Research On Video And Image Localization And Recognition Of Train Number

The key links in the automation of cargo inspection operations include the localization and recognition of train numbers,that is,the number information can be automatically recorded,saving manpower and material resources.In recent years,the application of computer vision technology to train number recognition can not only save costs,but also realize automatic monitoring of trains,which brings convenience to railway number recognition.However,there are also many difficulties in identifying the train number in the image or video: first,the area of the train number area in the panoramic image of the train is very small(less than 0.41%),and there are many text disturbances other than the train number;second,the train number It has the characteristics of changeable position and variable character spacing;again,the train number recognition is extremely susceptible to light,the structure of the train body is complex,the background is changeable,the cabin is stained and the perspective deformation is disturbed.Existing natural scene text recognition methods cannot be directly used for the tasks in this thesis,and it is difficult to achieve the desired results.On the one hand,it is easy to miss the car number target.On the other hand,the word-level text localization method is easy to locate the train number with a large number interval into two parts.At the same time,the accuracy of the train number recognition with perspective deformation is low.In view of the above difficulties,this thesis proposes video-oriented and image-oriented train number localization and recognition methods to achieve robust freight train number localization and recognition.The main contents of this thesis are as follows:(1)For static images,in the localization stage of freight train number,create a freight train number detection data set,and propose an image-oriented train number localization method.First,a basic model based on deep learning is used to locate the train number,which is not effective.Therefore,this thesis proposes three improvements based on this model to obtain the train number localization model.Aiming at the problem that small targets are easily missed,the feature maps of different scales are fused to generate text candidate regions to solve the multi-resolution problem in localization.For the problem that the complete train number cannot be accurately located,not only the horizontal regression layer is added,but also a border-sensitive fine-grained text box(BSF)and hard sample mining strategy(hard example mining,HEM).After improvements,the train number detection F1-Measure is 0.94.The experimental test data is composed of 2109 panoramic images of train cars marked in this thesis,including images such as night,day,different models,and variable train number scales.The detection speed of each image is 0.19s(2)For static images,at the stage of freight train number recognition,the train number recognition method based on the attention mechanism is implemented,that is,the image-oriented train number recognition method,and other advanced train number sequence recognition methods for comparison.The train number recognition method based on attention mechanism is composed of three parts: convolution layer,coding layer and decoding layer.After testing,the overall F1-Measure of train number recognition was 0.81.The car number character recognition F1-Measure is 0.94.The recognition speed of each image is 0.04s(3)In view of video data,in order to realize the localization and recognition of freight train number,a video number detection and recognition data set is constructed,and a video-oriented deformation number recognition method is proposed.Because the train number in the video has a perspective deformation caused by the perspective of the surveillance camera,it is not appropriate to use the static text recognition method.The method proposed in this thesis is a multi-stage train number recognition method.Firstly,the video train number localization method based on tracking is designed to reduce the false detection and missed detection in view of the problem of time redundancy and poor character definition of the video car number.Secondly,for deformed train numbers,an end-to-end train number recognition method with a built-in correction network is used.Finally,a video-based train number sequence recommendation strategy is proposed,which further utilizes the trajectory to optimize the recognition results.The experimental data is a video segment containing 7086 frames.After the experiment,the overall recognition number F1-Measure of the train number is 0.91.The car number character recognition F1-Measure is 0.99.The recognition speed of each car number image is 0.23s...