Applying fully convolutional neural networks for corrosion semantic segmentation for steel bridges: The use of U-Net
S.-K. Chen, I-Feng Huang, P.-H. Chen
Abstract
As the weather in Taiwan is warm and humid most of the time, steel bridges get rusted easily. Nowadays, bridges are significant infrastructure in most countries, and, thus, it is crucial to come up with an effective corrosion detection method for steel bridge inspection, so as to maintain the health of steel bridges and reduce the lifecycle costs of them at the same time. Browsing past research efforts, there were a number of image processing techniques (IPTs) proposed for quick and effective rust image recognition. A crucial issue on rust recognition is to distinguish real rust corrosion spots or areas from noises or patterns that look like rust. Also, the types of rust and different rust colors would affect the accuracy of rust recognition. In view of the above issues, a fully convolutional neural network, namely U-Net, will be explored in this paper to develop an image semantic segmentation model, which will be able to deal with a wide range of rust image recognition.