An Accurate Tunnel Crack Identification Method Integrating Local Segmentation and Global Fusion Detection

This paper proposes a novel approach for tunnel crack identification, employing local segmentation and global fusion detection. Initially, a local segmentation network is constructed using weights from an encoding layer pre-trained on numerous non-tunnel cracks, with only a limited number of tunnel crack samples. The input image is divided, and the local segmentation network performs pixel segmentation on these sub-images, with the sub-results stitched together to ensure accurate identification of all suspicious crack pixels. Subsequently, a global fusion detector is introduced, comprising two sub-models: Sub-model 1 extracts total crack targets within the stitched results, while Sub-model 2 detects possible false alarms from regular-shaped areas. The results from both sub-models are combined to effectively reduce the false alarm rate and ensure accurate segmentation results of cracks. Experimental findings on actual tunnel images demonstrate that the "segmentation before localization" method proposed in this paper achieves superior recognition accuracy and IOU ratio compared to the Unet3+, DeeplabV3+, and "localization before segmentation" Mask-RCNN algorithms. Specifically, the proposed method yields an accuracy improvement of 3.81% over the Unet3+ network, 2.71% over the DeeplabV3++ network, and 1.93% over the Mask-RCNN network. Moreover, noise interference from bolt repair areas is effectively mitigated, enhancing the method's engineering applicability.