Abstract: A novel detection method based on deep neural networks is proposed in this article to handle the tough problem of standardizing and quantifying in corrosion detection, and detection as well as quantitative analysis is achieved through semantic segmentation on corrosion images. The “encoder-decoder” architecture is selected in designing deep neural networks, which divides the process of computation into encoding part (downsampling) and decoding part (upsampling), and a segmentation mask with the same resolution as the input image is obtained after computation which indicates whether a pixel belongs to corrosion. Sutong bridge corrosion dataset including 440 corrosion images with resolution of 709×1067 is used to train the segmentation network, and raw images are augmented to 6156 manual labeled images with resolution of 256×256. The training process took about 7 h containing 50 epochs with binary accuracy of 92.55% in training set and 90.56% in validation set. Besides, segmentation network is also applied on raw images and the detection results show that majority of corrosion can be recognized. Corrosion area, corrosion rate and total corrosion rate are defined to quantitatively analyze the corrosion regions, and these indices can be directly calculated with segmentation mask which provide data support for daily maintenance for steel structures.

Key words: bridge engineering, corrosion detection, deep neural networks, computer vision, semantic segmentation