A Note on the Regularity of Images Generated by Convolutional Neural Networks
Not on the map. Why?
Abstract (arXiv)
The regularity of images generated by convolutional neural networks, such as the U-net, generative networks, or the deep image prior, is analyzed. In a resolution-independent, infinite dimensional setting, it is shown that such images, represented as functions, are always continuous and, in some circumstances, even continuously differentiable, contradicting the widely accepted modeling of sharp edges in images via jump discontinuities. While such statements require an infinite dimensional setting, the connection to (discretized) neural networks used in practice is made by considering the limit as the resolution approaches infinity. As practical consequence, the results of this paper in particular provide analytical evidence that basic L2 regularization of network weights might lead to over-smoothed outputs.
Key facts. Cited by 2 papers in the map (data through 14 Oct 2026). Builds on 20 of its 50 references that are in the map. First posted on arXiv 22 Apr 2022.
Builds on
20 of 50 references.
References · 20
- 2015U-Net: Convolutional Networks for Biomedical Image Segmentationcited by 19,713
- 2016Image-to-Image Translation with Conditional Adversarial Networkscited by 4,527
- 2016Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Networkcited by 1,908
- 2016Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoisingcited by 1,358
- 2018On the Spectral Bias of Neural Networkscited by 1,341
- 2017Deep Image Priorcited by 1,098
- 2017FFDNet: Toward a Fast and Flexible Solution for CNN based Image Denoisingcited by 510
- 2016Deep Convolutional Neural Network for Inverse Problems in Imagingcited by 433
- 2017Compressed Sensing using Generative Modelscited by 339
- 2018Deep Decoder: Concise Image Representations from Untrained Non-convolutional Networkscited by 120
- 2019Invertible generative models for inverse problems: mitigating representation error and dataset biascited by 94
- 2019Robust and interpretable blind image denoising via bias-free convolutional neural networkscited by 90
8 more
- 2016Understanding Deep Convolutional Networkscited by 76
- 2018Phase Retrieval Under a Generative Priorcited by 73
- 2021GANs for Medical Image Synthesis: An Empirical Studycited by 65
- 2017Automatic Brain Tumor Detection and Segmentation Using U-Net Based Fully Convolutional Networkscited by 42
- 2019Denoising and Regularization via Exploiting the Structural Bias of Convolutional Generatorscited by 40
- 2021A Generative Variational Model for Inverse Problems in Imagingcited by 14
- 2020Deep synthesis regularization of inverse problemscited by 7
- 2018A convex variational model for learning convolutional image atoms from incomplete datacited by 3
Built on by
Citing papers per year: 2022: 1; 2023: 1.
2 papers.