Generative Data Augmentation for Rare Corrosion Types Using Diffusion Models

Abstract

The application of machine learning and deep learning methods has been very useful in detecting corrosion in industrial operations, but their effectiveness is usually constrained due to extreme class skew in the accessible datasets. There are rare yet essential types of corrosion like pitting and stress-corrosion cracking (SCC) and so on and in that case the type of data is insufficient so the use of generalization in a model becomes biased. This paper will explore how generative data augmentation can be applied to overcome this weakness and compare the Generative Adversarial Networks (GAN) frameworks and diffusion-based networks. Three training setups are tested, namely (i) real-only data (ii) real data combined with GAN-generated samples (iii) real data combined with diffusion-generated samples. Both classification and segmentation tasks are evaluated with the help of model performance measured in terms of the macro F1-score and Intersection over Union (IoU). Moreover, Frechet Inception Distance (FID), Kernel Inception Distance (KID) and Learned Perceptual Image Patch Similarity (LPIPS) are used to analyze the quality of the generated images. Empirical findings indicate that diffusion-based augmentation avenues always achieve better performance in comparison to GAN-based strategies, especially in exotic corrosion crop categories, in producing a wider and more lifelike samples. This is associated with a considerable enhancement in the predictive accuracy and localization performance. The results indicate that to manage the class imbalance, it is necessary to not only increase the amount of data but also provide a high-quality and diverse synthetic samples and therefore diffusion-based augmentation is a promising solution to detect corrosion when little data is available in the industry.