Research On Multiclass And Multiscale Rail Surface Defect Detection Algorithm Based On Improved Faster R-CNN

As an efficient transportation mode,railway transportation brings great convenience to people’s life while driving economic development.Relevant statistics show that in the process of railway operation,rail surface defect deterioration caused by train accidents accounted for a relatively high proportion.So the accurate detection of rail surface defect has very important practical application value and research significance.The rail surface defect detection method can be divided into physical detection method and machine vision based method.Traditional physical detection methods are difficult to operate and require a lot of human intervention,so the detection results are subjective to a certain extent.Comparatively speaking,the detection method based on machine vision is more intelligent which has the advantages of fast detection speed and intuitive results,but it is easy to be affected by category diversity and scale variability,resulting in low detection accuracy.In this paper,a multiclass and multiscale rail surface defect detection algorithm based on improved Faster R-CNN is proposed by many aspects.The specific contents are as follows:1)Under the influence of illumination and weather conditions,there are many noise signals in the image of rail surface defect,which is not conducive to the effective extraction of defect features.To solve this problem,while using Gabor filtering algorithm for image denoising,the defect image in RGB color space is mapped to HSV color space to enhance the features of the defect image,which effectively reduces noise interference and provides a favorable condition for improving the detection accuracy of defect.2)The types of rail surface defect are varied and the number of different type defect is uneven,so it is difficult to accurately detect each type of defect.To solve this problem,firstly,the data of scarce categories are enlarged based on the generative adversarial network,which balanced the defect samples of different categories.Then,dilated convolutions with different bias are constructed for different defect types,which improve efficiency of feature extraction greatly.Finally,a semi-supervised loss function and the KL divergence are constructed,which increases the model generalization performance and the detection accuracy of multi-class defect.3)The scale of rail surface defect is varied and distribution position in the image is disorderly,so it is difficult to detect the location of each defect accurately.To solve this problem,a feature extraction module based on spatial attention mechanism is designed to improve the ability of feature extraction of small scale defects from the perspective of spatial variation.Then,the FPN detection network is chosen as the main structure,and the multi-level branch detector is constructed according to different feature extraction levels.Besides the corresponding multi-scale loss function is constructed,so as to improve the detection accuracy of multi-scale defect.The experimental results indicate that the improved defect detection algorithm can effectively detect crack defect,spalling defect and abrasion defect.Besides,the detection accuracy for different scale defect is high,which can provide technical support for the safety of railway transportation.