AquaSegNet: A real-time dynamic network for high-precision water leakage segmentation in shield tunnel lining

Long-term water leakage in shield tunnel linings degrades lining materials and compromises structural integrity, thereby severely threatening operational safety. This paper proposes AquaSegNet, a real-time framework derived from You Only Look Once version 11 segmentation (YOLOv11-seg), designed to tackle the geometric constraints of irregular leakage morphologies and the perceptual challenges of complex environmental interferences. Specifically, we formulate the Multi-Scale Dynamic Inception Mixer (MSDIM) to replace standard bottleneck units across the backbone and neck, enabling the adaptive capture of diverse leakage patterns. Simultaneously, the Spectral Enhanced Feed-Forward Network (SEFFN) is integrated to suppress environmental noise in the frequency domain while fusing multi-scale local details with global context. Furthermore, the prediction architecture is reconstructed using the Decoupled Fusion Shared Head (DFSH), which efficiently optimizes computational interaction through task decoupling. Experimental results on the self-constructed AquaSeg dataset and the public Water leakage dataset demonstrate that AquaSegNet achieves high-precision and an optimal computational trade-off with only 2.33 M parameters and 8.6 Giga Floating Point Operations (GFLOPs), thereby providing an efficient solution for real-time edge deployment.