Research On Locomotive Text Area Detection Based On Convolutional Neural Network

Nowadays,with the rapid development of railway industry,and the improvement of transport efficiency in China,the number of trains has increased dramatically.The maintenance and repair of locomotives(head of train)has become an important guarantee for the normal operation of railway trains.At present,locomotive maintenance points in China generally rely on manual fault detection,but the efficiency of manual detection is low,staffs are easy to fatigue,and there is a certain risk,which affects the normal operation of locomotives.In the rapid development of railway locomotive transportation,the traditional railway working mode urgently needs to realize automatic,intelligent and informationalized management.Based on the requirement of information management of railway locomotive maintenance points and the method of deep learning combined with image processing,this thesis extracts multiscale and multi-objective text area of locomotive side automatically.The work can be used as a basis of follow-up identification and automatic management and other information processing work.According to the actual situation of locomotive samples,this thesis realizes the precise localization of text area.The main work includes the following:(1)According to the characteristics of locomotive field samples,an image preprocessing method is designed.Because of the particularity of locomotive samples in railway field,the sample images with more interference factors need to be preprocessed to remove the noise such as rainwater and sediment,such as using Gauss filter to remove the noise of sample images.The quality of night samples needs to be enhanced.Because there are still spot effects after processing night samples,it is necessary to use filtering method to denoise.(2)The text areas of locomotive samples are roughly located based on convolutional neural network.A self-defined convolutional neural network is used to extract the features of the text area in the pre-processed sample image.After training,learning and discriminating,a text area discriminating grid is generated.Then the text area in the original sample image is roughly located,and then the rough location area is marked.(3)An algorithm for merging regional grids after rough localization is designed.The coarse localization areas are merged,and the labeled areas are merged into several vertical labeled areas by vertical merging.Then the vertical labeled areas are merged horizontally to mark the text areas accurately.(4)The text area detection module is designed.Then,the tests and comparisons using different scale samples and different deep learning network structures are conducted.The next,our algorithm is compared with the mathematical morphology method which is commonly used in text area detection.Finally,our algorithm is also compared with other methods of deep learning.In this thesis,the convolution network structure and sample size are debugged and optimized step by step based on deep learning and image processing methods.The locomotive text area can be detected very well,and the locomotive text area can be accurately located.The accuracy rate can reach 93.3%.