Highway tunnel crack recognition algorithm based on improved DeepLabv3+ and attention mechanism

Intelligent algorithms are promoting tunnel crack detection to automation and high precision, but the existing technology has some shortcomings in complex background and fine crack recognition. The study is based on DeepLabv3+ and improves the parallel structure of Atrous Spatial Pyramid Pooling (ASPP), proposing an Improved Atrous Spatial Pyramid Pooling (IASPP) module to further strengthen feature interaction; Multi-dimensional Hybrid Dynamic Attention (MHDA) mechanism is designed to enhance linear feature capture through multi-branch fusion and parallel attention. The DeepLabv3 +/MHDA algorithm is further constructed by combining the two improvements. Experimental results show that the proposed algorithm achieves the Accuracy (ACC) and F1 of 87. 32% and 0. 81 on the Crack Forest Dataset (CFD), and 88. 01% and 0. 87 on the DeepCrack dataset, respectively. The Frames Per Second (FPS) is 0.55 and 0.59, which is better than U-Net, Pyramid Scene Parsing Network (PSPNet) and the original DeepLabv3+. In the application test, the area of missed detection and false detection is significantly reduced, and the ability to deal with fine cracks and complex background is outstanding. The research provides an efficient scheme for automatic detection of tunnel cracks, and promotes the practical process of intelligent algorithm in engineering safety monitoring.