AUTHOR=Zhendong He , Xiangyang Gao , Zhiyuan Liu , Xiaoyu An , Anping Zheng 

TITLE=Rail surface defect data enhancement method based on improved ACGAN

JOURNAL=Frontiers in Neurorobotics

VOLUME=Volume 18 - 2024

YEAR=2024

URL=https://www.frontiersin.org/journals/neurorobotics/articles/10.3389/fnbot.2024.1397369

DOI=10.3389/fnbot.2024.1397369

ISSN=1662-5218

ABSTRACT=Rail surface defects are one of the common problems in railway operations and maintenance, posing a serious risk of accidents if not addressed promptly. However, the scarcity of rail surface defect samples and limited experimental data pose difficulties when employing deep learning methods for defect detection. These challenges often result in issues such as vanishing gradients and abnormality within the neural network, leading to instability and overfitting of training results. To address this problem, an improved ACGAN data augmentation method is proposed. Firstly, residual blocks are incorporated into the network structure to mitigate vanishing gradient problems. Additionally, a spectral norm regularization constrained discriminator is employed to enhance network stability, improve model generalization performance, and generate higher quality images. Furthermore, the generator's deconvolution layer is replaced with upsampling and convolution operations, while an undersampling layer is added to the discriminator for computational efficiency purposes. To tackle concerns regarding gradient abnormality, the gradient penalty mechanism based on maximum and minimum regret values is utilized. The proposed model has been validated experimentally by comparing it with ACGAN on the rail surface defect dataset. It has been found that the model generates clearer images with higher classification accuracy. Moreover, the proposed model reduced the FID value by 17.6% compared to ACGAN on the same dataset, reinforcing its ability to generate high-quality samples closely resembling real-world instances. Finally, experiments were conducted on the public dataset MNIST handwritten dataset, and good results were achieved, verifying the generalization ability of the network. This method has high practical value and can be applied to various image fields.