Skip to content

Degradation_Model.md

Latest commit

 

History

History
39 lines (28 loc) · 2.25 KB

Degradation_Model.md

File metadata and controls

39 lines (28 loc) · 2.25 KB
Raw

This document outlines the stochastic degradation pipeline used to transform High-Resolution (HR) Ground Truth images into degraded Low-Resolution (LR) pairs. The pipeline utilizes a dual-model approach to simulate real-world distortions. We provide a detialed example assuming a 2x scaling factor and an input of 1080p resolution

1. Pre-Processing (Global)

  • Initial Input: GT $1920 \times 1080$.
  • Augmentation: Random Flip and Rotation.
  • Initial Crop: Random crop to $400 \times 400$ to ensure uniform patch size for the degradation branches.

2. Degradation Models

The pipelinecan be splitted into two primary paths based on the first_older_degradation_only flag.

  • Path A: Model = 1 (Standard Degradation) This path focuses on traditional blur-and-noise sequences:

    • Filter2D (Blur 1): Initial blurring of the $400 \times 400$ GT.
    • Resize Down (0.5): Spatial downsampling to $200 \times 200$.
    • Noise/JPEG: Application of optional additive noise and JPEG compression artifacts.
    • Result: Produces a degraded $200 \times 200$ intermediate.

  • Path B: Model = 0 (Complex/High-Order Degradation)A more complex path involving multiple resizing and filtering steps:

    • Random Resize: Dynamic upscaling ($1-1.5\times$) or downscaling ($0.15-1\times$).
    • Noise & JPEG: Integration of noise and compression.
    • Blur 2 & Second Resize: A second round of blurring and resizing (down to $\sim 200 \times 200$).
    • Sinc Filter Branch: Application of Sinc filters to simulate ringing artifacts and spectral cutoffs.

3. Ground Truth Enhancement (USM)

To produce "sharp" Ground Truths for certain training objectives, the pipeline includes an Unsharp Mask (USM) branch:

  • The $400 \times 400$ GT is sharpened using USM to enhance edge definition.
  • This creates a disparity between the "Natural GT" and the "Sharpened GT," often used to push the model toward higher perceptual clarity.

4. Final Output: Paired Cropping

The final step ensures spatial alignment between the degraded input and the target:

  • LR Generation: The degraded stream is cropped to $48 \times 48$.
  • GT Generation: The corresponding reference stream (either Natural or USM Sharpened) is cropped to $96 \times 96$.
  • Scale Factor: This resulting pair supports a $2\times$ Super-Resolution training task.