Abstract: To addresses the challenges of insufficient tunnel lining crack image samples and low recognition accuracy of existing detection methods in complex backgrounds, an intelligent crack detection approach that integrates deep convolutional generative adversarial networks (DCGAN) with the dual-task (DT)-YOLO model is proposed. In data augmentation phase, DCGAN with regularization and enhancement generates crack images rich in texture details and authentic defect characteristics, effectively expanding the training sample size to construct a dataset encompassing diverse crack morphologies and background interferences. For model design, the DT-YOLO architecture is introduced. It incorporates a global feature pyramid to enhance multiscale semantic information fusion capabilities and integrates a wavelet subsampling module to preserve high-frequency detail features. This remarkably improves the model′s performance in recognizing complex crack morphologies and defect features within noisy backgrounds. Experimental results demonstrate that the proposed method achieves an F₁ score of 89.43% and an average precision of 88.10% on the self-built tunnel lining crack dataset. Compared to mainstream YOLO variants and other typical detection models, it exhibits distinct advantages, validating its robustness and generalization capabilities in complex tunnel environments characterized by varying illumination and stain interference. The results provide a new, highly accurate, and feasible approach for the automated detection of apparent structural defects in tunnels.

Key words: highway tunnels, lining cracks, adversarial generative networks, dual-task YOLO model, intelligent detection