Unsupervised video denoising for structured noise via Fourier masking (Invited Paper)

We present an unsupervised video denoising method designed to remove structured noise, such as stripes or other random patterns common in microscopy and remote sensing, without needing clean training data. Unlike existing spatial-domain approaches, our method operates in the Fourier domain, leveraging the observation that Fourier coefficients of structured noise exhibit independence. By randomly masking and replacing Fourier coefficients across frames, the model learns to recover clean video content. Results show effective noise suppression without prior noise modeling, offering a practical solution for real-world imaging systems where clean references are unavailable.