Towards automated 3D evaluation of water leakage on a tunnel face via improved GAN and self-attention DL model

Grasping the segmentation and three-dimensional (3D) positioning information of the water leakage area on a rock tunnel face is of great significance for determining the necessary construction arrangements to ensure the safety of tunnel excavation. This paper presents a novel method for automated 3D evaluation of a tunnel leakage area based on an improved Generative adversarial network (GAN) and Swin Transformer model. First, this paper solves the shortcomings of insufficient and unbalanced data in the original leakage image datasets obtained from mountain and submarine tunnel projects, by using new images generated by an improved lightweight GAN model that establishes the GAN-based WIIN-2 dataset. The leakage images in this dataset are then divided into five categories. Afterwards, a newly developed high-performance Swin Transformer model combines shift windows and a self-attention mechanism to produce intelligent segmentation of the leakage area. The segmentation results of the Swin Transformer model on the GAN-based WIIN-2 dataset achieves mAcc, mIoU, mF score, mPrecision and mRecall metrics of 93.1%, 91.5%, 82.83%, 85.62% and 80.3%, respectively. The segmentation results of the DL models (Swin Transformer, Deeplab V3+, Fast CNN and Unet) are subsequently compared. The Swin Transformer model performs better than the other three models in terms of the five evaluation metrics and segmentation efficiency, which indicates that the Swin Transformer model is an improvement on current methods for segmenting leakage areas on a rock tunnel face. Finally, the novel 3D leakage area location model proposed in this work is used to visualize and reconstruct the 3D coordinates of the leakage area on the rock tunnel face.