Tunnel lining water leakage image Segmentation based on improved BlendMask

Abstract

Water leakage of tunnel is a common disease in all kinds of existing tunnels. Automatic, timely, and accurate detection of water leakages is of great significance to the safe operation and maintenance for all kinds of tunnels. Due to the complicated background of tunnel surfaces, it is difficult for traditional tunnel defect detection algorithms to detect such defects quickly and accurately. To address these problems, this article presents an improved BlendMask image segmentation model with top-down and bottom-up architecture which accurately extract tunnel’s water leakage area from tunnel lining images. The proposed method is validated by an experimental study, and the results are compared with those obtained by the tunnel defect segmentation methods including Mask R-CNN and DeepCrack and other deep learning methods including CondInst, SoloV2, fully convolutional network and UNet. The Recall, Precision, F1-score, Dice and mean intersection over union (mIoU) for the proposed method are superior than those by the other methods above with respect to on test images. The subjective analysis on predicted results by different methods also shows the presented method can perform well.