Railroad track safety is critical to transportation operations, and track defects may lead to serious accidents. This paper proposes a railroad track defect detection method based on image feature extraction algorithm and BP neural network. Firstly, X-ray imaging is used to obtain the rail defect image, and then the image quality is enhanced by wavelet transform noise reduction processing and contrast and entropy. Then, the improved Otsu threshold method is used to segment the image and extract the geometric features of the defective image; finally, a BP neural network with 5-15-6 structure is constructed to classify and recognize the defects. The experimental results show that the running time of the proposed method in the threshold segmentation stage is only 0.1526 seconds, which is reduced by 62.3% compared with the maximum Shannon entropy multi-threshold segmentation method. For the six defects of wave abrasion, optical band anomaly, serration, abrasion, corrosion and pitting, five geometric features of frequency, amplitude, area, long axis and short axis are selected for the BP neural network training, and the network reaches an error of 0.01 after 112 trainings, which realizes the accurate classification of defects. The method improves the automation degree and accuracy of track defect detection through the combination of image feature extraction and machine learning, which provides an important guarantee for the safe operation of railroads.
