Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Rethinking Skip Connections: Additive U-Net for Robust and Interpretable Denoising

Vikram R Lakkavalli

TL;DR

This work addresses the lack of interpretability and channel inflation in U-Net denoisers caused by concatenative skip connections. It introduces Additive U-Net, a lightweight encoder–decoder that uses gated additive skips with learnable nonnegative scalars $\alpha_j$, enabling explicit control over cross-scale information flow without channel doubling or down/up-sampling. Empirical results on Kodak-17 show competitive PSNR/SSIM across Gaussian noise levels $\sigma \in \{15,25,50\}$, with qualitative evidence of improved edge preservation and noise leakage mitigation; interpretability is enhanced through gate analyses and frequency-domain observations that reveal a natural high-to-low frequency progression. The paper argues that controlling information flow can be as important as depth, offering a practical and transparent design for denoising and potential extensions to multi-task settings in medical and scientific imaging.

Abstract

Skip connections are central to U-Net architectures for image denoising, but standard concatenation doubles channel dimensionality and obscures information flow, allowing uncontrolled noise transfer. We propose the Additive U-Net, which replaces concatenative skips with gated additive connections. Each skip pathway is scaled by a learnable non-negative scalar, offering explicit and interpretable control over encoder contributions while avoiding channel inflation. Evaluations on the Kodak-17 denoising benchmark show that Additive U-Net achieves competitive PSNR/SSIM at noise levels σ = 15, 25, 50, with robustness across kernel schedules and depths. Notably, effective denoising is achieved even without explicit down/up-sampling or forced hierarchies, as the model naturally learns a progression from high-frequency to band-pass to low-frequency features. These results position additive skips as a lightweight and interpretable alternative to concatenation, enabling both efficient design and a clearer understanding of multi-scale information transfer in reconstruction networks.

Rethinking Skip Connections: Additive U-Net for Robust and Interpretable Denoising

TL;DR

This work addresses the lack of interpretability and channel inflation in U-Net denoisers caused by concatenative skip connections. It introduces Additive U-Net, a lightweight encoder–decoder that uses gated additive skips with learnable nonnegative scalars , enabling explicit control over cross-scale information flow without channel doubling or down/up-sampling. Empirical results on Kodak-17 show competitive PSNR/SSIM across Gaussian noise levels , with qualitative evidence of improved edge preservation and noise leakage mitigation; interpretability is enhanced through gate analyses and frequency-domain observations that reveal a natural high-to-low frequency progression. The paper argues that controlling information flow can be as important as depth, offering a practical and transparent design for denoising and potential extensions to multi-task settings in medical and scientific imaging.

Abstract

Skip connections are central to U-Net architectures for image denoising, but standard concatenation doubles channel dimensionality and obscures information flow, allowing uncontrolled noise transfer. We propose the Additive U-Net, which replaces concatenative skips with gated additive connections. Each skip pathway is scaled by a learnable non-negative scalar, offering explicit and interpretable control over encoder contributions while avoiding channel inflation. Evaluations on the Kodak-17 denoising benchmark show that Additive U-Net achieves competitive PSNR/SSIM at noise levels σ = 15, 25, 50, with robustness across kernel schedules and depths. Notably, effective denoising is achieved even without explicit down/up-sampling or forced hierarchies, as the model naturally learns a progression from high-frequency to band-pass to low-frequency features. These results position additive skips as a lightweight and interpretable alternative to concatenation, enabling both efficient design and a clearer understanding of multi-scale information transfer in reconstruction networks.
Paper Structure (21 sections, 3 equations, 6 figures, 1 table)

This paper contains 21 sections, 3 equations, 6 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Image denoising architectures
  4. Skip connection designs
  5. Multi-task extensions
  6. Proposed Method: Additive U-Net
  7. Design Motivation
  8. Architecture
  9. Loss and Training
  10. Variants
  11. Experimental Setup
  12. Architectural Variants
  13. Comparison Baselines
  14. DnCNN.
  15. Pseudo-additive U-Net.
  16. ...and 6 more sections

Figures (6)

  • Figure 1: Additive U-Net: each skip connection is scaled by a learnable coefficient $\alpha_j$, controlling information flow from encoder to decoder without channel concatenation. Corresponding encoder/decoder share same number of channels.
  • Figure 2: Visual comparison (Kodak-17, Test009 and Test005, $\sigma$=15). DnCNN oversmooths, Pseudo Add-U-Net leaves blotchy artifacts, while Real Add-U-Net preserves edges and natural textures.
  • Figure 3: Role of $\alpha$ in decoding. Varying $\alpha_3$ from 0–1 yields smooth changes in PSNR/SSIM, with peak near the learned value, confirming interpretable and optimized feature fusion.
  • Figure 4: Frequency analysis of filters. Encoder layers specialize from high- to low-frequency features, showing an emergent hierarchical progression without explicit multi-scale design.
  • Figure 5: Kernel schedule comparison ($\sigma$=15). The 9–7–5–3–1 variant emphasizes fine detail but retains grain, while 5–5–5–5–5 produces smoother backgrounds and sharper global contrast.
  • ...and 1 more figures