The Research On High Speed Train Number Recognition Algorithm Based On Image Processing

At present,China is being in the climax of high-speed railway construction and it will finally form the high-speed railway network of "eight horizontals and eight verticals"according to "the mid-long term planning for China's railway network".With the increase of mileage and speed,more attention has been paid to normal operation of the train.Pantograph-strip is the key component for the train to get the electric energy from the contact line.When monitor its abnormal situation using image processing technique,we need to quickly determine the train number information to promptly eliminate safety hazards and report to the higher authorities.The relationship between the pantograph state and train number can't be established using traditional RFID recognition technology,however it can satisfy the requirements by taking pantograph image and train numbe image to recognize number at the same time,which is the foundation for further detecting the abnormal state of pantograph fault.The work of this paper is based on the pantograph monitoring device.The method makes use of slide train number data captured by camera,accomplished the high-speed train number automatic recognition task after number location,number correction and number segmentation.The main work of the paper include the following aspects:?1?During the image preprocessing,images which are affected by the weather or the environment such as over exposure,darkness and fog are enhanced by the MSER.?2?In the process of accurately locating number,firstly,we filter the image contours and sets the maximum stroke threshold in advance aiming at the problems of large amount of calculation and inaccurate positioning in SWT.Then stroke width transform is executed.Finally,dilation operation makes the whole train number as an entirety to locate during the formation of number candidate region.Experiment indicates the improved stroke width transform algorithm computing time is reduced by about 70%compared with the original algorithm.?3?Focusing on the existing number perspective distortion problem,we firstly analysize that the traditional algorithm based on the vanishing point is not adaptive.Then TILT algorithm is introduced and mathematical model is established to correcte train number.Finally,a new algorithm for contour detection based on character spacing is introduced during a single number character segmentation process.This algorithm has a higher accuracy than the conventional contour segmentation algorithm for the segmentation accuracy of the fracture and the adhesion character.?4?In the process of single character recognition,firstly,the small sample data is enhanced to meet the needs of the training network.Then the CNN is designed based on the classic LeNet₅,which is proved to be reasonable.The experimental results show that the classification accuracy of 11 kinds of characters can be up to 99.9%.