Blocky artifacts on noisy fluorescence microscopy images — high-fidelity pre/post-processing recommendations (no retraining) #178
Description
Hi team, thanks for the great work!
I trained your network on a fluorescence microscopy (medical) dataset. Results are excellent on typical images, but when inputs are relatively noisy, the super-resolved outputs exhibit blocky/patch-like (tiling/checkerboard) artifacts.
Constraints
Use case: medical fluorescence microscopy.
Fidelity is critical: preserve morphology and intensity relationships; avoid hallucinations.
Scope: looking for inference-time solutions only (pre-/post-processing); I’d prefer not to retrain or fine-tune the model.
What I’ve tried
Pre-denoising (Gaussian blur, Non-Local Means, BM3D): reduces noise but can oversmooth fine structures.
Adjusting inference tiling (larger tile size/overlap, reflective/symmetric padding): modest improvement; artifacts persist on heavy noise.
Test-time augmentation + ensembling: small gain.
Light bilateral/guided filtering as post-processing: helps a bit but risks blurring thin filaments.
Questions
Recommended pre-processing pipelines for noisy fluorescence images that preserve structure and intensity? (e.g., background estimation/subtraction, variance-stabilizing transforms, edge-preserving denoising, normalization strategies)
Effective post-processing to suppress blockiness/tiling without degrading small structures? (e.g., deblocking, mild edge-aware smoothing, frequency-domain de-ringing)
Any inference settings in this repo known to mitigate patch artifacts? (tile size/overlap, padding modes, input dynamic-range normalization, anti-alias options)
Pitfalls to avoid to maintain medical-grade fidelity (operations that might distort intensity ratios or micro-morphology)?
I can share input/output crops that illustrate the artifacts if helpful.
Thanks in advance for any guidance!
