Research On Recognition Technology Of Freight Train External Inspection Images Based On Deep Learning

The Cargo inspection operation plays a very important role in railway freight transportation and is an important way to ensure the safe operation of railway freight.At present,the main cargo inspection operation method still relies on manual labor.Even in some large marshalling stations that have installed high-definition video surveillance systems,the cargo inspection operation is only transferred from the outdoor to the indoor,and essentially relies on manual labor.Therefore,current cargo inspection operations still have problems such as low cargo inspection efficiency,high rate of missed inspections,unstable working conditions,and difficulty in achieving intelligent management.The target detection technology based on deep learning has been widely used in many industrial inspections and has achieved excellent detection results.In this paper,the recognition of abnormal targets in the external detection image of railway freight trains is studied by target detection technology.The main research contents include:(1)Based on the typical truck shape images collected at the freight marshalling station,analyze the types and characteristics of cargo inspection failures.Taking common Gondola and Caravan as examples,the basic data set of cargo inspection images is constructed through methods such as data collation and analysis and manual accurate labeling,and finally the belts are obtained.There are more than 2500 accurately labeled basic data sets of VOC2007 format.(2)Analyze and discuss data enhancement technology,and adopt multiple data enhancement methods such as image flipping,contrast enhancement,color enhancement,and noise addition for different goods inspection pictures to simulate the difference of the goods inspection site environment,expand and enrich the goods inspection image data set,and improve In order to improve the generalization ability and robustness of the training model,the experimental verification data enhancement can effectively improve the detection accuracy of the model.In addition,this article also compares the effects of Mosaic and CutMix data enhancement methods on cargo inspection data sets,and experiments verify that Mosaic data enhancement methods are better than CutMix data enhancement methods.(3)Based on the target detection algorithm of convolutional neural network,considering the high-efficiency and real-time requirements of data processing on site,the detection effects of the four target detection models are compared through experiments.The results show that the YOLOv4 target detection model has more efficient and accurate recognition capabilities.YOLOv4 is used as the basic algorithm for intelligent inspection of goods inspection images.(4)Aiming at the problem of the size of a priori box,the K-means++clustering algorithm is introduced,considering the multi-scale features of YOLOv4,the size of the a priori box is further optimized through the parameter adjustment algorithm,and the gap between the parameters of the a priori box is increased.The matching degree between the a priori box and the different feature layers is improved,and the performance of the model is further improved.(5)For the single-stage target detection algorithm of YOLOv4,which often has the problem of class imbalance,a multi-scale focal loss strategy is introduced to improve the loss function of YOLOv4.Experiments verify that the improved YOLOv4-MF model can achieve higher AP and Fβ on the data set in this paper.