Contrastive GAN for shearography phase denoising: Unsupervised single-image training on wrapped fringe patterns

Abstract

Shearography (Speckle Shearing Interferometry) is widely used in industrial non-destructive testing due to its advantages of being non-contact and providing full-field measurement. However, its phase fringe patterns are corrupted by speckle noise, which severely affects the accuracy of dynamic monitoring. Traditional filtering methods (e.g., sine/cosine filtering) suffer from issues such as low computational efficiency and strong parameter dependency. Although existing deep learning solutions are effective, they rely on paired training data and often lack sufficient capability for high-resolution processing. To address this, this paper innovatively applies Single-image Contrastive Unpaired Training (SinCUT), a lightweight unidirectional style transfer denoising algorithm based on contrastive learning. This-method uses a single high-noise experimental phase map as the source domain (noisy image) and a single simulated ideal fringe pattern as the target domain (clean image), constructing an improved generative adversarial network to perform denoising on high-resolution experimental phase maps. Experimental results demonstrate that the SinCUT algorithm achieves fast processing time and excellent denoising performance, providing a viable solution for real-time non-destructive testing in industrial field applications.