
Graphical abstract

Fig. 1. Framework of the proposed two-step deep learning-enabled metro tunnel lining defect recognition.

Fig. 2. Overall architecture of TLCNet.

Fig. 3. Structures of (a) MBConv and (b) SE modules.

Fig. 4. Structures of (a) visual Transformer module and (b) feed forward network.

Fig. 5. Overall architecture of TDDNet.

Fig. 6. Details of the modules in the TDDNet model.

Fig. 7.
An illustration of the predicted box, real bounding box, and minimum enclosing box.

Fig. 8. Structures of (a) SimAM block and (b) ACmix block.

Fig. 9. A schematic diagram of the metro tunnel inspection drone (MTID).

Fig. 10. Typical examples of (a) defect-free and (b) defective images in the classification dataset.

Fig. 11. Schematic diagram of the 5-fold cross-validation experiment.

Fig. 12. Dataset for object detection experiments.

Fig. 13. Data augmentation methods.

Fig. 14. Loss curves of the TLCNet in the 5-fold cross-validation experiment.

Fig. 15. Classification results of several representative testing images.

Fig. 16. Confusion matrices generated during cross-validation of (a) TLCNet, (b) ViT, and (c) DenseNet.

Fig. 17
. Loss curves of the TDDNet on the training and validation datasets.

Fig. 18. Identification results of different models for large-scale water leakages (a) (b), medium-scale spalling defects (c) (d), small-scale cracks (e) (f), and mixed defects (g) (h).

Fig. 19. Predicted results and heatmaps of different models for testing images.

Fig. 20. AP calculation results of seven models, including (a) TDDNet, (b) Faster R-CNN, (c) YOLOv4, (d) YOLOv5, (e) YOLOX, (f) YOLOv7-x, and (g) SSD.


Fig. 21. Predicted results of seven models for (a) simple and (b) complex samples.


Fig. 22. Metro tunnel lining defect intelligent recognition platform: (a) home page, (b) interface for automatic classification of lining images, and (c) interface for detection of multi-defect.


Fig. 23. Practical engineering applications based on defect intelligent detection platform: (a) metro tunnel lining image classification application example 1, (b) metro tunnel lining image classification application example 2, (c) metro tunnel lining multi-defect detection application example 1, and (d) metro tunnel lining multi-defect detection application example 2.