Automated Bridge Coating Defect Recognition Using U-net Fully Convolutional Neural Networks

As the weather in Taiwan is mostly warm and humid, steel bridges get rusted easily. For rusting is one of the most significant factors in steel bridge maintenance, together with the crucial role that steel bridges play in most countries, it is important to develop effective steel bridge rust detection methods to enhance steel bridge health and safety, and lower the lifecycle cost of steel bridges. Some image processing techniques (IPTs) have been developed in prior research to quickly and effectively detect rust defects of steel bridges. The keys to rust defect recognition are the discrimination of rust spots from the background that may contain rust-like noises and the handling of non-uniform illumination. In order to detect rust spots in a more effective and efficient fashion, this paper explored a new rust recognition method that integrates a deep-learning-based fully convolutional neural network, namely U-net, and a newly developed image semantic segmentation model, to provide pixel-wise steel bridge rust defect recognition. © 2021, Chinese Institute of Civil and Hydraulic Engineering. All right reserved.