Fatigue cracks are critical types of damage in steel structures due to repeated loads and distortion effects. Fatigue crack growth may lead to further structural failure and even induce collapse. Efficient and timely fatigue crack detection and segmentation can support condition assessment, asset maintenance, and management of existing structures and prevent the early permit post and improve life cycles. In current research and engineering practices, visual inspection is the most widely implemented approach for fatigue crack inspection. However, the inspection accuracy of this method highly relies on the subjective judgment of the inspectors. Furthermore, it needs large amounts of cost, time, and labor force. Non-destructive testing methods can provide accurate detection results, but the cost is very high. To overcome the limitations of current fatigue crack detection methods, this study presents a pixel-level fatigue crack segmentation framework for large-scale images with complicated backgrounds taken from steel structures by using an encoder-decoder network, which is modified from the U-net structure. To effectively train and test the images with large resolutions such as 4928 × 3264 pixels or larger, the large images were cropped into small images for training and testing. The final segmentation results of the original images are obtained by assembling the segment results in the small images. Additionally, image post-processing including opening and closing operations were implemented to reduce the noises in the segmentation maps. The proposed method achieved an acceptable accuracy of automatic fatigue crack segmentation in terms of average intersection over union (mIOU). A comparative study with an FCN model that implements ResNet34 as backbone indicates that the proposed method using U-net could give better fatigue crack segmentation performance with fewer training epochs and simpler model structure. Furthermore, this study also provides helpful considerations and recommendations for researchers and practitioners in civil infrastructure engineering to apply image-based fatigue crack detection.
Keywords: computer vision; deep learning; fatigue crack; semantic segmentation; steel structures.