Precise tunnel lining crack detection using multimodal fusion transformer and metaheuristic-optimized ResNet-50

Automated crack detection in tunnel linings is critical for ensuring infrastructure safety and facilitating efficient long-term maintenance. Manual inspection methods are typically slow, labour-intensive, and inconsistent, highlighting the necessity for advanced automated solutions. This research presents a Mathematical Multimodal Fusion Transformer-Based Optimization (MMFTO) framework that integrates multiple tunnel-lining image modalities to enable robust crack identification. The framework utilizes an optimized multimodal fusion transformer encoder to extract and combine complementary features that are often missed by unimodal systems. Hyper-parameter tuning is conducted using an Improved Piranha Foraging Optimization Algorithm (IPFOA) algorithm, followed by feature refinement with a Modified Sunflower Optimization Algorithm (MSFOA). Final classification is performed by a DeepResNet-50 model. The proposed MMFTO–MSFO–DeepResNet50 system is evaluated on a newly curated multimodal tunnel-lining dataset and compared with existing deep-learning and unimodal approaches. The model achieves 95 % accuracy, 92 % precision, 92 % F1-score, 91 % recall, a mean average precision (mAP) of 94.7 %, and a mean intersection over union (mIoU) of 93.6 %. Further evaluation on Crack500, Tunnel200, and DeepCrack datasets demonstrates strong robustness and generalizability. These results indicate that the framework offers significant potential for real-world tunnel monitoring and intelligent infrastructure management.