Research on Enhancement Method of Track Defect Sample Based on Deep Convolution Generative Adversarial Network

The scarcity of rail surface defect samples is the main factor that affects the direct application of deep learning methods to rail surface defect recognition. This paper proposes a method for enhancing orbital defect samples based on Deep Convolution Generating Adversarial Network (DCGAN). First, analyze the reasons for the disappearance of the gradient of the DCGAN generator network, and demonstrate from the mathematical deduction that DCGAN does not completely solve the problem of the disappearance of the gradient; then, by adding noise to the real image and the generated image at the same time to guide the learning process of the sample distribution, JS is realized (Jensen-Shannon) Divergence optimization; Finally, experimental demonstration. The experimental results show that this method effectively realizes the feature restoration of the defect samples in addition to improving the stability of the training process, provides sufficient and effective sample data for the application of deep learning in the detection of track defects, and has certain application value.