| With the rapid development of high speed railway in China,it is essential to ensure the safety of railway operation.As the key equipment to ensure the safety of railway operation in the station,the safety and reliability of computer interlocking system directly affects the efficiency and safety of railway operation.In order to ensure that the interlocking system can strictly complete the functions of each part,computer interlocking test is very important.In recent years,with the rapid development of deep learning and image processing technology,in this thesis,the result judgment module in interlocking automatic test is studied,and the deep learning algorithm is used to realize the location and recognition of interlocking interface signals,switches,section names and icons on the interlocking interface.First,the theoretical basis of interlocking test and deep learning related technologies are introduced,and the advantages and disadvantages of deep learning and machine learning as well as the framework of deep learning are analyzed and summarized,and deep learning algorithms and development framework based on Tensor Flow is determined to adopt.By analyzing the structural characteristics,optimization methods and training process of Neural Network,Convolutional Neural Network(CNN)and Recurrent Neural Networks(RNN),the location and recognition algorithm and processes of text and icons on the interlocking interface are determined.Secondly,in the text location algorithm of interlocking interface based on CTPN,the key frames is firstly extracted from the test video data to obtain text location images.VGG16 and Bi LSTM(Bi-directional Long Short Term Memory neural network)feature extraction networks were used to extract spatial and sequential features of text respectively.Then,the text proposal and text line construction modules are used to detect and connect text proposals respectively,and the line text to be located is obtained to complete the text location task of interlocking interface.The experimental results show that the CTPN text location algorithm combining CNN,RNN and RPN(Region Proposal Network)can better complete the text location task of interlocking interface,and the location speed and accuracy of the model meet the requirements of interlocking automatic test.Thirdly,a character recognition algorithm combining traditional character segmentation and improved Alex Net network is proposed in view of few types of text recognition and character distribution law of interlocking interface.By analyzing the advantages of common character segmentation algorithms,a two-level character segmentation method combining horizontal projection and vertical projection is determined.The effect of character segmentation by two-level character segmentation method is good,which resolves various disturbances on the text image of the interlocking interface;Alex Net network is improved to meet the character characteristics,categories and recognition accuracy requirements of interlocking interface.The experimental results show that the improved network not only improves its own performance,but also shows advantages compared with other models.Fourthly,in the interlocking interface icon location algorithm based on Faster R-CNN,according to the characteristics of interlocking interface icons,common target detection algorithms are selected for comparative experiments,and the Faster R-CNN network is selected as the interlocking interface icon location algorithm.By comparing the effects of different convolutional neural networks on extracting image features,it is determined that ZF network is used to extract image features,which solves the disadvantages of traditional detection networks in feature extraction.Finally,the network structure is designed and the model is trained.Finally,in the interlocking interface icon recognition algorithm based on improved Le Net-5,the advantages and disadvantages of the classical Le Net-5 network are analyzed,and the algorithm suitable for interlocking interface icon recognition is obtained by greatly improving the network structure.Through the comparison experiment of all the networks before and after the structure change,the superiority of the improved model is proved,and its recognition accuracy is greatly improved. |
