A lightweight fully convolutional neural network for efficient and accurate tunnel crack detection

Crack damage is a significant issue in tunnel environments, requiring efficient and reliable detection to ensure structural integrity. Many existing solutions are computationally expensive, making them unsuitable for the tight "sky window" inspection periods common in tunnel maintenance. This study introduces FCN CrackDetect to provide a practical and high-speed solution for automated tunnel crack detection that balances high accuracy with low computational requirements.

The goal of this study is to develop a lightweight model that provides high accuracy with low computational cost, while effectively addressing the "domain gap" to ensure reliable crack detection across diverse and unseen tunnel environments. This FCN CrackDetect is developed to achieve strong generalization through strategic architectural design rather than explicit adaptation modules.

We designed a lightweight U-Net-based architecture featuring an improved EfficientNetV2S, optimized for feature extraction. The model's novelty lies in the integration of DropBlock regularization and Efficient Channel Attention (ECA) to promote the learning of domain-invariant features, a concept known as "implicit domain generalization". This forces the model to learn more resilient representations, enhancing its performance across source domains without requiring data from an unseen tunnel environment with cracks during training.

Evaluated on a dataset of 6000 tunnel crack images, FCN CrackDetect demonstrates superior accuracy and a significantly faster runtime (29 ms) compared to baseline models validated through both CPU and GPU evaluations. The ablation study confirms the synergistic effects of integrated components, such as EfficientNetV2S, residual connections, and CBAM attention, which enhance crack detection capabilities and improve feature learning.

FCN CrackDetect provides a feasible and fast method for automated tunnel crack detection, serving as a robust alternative to complex adaptation frameworks.