Font Size: a A A

Study On The Detection Of Rail Tread Block Defect Based On Improved Faster R-CNN

Posted on:2021-05-21Degree:MasterType:Thesis
Country:ChinaCandidate:C Y LuFull Text:PDF
GTID:2392330611479835Subject:Information and Communication Engineering
Abstract/Summary:PDF Full Text Request
Due to the impact of the rail wheel's strong friction,extrusion and impact,the rail tread,as the most frequent contact area between the rail and the wheel,is prone to defect to train safety and block defect is the most common of these defects.Therefore,in order to ensure the running safety of trains in the railway operation,this paper studied the detection of rail tread block damage.The main research contents are as follows:1)Considering that the detection technology based on machine vision has the advantages of fast speed and intuitive results,this paper analyzes the rail damage detection research based on traditional machine vision and deep learning,and further studies the target detection research based on deep learning.Considering that the target detection based on region nomination has the advantages of high detection accuracy,the research of rail tread damage detection based on region nomination is proposed.2)Considering the characteristics of small difference and large scale variation in the block damage of rail tread,a detection study of the block damage of rail tread based on FPN(Feature Pyramid Network)and Faster R-CNN was proposed,and the effectiveness of the detection of the block damage of rail tread was verified by the design and training of the detection Network3)In order to solve the problem that the regression loss of Smooth L1 in Faster R-CNN is defined as the distance between the predicted border and the actual border,and is insensitive to the predicted border position,a novel novel algorithm is proposed to replace the loss of Smooth L1 with a Generalized Intersection over Union(GIo U)loss.In addition,aiming at the problem that the large number of redundant anchors generated by Region Proposal Network(RPN)in Faster R-CNN lead to the imbalance of positive and negative samples in the detection Network training,and then affect the Network detection effect,an optimization scheme of Faster R-CNN using Region Proposal by Guided Anchoring(GA-RPN)to replace RPN was proposed.Finally,through experiments,it is verified that the improved network can improve the detection accuracy of rail tread block damage and realize the accurate detection of damage.
Keywords/Search Tags:Rail tread, block defect, Object detection, Deep learning, Faster R-CNN
PDF Full Text Request
Related items