Abstract Accurate identification of surface defects in tunnel linings is essential for assessing security risks and devising targeted maintenance strategies. Despite the extensive use of deep learning‐based technology for this purpose, achieving satisfactory identification accuracy from insufficient and imbalanced samples remains challenging. This paper proposed a novel approach that combines generative adversarial network (GAN)‐assisted data augmentation with a convolutional neural network (CNN) for image‐level identification of tunnel defects in scenarios with limited data. First, a deep convolutional GAN was built to automatically generate synthetic defect images, which were then incorporated into the initial dataset to create a more sufficient and balanced dataset. Next, an innovative CNN‐based model called parallel multiscale attention convolutional neural network (PMACNN) was developed to enable efficient defect image classification. This model integrates multi‐scale parallelism and attention mechanisms, which can effectively improve the feature extraction capability and key feature attention. The PMACNN was trained, validated and tested on both the initial and augmented datasets. Its performance was compared with seven state‐of‐the‐art deep learning models, and the results show that PMACNN achieves superior computational efficiency and classification accuracy over the baselines. Meanwhile, the ablation studies clearly demonstrate the contributions of GAN augmentation, multi‐scale parallelism and attention mechanisms to the model's performance, increasing accuracy by 8.89%, 2.31%, and 1.18%, respectively. Finally, the gradient‐weighted class activation mapping was applied to interpret the model, and its visualized heatmaps proved to facilitate the exact defect localization without pixel‐level image labeling.
Yan et al. (Fri,) studied this question.