SDIP: Zero-shot AI restores blurry microscope images without training data
No training data? No problem – this AI restores blurry microscope images from scratch.
A new zero-shot deep image prior framework called SDIP sequentially performs denoising and deconvolution in fluorescence microscopy without external training data. It uses an aSeqDIP-based module for noise suppression and a wavelet-based background correction followed by RLG-DIP, which integrates the Richardson-Lucy deconvolution result as a physically consistent guidance prior. Tested on the BioSR dataset across multiple cellular structures, SDIP improves both signal-to-noise ratio and resolution, achieving superior visual quality and improved quantitative performance on most evaluated structures.
- Zero-shot: requires no paired training data, uses only the degraded image itself
- Combines sequential autoencoding for denoising and Richardson-Lucy guided DIP for deconvolution
- Outperforms supervised methods on the BioSR dataset, improving SNR and resolution across multiple cellular structures
Why It Matters
Enables high-quality microscopy restoration without costly paired datasets, accelerating biological and medical imaging research.