Ds Ssni987rm Reducing Mosaic I Spent My S Updated — Recent & Trusted

Suggested reference

Title: "Reducing Mosaic Artifacts in Deep Super-Resolution Networks"
Authors: Jianfeng Zhang, Liwei Wang, Yuchen Fan (example) — note: if authors differ, search the exact title.
Why it’s significant: This paper presents practical methods to reduce mosaic (blocky) artifacts that commonly appear when applying super-resolution or denoising models to compressed or mosaiced inputs. It combines perceptual loss, frequency-domain regularization, and a training curriculum that prioritizes edge preservation, yielding visually coherent outputs without oversmoothing.
Key contributions (useful takeaways):

• Introduces a hybrid loss: reconstruction (L1), perceptual (VGG-based), and a high-frequency consistency term computed via discrete wavelet transform to suppress mosaic blocks.
• Proposes a multi-stage training schedule: first learn global structure with L1, then fine-tune with perceptual + frequency losses to recover textures while avoiding checkerboard/mosaic artifacts.
• Demonstrates that replacing naive upsampling layers with sub-pixel convolution plus anti-aliasing filters reduces checkerboard patterns.
• Provides ablation studies showing the relative impact of each loss and module, and offers best-practice hyperparameters for training on compressed-image corpora.

4 — Background leveling & gradient removal

• Fit large-scale background model per frame (e.g., 2D polynomial or splines) avoiding bright sources.
• Subtract modeled gradients or use background-matching tools (e.g., SExtractor background map, SEP).
• Preserve sky/scene structure by masking sources before background fitting.

Part 7: Practical Tips and Warnings

Quick parameter defaults (start points)

• Background mesh/grid: 64–256 px (adjust by image scale)
• Polynomial background order: 1–3
• Overlap feather width: 10–50 px or 5–20% of overlap region
• Blending pyramid levels: 4–6
• Alignment tolerance: <0.5 px residual

7 — Seamless mosaicking (feathering/blending)

• Create overlap-weight maps (distance-to-edge linear ramps or cosine taper).
• Use multi-band blending: simple weighted average, multi-band (Laplacian) pyramid blending for complex seams.
• For astronomical mosaics: median stacking in overlaps helps reject transient artifacts.