Automatic crack segmentation using deep high-resolution representation learning

Cracks are one of the most common types of surface defects that occur on various engineering infrastructures. Visual-based crack detection is a challenging step due to the variation of size, shape, and appearance of cracks. Existing convolutional neural network (CNN)-based crack detection networks, typically using encoder-decoder architectures, may suffer from loss of spatial resolution in the high-to-low and low-to-high resolution processes, affecting the accuracy of prediction. Therefore, we propose HRNet(e), an enhanced version of a high-resolution network (HRNet), by removing the downsampling operation in the initial stage, reducing the number of high-resolution representation layers, using dilated convolution, and introducing hierarchical feature integration. Experiments show that the proposed HRNet(e) with relatively few parameters can achieve more accuracy and robust performance than other recent approaches. (C) 2021 Optical Society of America