Source-linked AI summary
Invertible Denoising Network: A Light Solution for Real Noise Removal
Yang Liu, Zhenyue Qin, Saeed Anwar, Pan Ji, Dongwoo Kim, Sabrina Caldwell, Tom Gedeon
TL;DR
Real-image denoising is difficult for invertible networks because noisy inputs and clean outputs follow different distributions. InvDN separates clean content and noise in latent representations, then replaces the noisy representation for denoising and manipulates it for noise generation. It achieves state-of-the-art SIDD performance with substantially fewer parameters and also generates noise more similar to the original.
Problem
Invertible denoising must reconcile noisy inputs with clean outputs from different distributions while preserving the benefits of lightweight, information-lossless, memory-saving models.
Method
InvDN transforms a noisy input into a low-resolution clean image and noise-containing latent representation, replacing or altering the noisy latent representation for denoising or noise generation.
Results
InvDN achieves a new state-of-the-art result on the SIDD test set, using 2.64M parameters and less than 4.2% of DANet’s parameters.
Takeaways & Limitations
InvDN provides a lightweight invertible approach that restores clean images and generates new noisy images more similar to the original noise.
Takeaways & Limitations
Visual comparisons on RNI15 are provided without ground-truth images, so that dataset is evaluated only visually.
Abstract
from arXiv · showhide
Invertible networks have various benefits for image denoising since they are lightweight, information-lossless, and memory-saving during back-propagation. However, applying invertible models to remove noise is challenging because the input is noisy, and the reversed output is clean, following two different distributions. We propose an invertible denoising network, InvDN, to address this challenge. InvDN transforms the noisy input into a low-resolution clean image and a latent representation containing noise. To discard noise and restore the clean image, InvDN replaces the noisy latent representation with another one sampled from a prior distribution during reversion. The denoising performance of InvDN is better than all the existing competitive models, achieving a new state-of-the-art result for the SIDD dataset while enjoying less run time. Moreover, the size of InvDN is far smaller, only having 4.2% of the number of parameters compared to the most recently proposed DANet. Further, via manipulating the noisy latent representation, InvDN is also able to generate noise more similar to the original one. Our code is available at: https://github.com/Yang-Liu1082/InvDN.git.
1. Introduction
InvDN applies invertible networks to real image denoising, addressing the mismatch between noisy inputs and clean outputs while retaining lightweight, information-lossless, and memory-saving properties. It introduces separate latent variables for clean content and noise, enabling denoising and noise generation.
- Motivation: Invertible networks are lightweight, information-lossless, and memory-saving, making them suitable for denoising on small devices.Encoding and decoding share parameters, input details are preserved, and gradient computation uses constant memory regardless of network depth.
- Challenge: Real-image denoising is challenging for invertible models because noisy inputs and clean outputs follow different distributions.Traditional invertible models assume the input and reversed output share a distribution, whereas denoising must remove noise before reversion.
- Method: InvDN uses two latent variables: one encodes clean content with noise and high-frequency details, while the other encodes only clean content.The forward pass produces a downscaled, channel-expanded latent representation whose first three channels match the low-resolution clean image; remaining channels retain noisy signals.
- Contribution: InvDN is presented as the first invertible network designed for real image denoising.The contribution targets real-image denoising rather than synthetic-noise removal or noise generation alone.
- Results: InvDN achieves a new state-of-the-art result on the SIDD test set while using far fewer parameters and less runtime than previous state-of-the-art methods.The introduction reports this result alongside the model’s ability to generate new noisy images more similar to the original noise.
2. Related Work
Prior denoising work includes model-driven methods based on noise and image-prior assumptions and data-driven methods for real noise. Invertible approaches offered model compression and noise modeling, but no reported invertible network had addressed real-image denoising.
- Traditional Methods: Traditional denoising methods formulate maximum a posteriori optimization using noise-distribution assumptions and clean-image priors.Examples include mixture-of-Gaussian noise models, total variation, sparsity, and non-local similarity methods such as BM3D and NLM.
- Data-driven Deep Learning Denoising: Deep-learning denoising advanced from synthetic-noise tasks toward blind real-image denoising after datasets such as DND and SIDD became available.Real-noise methods include adapting existing approaches and modeling real noise with more complex distributions and architectures.
- Data-driven Deep Learning Denoising: VDN models real noise with distributional assumptions, whereas DANet avoids explicit noise assumptions and uses a GAN framework for denoising and noise generation.The related-work discussion notes potential limitations from unsuitable noise assumptions in VDN and unstable, slower GAN training in DANet.
