Surface Defect Detection of Rails Based on Convolutional Neural Network Multi-Scale-Cross FastFlow

Abstract

Aiming at the problems of difficulty in obtaining images of rail surface defects, wide range of defect types, unpredictability of defects and slow defect detection speed, we propose an algorithm for rail surface defect detection based on a convolutional neural network multi-scale-cross FastFlow model (MSC-FastFlow). The pre-training model is used to extract multi-scale features to ensure the extraction speed while doing a good job of image feature extraction. A scale normalized cross fast flow is used as a probability distribution estimation to obtain a standard Gaussian distribution of the image, and the anomaly score is achieved by the distance from the center of the distribution for detect detection as well as locating defects. The comprehensive experimental results show that the unsupervised defect detection method can automatically detect defects without any defect samples, achieving 98.2% and 97.42% detection accuracy at the image level and pixel level, respectively, with a detection speed of 0.08s per image, which is better than existing algorithms for rail surface defect detection and can be well applied in practical projects.