Abstract
With the rapid development of urban metros, the detection of shield tunnel leakages has become an important research topic. Progressive technological innovations such as deep learning-based methods provide an effective way to detect tunnel leakages accurately and automatically. However, due to the complex shapes and sizes of leakages, it is challenging for existing algorithms to detect such defects. To address these issues, this paper proposes a novel deep learning-based model named segmenting objects by locations network v2 for tunnel leakages (SOLOv2-TL), which is enhanced by ResNeXt-50, deformable convolution, and path augmentation feature pyramid network (PAFPN). In the SOLOv2-TL, ResNeXt-50 coupled with deformable convolution is the backbone for boosting feature extraction ability that would enable the model sensitivity to leakages of different shapes. The PAFPN is introduced as the neck to reduce the loss of leakage information and more accurately assign leakages of different sizes to their corresponding feature levels. The superior performances of ResNeXt-50 with deformable convolution and PAFPN were validated by ablation tests. Moreover, the segmentation results obtained by SOLOv2-TL were compared with those by the mask region-based convolutional neural network (Mask R-CNN), Cascade Mask R-CNN, and SOLO which demonstrated that the mAP, mAP50, and mAP75 of SOLOv2-TL are higher than those of the other methods, where mAP indicates the mean mask average precision (AP) at intersection over union (IoU) = 0.50:0.05:0.95, mAP50 refers to the mean mask AP with an IoU threshold of 0.50, and mAP75 denotes the mean mask AP with an IoU threshold of 0.75. Finally, a leakage area quantification method is presented.