Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Hybrid Training of Denoising Networks to Improve the Texture Acutance of Digital Cameras

Raphaël Achddou, Yann Gousseau, Saïd Ladjal

TL;DR

The paper addresses how to better preserve texture in digital camera outputs by introducing a texture acutance metric rooted in a cross-spectrum based modulation transfer function, weighted by a human contrast sensitivity function. It combines this acutance objective with conventional fidelity losses in a mixed training regime that uses both natural and synthetic dead leaves images to train denoising networks. Empirically, the method improves texture preservation (acutance) without compromising standard fidelity measures on AWGN denoising with FFDNet and also extends to RAW image denoising, where RAW acutance and RGB acutance are enhanced for moderate loss weights, with some trade-offs at high weights. The work offers a practical route to automatic texture-quality improvements and informs camera evaluation by integrating texture-focused metrics into training objectives.

Abstract

In order to evaluate the capacity of a camera to render textures properly, the standard practice, used by classical scoring protocols, is to compute the frequential response to a dead leaves image target, from which is built a texture acutance metric. In this work, we propose a mixed training procedure for image restoration neural networks, relying on both natural and synthetic images, that yields a strong improvement of this acutance metric without impairing fidelity terms. The feasibility of the approach is demonstrated both on the denoising of RGB images and the full development of RAW images, opening the path to a systematic improvement of the texture acutance of real imaging devices.

Hybrid Training of Denoising Networks to Improve the Texture Acutance of Digital Cameras

TL;DR

The paper addresses how to better preserve texture in digital camera outputs by introducing a texture acutance metric rooted in a cross-spectrum based modulation transfer function, weighted by a human contrast sensitivity function. It combines this acutance objective with conventional fidelity losses in a mixed training regime that uses both natural and synthetic dead leaves images to train denoising networks. Empirically, the method improves texture preservation (acutance) without compromising standard fidelity measures on AWGN denoising with FFDNet and also extends to RAW image denoising, where RAW acutance and RGB acutance are enhanced for moderate loss weights, with some trade-offs at high weights. The work offers a practical route to automatic texture-quality improvements and informs camera evaluation by integrating texture-focused metrics into training objectives.

Abstract

In order to evaluate the capacity of a camera to render textures properly, the standard practice, used by classical scoring protocols, is to compute the frequential response to a dead leaves image target, from which is built a texture acutance metric. In this work, we propose a mixed training procedure for image restoration neural networks, relying on both natural and synthetic images, that yields a strong improvement of this acutance metric without impairing fidelity terms. The feasibility of the approach is demonstrated both on the denoising of RGB images and the full development of RAW images, opening the path to a systematic improvement of the texture acutance of real imaging devices.
Paper Structure (12 sections, 8 equations, 4 figures, 2 tables)

This paper contains 12 sections, 8 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Texture acutance : a frequential loss assessing texture preservation
  4. Dead leaves images
  5. Texture acutance
  6. Acutance loss for image restoration CNNs
  7. Image denoising results with FFDNet
  8. Quantitative evaluation
  9. Spectral preservation
  10. RAW image denoising
  11. Conclusion
  12. Aknowledgements

Figures (4)

  • Figure 1: Grey level dead leaves image and its associated digital spectrum in logarthmic scales. The theoretical value is a straight line.
  • Figure 2: Diagram explaining the computation of the acutance metric
  • Figure 3: Comparison of FFDNet results on two natural images and on a dead leaves image. From left to right: original image, noisy image, image denoised with standard FFDNet, image denoised with FFDNet trained on a mixed database without the acutance loss, and finally with the acutance loss.
  • Figure 4: Comparison of the MTF evaluated with FFDNet trained on a mixed database with or without the acutance loss, on the whole dead leaves image test set.