Two-stage fatigue crack detection framework with crack-preserving downsampler

Inspection of steel bridges is essential for maintaining structural integrity and ensuring public safety. Automation of such inspections using neural networks for the visual detection of fatigue cracks is a prominent way to improve structural reliability and operational efficiency. This is often done using multiple neural networks to ensure the reliability of the results. Therefore, in this work, a two-stage crack detection and sizing framework for images of steel bridges is proposed and analysed in detail, which combines two neural networks. Additionally, it is shown that standard image downsampling methods can be non-optimal for the crack detection task because of the small width of the cracks at the surface. Hence, image downsampling is an important step for automatic crack detection. This is applied to the images contained in the Cracks in Steel Bridges (CSB) dataset In this work, a crack-preserving downsampling method is introduced, which is designed to downsample images in such a way that (our two-stage) crack detection in images of steel bridges shows higher performance.