Semantic Segmentation of Shield Tunnel Leakage with Combining SSD and FCN

Abstract

With the rapid development of maintenance of the urban metro tunnel, the structural defects of the metro shield tunnel, especially the water leakage, need to be recognized quickly and accurately. Mask R-CNN is one of the state-of-the-art instance segmentation methods, which has achieved the automatic segmentation of shield tunnel leakage. Although the error rate of the Mask R-CNN algorithm is very low due to a series of complex network structures such as feature pyramid network (FPN) and region proposal network (RPN), the inference cost is 3.24 s per image. Because the structural inspection usually takes only 2–3 h, quick processing of defect images seems necessary. Inspired by a real-time detection method called Single Shot MultiBox Detector (SSD) and the generation of Mask R-CNN, this study constructed a novel convolutional network for fast detection and segmentation of the water leakage. Taking into account the unique appearance and features of water leakage area, it was divided into five groups of different backgrounds to evaluate the interference caused by the complex background and its various shapes. Finally, 278 images were used to test the network, and the average IOU was found as 77.25%, which was close to that of Mask R-CNN. Additionally, the average segmentation time was calculated as 0.09 s per image, far less than Mask R-CNN, which meets the actual requirement of engineering.