High-precision corrosion degree nondestructive segmentation method with virtual and real synthetic data labeled by unsupervised learning

Corrosion is a significant issue for materials, leading to economic losses and potential safety accidents. Corrosion degree detection allows the assessment of its impact on materials, providing crucial safety and performance information essential for maintaining and managing asset integrity. This study proposes an intelligent detection technology based on the pixel-level location of surface corrosion area and corrosion degree recognition of carbon steel samples. First, a corrosion acceleration test was employed to corrode the samples to various degrees. A generative adversarial network (GAN), StyleGAN3-t expands the corrosion image, reducing the experimental workload and sample requirements. A semi-automatic labeling approach using the Segment Anything Model (SAM) was introduced for rapid and high-resolution identification of corroded regions with complex shapes. Lastly, this paper presents the MN-DeepLabv3, which replaces the DeepLabv3 backbone network Xception with MobileNetV2, for training real corroded and generated virtual images, respectively. Experiments show that MN-DeepLabv3 outperforms other algorithms in segmenting the corrosion area and recognizing the corrosion degree. This approach presents a promising technical strategy for intelligent detection of carbon steel surface corrosion.