🤖 AI Summary
This work addresses the limitations of standard U-Net’s concatenation-based skip connections, which often induce channel dimension explosion and fail to suppress noise propagation, thereby compromising denoising performance and interpretability. To overcome these issues, the authors propose Additive U-Net, which replaces conventional concatenation with additive skip connections modulated by learnable non-negative gating scalars. This lightweight and interpretable mechanism explicitly controls the contribution of encoder features, effectively avoiding channel expansion while revealing a natural progression of multi-scale features from high- to low-frequency components. Evaluated on the Kodak-17 benchmark, the model achieves competitive PSNR and SSIM across noise levels (σ = 15, 25, 50) and demonstrates strong robustness to variations in network depth and convolutional kernel scheduling.
📝 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 {\sigma} = 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.