- Invertible Networks: Noise Flow uses an invertible architecture to learn real noise distributions for augmentation but requires raw-RGB, ISO, and camera-specific information.It does not propose a new denoising backbone, and no invertible network for real-image denoising had been reported.
3. Invertible Denoising Network
InvDN adapts invertible networks to denoising by separating low-resolution clean content from high-frequency information that may contain noise, then replacing the noisy latent representation during reconstruction. Its architecture combines invertible wavelet downsampling with invertible blocks while preserving lightweight, information-lossless computation.
- Concept of Design: Invertible denoising is challenging because noisy inputs and clean outputs follow different distributions, requiring noise to be abandoned in latent space before reversion.Traditional invertible applications generally map between representations with the same distribution, unlike denoising.
- Network Architecture: Each DownScale Block applies an invertible wavelet transformation that halves spatial resolution and increases channels by 4×, followed by invertible blocks.Haar wavelets decompose representations into low- and high-frequency components; the resulting feature map changes from (H, W, C) to (H/2, W/2, 4C).
- Concept of Design: Because high-frequency clean details and noise are difficult to disentangle, InvDN discards the corresponding latent channels and samples zHF from N(0, I) to reconstruct the clean image.The sampled latent variable is combined with the learned low-resolution representation during inverse transformation.
- Concept of Design: InvDN transforms a noisy image into a low-resolution clean image and a latent representation containing high-frequency content and noise.The first three output channels are trained to match the downsampled clean image, while the remaining channels encode the other information.
- Network Architecture: The forward transformation learns the low-resolution component with a supervised objective, while the inverse transformation is trained to restore the clean image using sampled high-frequency latent variables.InvDN trains the invertible transformation jointly with forward and backward objectives.
- Network Architecture: Its coupling-based invertible blocks require inversion only of addition and multiplication, allowing the internal φ networks to use arbitrary architectures, including padded convolutions.Residual blocks are used for several φ operations, and skip connections pass low-frequency features to deeper layers.
4. Experiment
InvDN is evaluated on three real-world denoising benchmarks using quantitative, visual, and ablation studies. It achieves strong denoising with a lightweight architecture, while component choices and Monte Carlo self-ensembling further improve performance.
- Datasets and evaluation: InvDN is evaluated on SIDD, DND, and RNI15, using quantitative metrics where ground truth is available and visual comparisons for RNI15.SIDD and DND provide online test evaluations, whereas RNI15 has no ground truths.
- Quantitative results: 2.64M parameters give InvDN fewer than 4.2% of DANet’s parameters while achieving a new SOTA result on SIDD.Its DND performance is comparable to recent competitive models, indicating generalization of the lightweight model.
- Qualitative results: InvDN restores sharper patterns and subtle edges, while competing methods produce blur, artifacts, over-smoothing, blockiness, or random dots.The visual comparison covers images from SIDD, DND, and RNI15.
- Ablation study: Haar wavelets converge faster and more stably than squeeze transformations during training.The comparison reports PSNR on the validation set at the same training iteration.
- Ablation study: Residual blocks outperform dense blocks for denoising accuracy and use 2.6M rather than 4.3M parameters.The comparison concerns the network blocks φi.
- Ablation study: Increasing invertible blocks consistently improves validation accuracy, while two down-scale blocks perform best when invertible-block count is fixed.The study reports PSNR on the validation set at the same iteration.
- Ablation study: Using ℓ2 loss for low-resolution images and ℓ1 loss for high-resolution images achieves the highest performance; gradient and SSIM losses do not improve it.The comparison uses PSNR on the SIDD validation set.
- Monte Carlo self-ensemble: Monte Carlo self-ensembling averages multiple latent-variable samples and improves performance on 83.35% of SIDD validation images.The reported visualization uses an MC size of 16, and the averaged output further reduces residual noise.
5. Conclusion
InvDN addresses the distribution mismatch in invertible real-image denoising by separating clean content from noise, enabling both noise removal and generation.
- InvDN transforms noisy input into a low-resolution clean image and a noise-containing latent representation.
- During noise removal, InvDN replaces the noisy latent representation with a prior-sampled representation to restore clean images.
- InvDN can generate noise by altering the noisy latent vector and reconstructing new noisy images.
- Extensive experiments on three real-noise datasets demonstrate effectiveness in both noise removal and noise generation.