Abstract Accurately segmenting and analyzing the muck generated during TBM excavation can reflect the geologi? cal conditions of tunnels and the operation of equipment, which is of great significance for construction risk warning and improving construction efficiency. To address issues such as incomplete detection of large muck blocks, missed detection of small muck blocks, and unclear edge segmentation during the segmentation process, a muck segmentation network based on global perception and edge refinement is proposed. A global perception module is designed to utilize deep strip convolutional attention networks of different sizes to expand the network's receptive field and enhance the integrity of muck block segmentation. An edge refinement module is introduced to aggregate spatial attention and channel attention, and a channel shuffle method is used to promote information exchange between different channels, thereby improving the network's perception ability for image details and the accuracy of muck edge segmentation. Through testing on a self-made dataset, compared with other classic algorithms, the proposed network achieves improvements in objective evaluation metrics, with recall, precision, intersection over union, and F1 score reaching 98.37%, 91.48%, 90.11%, and 94.80%, respectively. Additionally, the segmentation effect images are more complete, and the edges are clearer.
Abstract：
Accurately segmenting and analyzing the muck generated during TBM excavation can reflect the geologi? cal conditions of tunnels and the operation of equipment, which is of great significance for construction risk warning and improving construction efficiency. To address issues such as incomplete detection of large muck blocks, missed detection of small muck blocks, and unclear edge segmentation during the segmentation process, a muck segmentation network based on global perception and edge refinement is proposed. A global perception module is designed to utilize deep strip convolutional attention networks of different sizes to expand the network's receptive field and enhance the integrity of muck block segmentation. An edge refinement module is introduced to aggregate spatial attention and channel attention, and a channel shuffle method is used to promote information exchange between different channels, thereby improving the network's perception ability for image details and the accuracy of muck edge segmentation. Through testing on a self-made dataset, compared with other classic algorithms, the proposed network achieves improvements in objective evaluation metrics, with recall, precision, intersection over union, and F1 score reaching 98.37%, 91.48%, 90.11%, and 94.80%, respectively. Additionally, the segmentation effect images are more complete, and the edges are clearer.
ZHANG Yan HUO Tao ZHANG Zhongwei MA Chunming
.TBM Muck Segmentation Method Based on Global Perception and Edge Refinement[J] MODERN TUNNELLING TECHNOLOGY, 2024,V61(3): 141-147
2024, Vol. 61 