Research Of Train Number Localization And Recognition Method

With the increasing demand for freight transportation,railway freight transportation has become the main component of railway transportation.Promoting the construction of railway safety monitoring and management system has also become an important part of railway freight management.The original cargo inspection work mainly relies on manual inspection,which is outdated and time-consuming,and no longer meets the modern requirements of freight management.Realizing the automation of freight inspection operation is the key point in the construction of railway safety monitoring and management system,and realizing the localization and recognition of freight train number is the most basic step in the automation of freight inspection operation.Freight train number(about 400*80 pixels)occupies a very small area in the whole input image(about 8375*2048 pixels),and there are a large number of interference texts.In addition,due to the influence of environment,background,illumination,vehicle type and other factors,it brings great challenges to the localization of freight train number.Because of the uncertain number interval,the broken or adhered or missing train characters,it is difficult to identify the number of freight train.In this paper,the methods based on visual image processing technology and scene text recognition technology are proposed to realize the localization and recognition of freight train number.The main contents of this paper are as follows:(1)In the localization stage of freight train number,a two-stage localization method of freight train number is proposed.In the first stage,in order to reduce the missing detection rate of complex train number data with large spacing,a deep learning method based on improved scene target detection is adopted to detect train number.In the second stage,the candidate train number regions are processed,and the second screening of candidate train number regions is realized by using the technology of "histogram of oriented gradient + support vector machine"."We collected and labeled 4500 train image of difficult and easy for testing and analysis.Compared with the original detection method,the experimental result has improved by nearly 10%and achieved 90%localization precision.Each image only needs 0.3 seconds of processing time.(2)In the recognition stage of freight train number,two kinds of train number recognition algorithms are studied.Firstly,a method of segmentation and recognition is proposed.For solving the problem of breaking and cohesion of train number characters effectively,a dynamic window segmentation algorithm is proposed to segment the area of train number into independent characters.Then,support vector machine is used to recognize train number.In addition,an end-to-end recognition method based on deep learning is also used.Firstly,the convolutional neural network is used to extract the image features,and the recurrent neural network is used to learn the sequence features of train number.Finally,the connectionist temporal classification is used to transform the final recognition results.We labeled 2000 train number images and compare the two recognition methods.The dynamic window segmentation is more in line with the characters with uneven width and internal breakage.The freight train number recognition based on deep learning avoids error transmission,and achieves 91%accuracy,with recognition speed of 0.04 seconds per train number image.(3)In order to make full use of the experimental data,a multi-angle fusion algorithm is proposed to improve the system recognition rate in our freight train number recognition system.The fusion strategy includes the fusion of localization results and recognition results,aiming to output the only result.We quantitatively analyzed the multi-stage freight train number localization and recognition system and two-stage freight train number localization and recognition system on 3000 images.Experiments show that the recognition accuracy of the two-stage freight train number system reaches 91%,and the fusion strategy improves the recognition accuracy of the system by at least 3%.