Abstract: The leakage defect is one of the most common surface defects during the operation of shield tunnels,which has adverse effect on the tunnel structure safety and surrounding ground stability.According to the deep learning-based image defect detection methods,this study constructed a mixed dataset,including the images collected by both detection device and manual inspection.Taking the average precision (AP) as evaluation metric,the trained Mask R-CNN achieves better recognition performance (0.447) than the original (0.386) and expanded (0.403) datasets.Considering the complex characteristics of leakage defect and the large difference between different defects,the model parameters dynamically generated by conditional convolution were further adopted to replace the static model parameters in Mask R-CNN,and improve the model segmentation speed and accuracy.With frames per second (FPS) as the evaluation metric,the model segmentation speed is increased from 7 to 10FPS,and the segmentation results are closer to the real defects contour,which is beneficial to realize quantitative analysis of the leakage defects severity.