Paper ID 227 - ISARC 2024 Keynote (Best Paper) by Honghu Chu

Title: CrackGauGAN: Semantic Layout-based Crack Image Synthesis for Automated Crack Segmentation
Authors: Honghu Chu, Weiwei Chen and Lu Deng
Abstract: Automated crack inspection, particularly deep learning (DL)-based crack segmentation, is crucial for the effective and efficient maintenance, repair, and operation of civil infrastructure. However, the performance of DL-based segmentation methods is often limited by the scarcity of pixel-wise labeled crack images. This paper presents CrackGauGAN, an automated crack image synthesis network that can be used to generate realistic and diverse crack image and mask pairs, which are instrumental in improving the performance of DL-based crack segmentation models. The CrackGauGAN is developed with three customized improvements based on the original GauGAN architecture. Firstly, a Criminisi-based crack image inpainting operator is introduced before the image encoder, enabling the exclusion of crack noise interference during background color feature extraction. Secondly, a background texture extraction method is proposed, assisting the SPADE-based generator in decoupling background textures as prior information. Lastly, an adaptive pseudo-augmentation strategy is introduced in the discriminator, allowing the model to be effectively trained on small-scale crack datasets. Ablation studies are conducted to prove the effectiveness of each component, and further crack image generation experiments demonstrate that the CrackGauGAN can synthesize various cracks with excellent diversity and fidelity. The CrackGauGAN-generated crack images show average improvements of over 1.97 and 7.91 in the Inception Score (IS) and Fréchet Inception Distance (FID), respectively, compared to the previously most advanced GauGAN and Pix2PixHD. As a fully automated crack image mask pair generation architecture, the CrackGauGAN can be used to provide reliable data support for the application of DL-based segmentation models in crack inspection tasks.
Keywords: Deep Learning, Generative Adversarial Network, Crack Image Synthesis, Feature Decoupling, Crack Segmentation
DOI: doi.org/10.222...