Steel rail surface defect image expansion method based on attention mechanism and DCGAN

Abstract

The invention relates to the technical field of steel rail surface defect detection, in particular to a steel rail surface defect image expansion method based on an attention mechanism and DCGAN, comprising the following steps: making a rail surface defect data set; building a self-attention module; building a channel attention module; building a DCGAN network and adding self attention and channel attention; training the model and obtaining a generation result; the method has the beneficial effects that the depth of the DCGAN is properly increased, a self-attention mechanism is added to capture the long-distance size dependence in the image, and the defect that convolution can only process local neighborhood information is overcome; and a channel attention mechanism is added to recalibrate features of a convolution channel interdependence relationship, and more important channel information is paid attention to. The rail surface defect samples generated by the method are good, many new defects are generated on the basis of the original data set, the diversity of the defects is increased, the method can be used for large-scale expansion of the rail surface defect samples, and application of supervised deep learning to steel rail detection is improved.

Description

technical field [0001] The invention relates to the technical field of rail surface defect detection, in particular to a rail surface defect image expansion method based on an attention mechanism and DCGAN. Background technique [0002] During train operation, the wheel hub and the rail have long-term contact friction and impact, and the defects of the rail mostly start from the surface and gradually increase. When the train runs on the surface of the defective track, the periodic impact will cause the coupled vibration of the entire vehicle and the rail system. Running under the environment of extreme conditions will shorten the service life of each part of the train, and in severe cases, it will cause the vehicle to be overturned, burned and cut. Therefore, it is of practical significance to detect track surface defects in time. Physical detection methods such as eddy current detection and magnetic flux leakage detection were once the main means of rail detection, but thei...

AI Technical Summary

Problems solved by technology

However, supervised learning needs to train a large number of defect samples. Due to timely maintenance, there are few defect samples in the track inspection images obtained from the railway engineering department, so that supervised deep learning such as target detection and semantic segmentation does not perform well in actual scenarios. At the same time, there are many defects. Scale and diversity bring challenges to detection tasks

Currently commonly used data augmentation methods such as rotation, cropping, and adding noise essentially only increase the number of data sets, and the diversity of defect samples generated by the generative confrontation network GAN is low, making it difficult for the algorithm to detect untrained defects.

Images