Research On Text Localization And Recognition Of Train Number In Natural Scenes

The detection and identification of freight train number is an important link in the informatization and intelligent construction of the Railway Transportation Management Information System(TMIS).Its main function is to realize the automatic identification of train number and train number information,and provide accurate,real-time and complete basic train information such as train number for the railway transportation management information system and the comprehensive management information system.Since entering the information age,computer vision technology has been widely used in various neighborhoods,bringing about no small industrial upgrading.Applying this technology to the detection and identification of train number can not only reduce part of the cost,but also accurately and real-time detect train information.However,there are still many difficulties in the detection and identification of train numbers operating in natural scenes: firstly,the monitoring equipment at the station is not a professional high-speed camera,and the train monitoring recorded by this equipment will have serious motion smear;secondly,the installation of monitoring equipment The angle is not directly facing the number area of the train,so most of the car numbers in the surveillance video are tilted and distorted;in addition,due to the different types of cars,the position of the painted car number in the car body is not fixed,and the character interval of the car number will also be affected by the conditions of the car body.Influence;Finally,during the long-term operation of the train,the car number will be defaced,and the ideal lighting conditions cannot be fully obtained,which has a great impact on the car number recognition.Many existing natural scene text recognition methods cannot be directly applied to train operation scenarios,and it is difficult to achieve ideal results.In response to the above problems,this paper studies the positioning and identification of train numbers to realize the positioning and identification of train numbers in high-performance natural scenes.In the detection and location phase of freight train number,this paper proposes a train number location method based on region segmentation.First,the Res Net which introduces the attention of the channel domain is used to extract features,so that the model pays more attention to the convolution channel information,redistributes the weight from the perspective of the channel domain,and improves the positioning accuracy of the vehicle number;Then,the feature pyramid and bottom-up path enhancement module are used to fuse the multi-scale feature map,and the strong positioning features in the shallow network are spread to the deep network,so as to accurately locate the vehicle number area from the complex environment and reduce the vehicle number detection rate;Finally,the progressive scale expansion module based on the breadth-first algorithm is used to expand and segment the fused feature map from small scale to large scale,and the loss function combining the set similarity measure Dice coefficient is used to classify and regression the segmentation results,and the positioning results are output.The experimental results show that the positioning accuracy of the network model proposed in this paper is 97.47%,the recall rate is 94.14%,the comprehensive F1-score is 95.78%,and a prediction time of about 0.2seconds for a single image.In the identification phase of freight train numbers,this paper proposes a connectionist temporal classification based train number identification method.Using the idea of rectification first and then recognition,the psyllium images were first corrected using spatial transformation network,then residual networks were used to extract features,and finally the predicted psyllium results were exported using a combined output of a two-way long and short memory network and connectionist temporal classification.The experimental results show that through training and validation on a real train number image dataset,the proposed network model has a recognition accuracy of 94.66% and can predict 299 train number cropped images per second.