Research On Image Detection Method Of Rail Surface Defects Under Unbalanced Small Samples

Railroad system is an important part of Chinese transportation system,its operation environment is complex and changeable,often affected by rain,snow and dirt,and the rail surface load also changes in real time with the different weight of locomotives,thus making the types of defects on the rail surface very complicated.Due to the supervised deep learning methods in the application of rail surface defects detection there is a problem of insufficient defect samples and imbalance between positive and negative samples.A three-stage rail surface defect detection algorithm RSDD-Net(rail surface defect detection network)is proposed based on the frontier deep learning method,including RC-Net(rail surface cropping network)for rail surface localization and extraction,DR-VAE(defect removal variational autoencoder)for rail surface defect segmentation and MSDC(defect mask classifier)for rail surface defect classification.Firstly,based on the feature complexity of the original rail inspection image,a simple and effective RC-Net regression model is proposed to achieve accurate extraction of rail surface regions;secondly,based on the depth generation model Soft-Intro-VAE(soft introspective variational autoencoder),a lightweight semantic segmentation architecture DR-VAE is proposed,which only requires normal samples for selfsupervised training to realize the semantic segmentation of railroad surface defects.Then,based on the main architecture of twin neural network,combined with the idea of selfsupervised learning,we propose a more applicable MSDC for rail surface defect classification for three specific categories of rail surface defects.First,by analyzing the complexity of the original rail inspection image,the target detection network is reduced and a regression model RC-Net,which outputs only the transverse coordinates of the rail vertices,is proposed to locate and extract the rail surface area from the rail inspection image for subsequent defect segmentation and defectSecond,the defect segmentation model DR-VAE masks random pseudo-defects to random locations in the normal track surface image to generate self-supervised signals by DRM(defect random mask)during the training phase.Meanwhile,the decoder in DR-VAE also acts as a discriminator to implement introspective adversarial training.To optimize the loss function for complex and variable track surface backgrounds,DR-VAE uses SSIM(Structured Similarity Metric)as a pixel loss to reconstruct the defect-removed track surface background images.In the inference stage,DR-VAE reduces the error in the reconstruction by introducing a distributed capacity attenuation factor to improve the restoration of the rail surface background,and finally uses the residual map of the original and reconstructed images to achieve semantic segmentation of rail surface defects.Then,the MSDC model for rail surface defect classification is proposed.The model architecture is divided into three main parts,including a feature extraction framework,a parameter fine-tuning framework,and a classification inference framework.Among them,the feature extraction framework enables the model to learn more general feature vectors through the process of encoding and decoding images by autoencoder.The parameter fine-tuning framework fine-tunes the weight parameters of the feature extraction network on the rail defects dataset to make the model more suitable for extracting abstract features that are beneficial for rail defects classification.The inference framework identifies defect types by comparing the cosine similarity of the support set with the normalized feature vectors of the defect images to be classified.Finally,a three-stage rail surface defect detection model RSDD-Net is obtained by cascading RC-Net,DR-VAE and MSDC.the influence of hyperparameters in each sub-model on the model performance is investigated by P-R curves and the optimal hyperparameters of the sub-models are determined.The main performance indexes of each sub-model are investigated by comparison experiments.The comprehensive experiments show that the defect detection results of RSDD-Net can meet the practical requirements and have certain application value.