Research On TEDS Character Recognition Algorithm Based On Deep Learning

With the rapid development of high-speed railway in China, high-speed EMU’s security issues are increasingly paid attention to. Trouble of moving EMU Detection System(TEDS) as a technology of moving EMU monitoring have been gradually implemented in some important railway stations during the last two years. The existing TEDS identify types by the numerical characters that printed on both sides of EMU, and establish a correspondence between real-time images and standard model images to achieve fault diagnosis automatically. Therefore, correct character identification is a prerequisite for TEDS to work normally. In some extreme cases, those EMU characters are difficult to identified by existing image recognition algorithm due to uneven illumination, distortion and mirror reflection.In recent years, the deep learning theory becomes a hot research topic both in academia and applications, and it is widely used in big data analysis. As a deep learning model, convolutional neural networks becomes an important model on both image classification and speech recognition by virtue of a multilayer feature extraction and sharing weight. After analyzed the data images, the research is mainly concentrated on convolutional neural network identification which includes the following three aspects:Firstly, this paper bring up an improved stroke width transform algorithm for locating character region under complex illumination after analyzed some existing localization algorithms. Then, in character segmentation part, this paper decided to use vertical projection based on gray level correction to divide image into single characters. Finally, considering the data quantity, complexity and other factors, this paper bring up a character recognition plan based on convolutional neural network after optimizated input and output layer, hidden layer structure and activation function, and analyzed the results of parameter adjustment.Experimental results show that optimized and improved convolution neural network can recognize characters efficiently. Under complicated illumination, the correct rate of EMU character recognition was 99.22%.