Research On Multi-object Recognition Of Video Image Of Traction Substation Based On Deep Learning

With the development of high-speed railway technology in China and the improvement of automation requirements,more and more traction substations are developing to unattended,and the intelligence of video surveillance and inspection system in substations has become particularly important.Therefore,the purpose of this paper is to provide technical support for remote intelligent patrol inspection of traction substation,and mainly study how to realize multi-object recognition of video image of traction substation.Most of the existing traction substation monitoring image recognition is for a specific target feature extraction,unable to complete multi-target recognition.Moreover,the target recognition task with good robustness,fast speed and high accuracy cannot be completed simultaneously.This paper is aimed at the study of this task,which can not only achieve the multi-target recognition task with good robustness,fast speed and high accuracy,but also solve the problem that the color of all kinds of indicator lights in a very dark state are all white,so that they cannot be correctly recognized.Based on the theory of deep learning,this paper makes use of the mobility of deep neural network and applies the idea of transfer learning to overcome the dependence of deep learning on data and hardware computing power.The multi-object recognition model of video image of traction substation with good performance can be used in engineering practice under the condition of limited computing power and data quantity.Based on the theoretical study of transfer learning,this paper explored the SSD(Single Shot Mutibox Detector)and YOLOv2(You Only Look Once v2)model,and combined the Python language and TensorFlow platform to realize the multi-target identification method with high accuracy,fast speed and good robustness.This model transfer method can automatically identify the meters,on-off indicator light of High Voltage Switchgear and on-off state of highvoltage switch in video image of traction substation.Moreover,the recognition function can be completed accurately and quickly on the changes of pixel image,external environment,light and different camera angles.It is of great engineering significance for the unattended development of substations.