorithms can be used in more unsupervised learning problems.