An intelligent segmentation and quantification method for apparent cracks in tunnels based on improved Faster R-CNN and pixel threshold algorithms

This work presents a deep learning framework for the automatic detection and quantitative evaluation of tunnel cracks. Initially, a systematic data pre-processing method is proposed. By jointly leveraging SRGAN, Histogram Equalization, Contrast Limited Adaptive Histogram Equalization, the raw data is normalised and augmented prior to model training. A semantic segmentation method for tunnel lining cracks is developed, which is based on the object-detection network, the OTSU algorithm, and morphological algorithms. In the backbone network of the object detection, Swin Transformer and Sand Glass Block are integrated to concurrently account for both global perception and local feature representation capabilities. Furthermore, a Parallel Three-Channel Feature Extraction module (PTFE) and a Hierarchical Semantic Broadcast Feature Extraction module (HSB) are combined to enhance the responsiveness of the feature maps to critical regions. Based on the image data collected from actual tunnel engineering projects, the performance of the proposed method is evaluated using multiple evaluation indicators. The results demonstrate that the proposed method exhibits superior performance compared to Faster R-CNN, RetinaNet, and YOLO v5. Specifically, investigations are carried out to explore the impact of the PTFE and HSB module, data preprocessing techniques, and transfer learning on the proposed method.