Research On Rail Surface Defect Detection Technology Based On Few-Shot Learning

Rail is an important safeguard of transportation.The research of rail surface defect detection is of great significance.At present,rail inspection is mainly conducted by manual visual inspection.This method relies too much on subjective judgment and has low efficiency,as well as security risks.Research on defect detection based on deep learning develops rapidly and has been applied to many practical environments,but usually requires a large number of samples for training.Rail surface defect samples are not easy to collect and the number is small.If they are directly used in network training,network overfitting will be caused,which cannot meet the needs.Aiming at the problems of few samples and low accuracy in detecting fine defects,this paper proposes a rail surface defect detection method based on few shot learning.The specific methods are as follows:(1)A data enhancement method for rail defect samples based on improved auxiliary classifier generative adversarial network(ACGAN)is proposed.From the perspective of sample expansion to solve the problem of fewer samples,and the ACGAN is improved by combining the characteristics of rail defect samples.Residual block and spectral norm regularization are added to the network structure to solve the problem of gradient disappearance and gradient anomaly.The deconvolution layer in the generator is replaced by up-sampling plus convolution,and the down-sampling layer is added to the discriminator to reduce the computation.The problem of determining sample authenticity in GAN is regarded as PU learning method,which makes the network pay attention to the quality of the generated image.The gradient penalty mechanism based on the maximum and minimum regret method is added to the objective function to limit the gradient change amplitude.The model presented in this paper is compared with ACGAN on rail surface defect data set and MNIST handwritten digit data set.Experiments show that the method proposed in this paper is better than ACGAN in image quality and can be applied to a variety of environments.(2)An improved feature pyramid network(FPN)for rail surface defect localization method is proposed.Considering that the deep learning network is not effective in detecting fine defects on rail surface,this paper improves the FPN.Deformable convolution is added to FPN to cope with the geometrical deformation of rail surface defects.Adding attention mechanism to increase the weight of defect features;ROI Align was added to optimize defect feature extraction.The improved FPN was pre-trained on MS COCO data set,and then the parameters were transferred to the rail detection model using transfer learning method.The method presented in this paper can effectively solve the difficult problem of detecting small rail defects and realize the accurate positioning of defect detection.(3)A metric learning-based method for classifying rail surface defects is proposed.Metric learning is used to solve the few-shot problem.Based on the representation-based measurement learning method,a multi-modal network and a feature embedding module are designed to classify defects by calculating the distance between category representation and feature vector.Comparison and ablation experiments are carried out on mini Image Net data set and rail defect data set,and the results were analyzed to verify the feasibility of the proposed method.