Tunnel Lining Defect Identification Method Based on Small Sample Learning

Aiming at the problem of insufficient number samples due to difficulty data acquisition in identification tunnel lining defects, a generative adversarial network was introduced expand data, and improved for mode collapse traditional that generated image features were not obvious. On basis WGAN-GP network, deep convolutional is selected as its backbone effectiveness feature extraction by Lv et al. (2022) used improve quality images network. In addition, residual module into discriminator upsampling generator which further solves gradient disappearance two networks during update iteration process through idea cross-connection, while better retaining underlying features, effectively low Compared with original improved, losses converge faster. At same time, recognition accuracy YOLOv5 4.4% overfitting phenomenon alleviated, proves method under limited training set.