Electronic Science and Technology ›› 2020, Vol. 33 ›› Issue (9): 63-68.doi: 10.16180/j.cnki.issn1007-7820.2020.09.011

Previous Articles     Next Articles

Research on Rail Surface Defect Recognition Based on Faster R-CNN

SU Ye1,LI Jun1,YANG Haima1,LIU Jin2,JIANG Shenghua3   

  1. 1. School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China
    2. School of Electronic and Electrical Engineering,Shanghai University of Engineering Science,Shanghai 201620,China
    3. Shanghai Ruiniu Machinery Crop,Shanghai 201314,China
  • Received:2019-06-17 Online:2020-09-15 Published:2020-09-12
  • Supported by:
    National Natural Science Foundation of China(61701296);National Natural Science Foundation of China(U1831133);Shanghai Natural Science Foundation(17ZR1443500);Shanghai Aerospace Science and Technology Innovation Fund(SAST2017-062);Baoshan District Science and Technology Innovation Special Fund(17-C-21)


External factors usually have effects on the instrument used to detect rail surface defect, resulting in poor accuracy and efficiency of instrument. For this problem, a method for detecting rail surface defects based on the Faster R-CNN network is investigated. The method reverses the preprocessed image and realizes the projection of rail image with Radon transform. In the projection curve, the rail surface area is extracted by using the characteristics that the rail length is fixed and the gray value is less than the average value of the image. Then, the candidate region is extracted through the regional proposal network and compared with the regional recommendation of Fast R-CNN network for the detection of surface defects of rail by the Faster R-CNN networks. According to the test data, the accuracy of crack, scar, abrasion and scratch is 92.17%, 91.85%, 93.45% and 93.27% respectively, which verifies the efficiency and accuracy of the proposed method in identify the surface defects of rail.

Key words: rail surface defect, preprocessing, radon transform, grey value, regional proposal network, Faster R-CNN networks

CLC Number: 

  • TP278