Dual-Path Adversarial Denoising Network Based on UNet

Jinchi Yu, Yu Zhou, Mingchen Sun, Dadong WangSensors 2025 · DOI

TL;DR

This paper introduces a novel "dual-path adversarial denoising network" that uses a UNet-based denoiser and a generator to create realistic noisy images. This approach effectively removes noise while preserving fine image details and global structures, outperforming existing methods on benchmark datasets.

What Problem Does This Paper Address?

Digital images are prone to noise, which degrades their quality and hinders analysis in critical applications like medical imaging and remote sensing. Traditional denoising methods often struggle to strike a balance between removing noise and preserving essential image details. They also lack adaptability to diverse noise types and can fail to capture global image structures. Existing deep learning approaches, while promising, can be data-hungry and sensitive to discrepancies between simulated and real-world noise.

Methods

The proposed solution is a three-module architecture:

  • Generator (U-shaped Imitation Noise Generation Network - UING): This module uses a UNet-like structure to generate synthetic noisy images. By learning the distribution of real noise, it creates realistic "fake" noisy images that augment the training data. This helps the denoiser become more robust to various noise types.
  • Denoiser (Double-Layer U-Denoise Network - DUNetD): This is the core denoising component, featuring a dual-UNet architecture. It employs two UNet branches that process image features at different levels. Crucially, it incorporates:
  • Multi-Receptive Field Perception Denoise Module (MRDB): This module uses dilated convolutions with varying rates to capture multi-scale contextual information, enhancing the ability to preserve fine details.
  • Atrous Spatial Pyramid Pooling (ASPP): Integrated within the MRDB, ASPP further expands the receptive field and captures multi-scale information without increasing computational cost or reducing resolution.
  • Discriminator (Dual-Channel Adversarial Discriminator - D2P): This module acts as a judge, distinguishing between real noisy images and those generated by the UING, and also between real clean images and the output of the DUNetD. This adversarial feedback loop drives the generator to produce more realistic noise and the denoiser to achieve better noise removal and detail preservation.

The training process involves a "dual-path adversarial training" where the generator and denoiser work in tandem, guided by the discriminator. This creates a closed-loop optimization for "noise synthesis-noise removal."

Datasets

The paper utilizes three established datasets for training and evaluation:

  • SIDD (Smartphone Image Denoising Dataset): Contains 30,000 noisy images across 10 scenes, captured under various lighting conditions, with corresponding ground-truth images.
  • DND (Darmstadt Noise Dataset): Features 50 images from consumer cameras, with noisy images captured at higher ISO settings. Ground truth images were carefully processed but are not publicly available for all images.
  • PolyU Dataset: Comprises 40 scenes, with an average of multiple images used to create a near "ground truth." This dataset includes both noisy images and noise-free references.

The SIDD-Medium dataset was used for training, with 256x256 patches extracted. Testing was performed on the DND, SIDD, and PolyU datasets.

Key Insights and Ideas

  • Dual-Path Adversarial Learning for Denoising: The core innovation is using adversarial training not just for image generation but specifically for noise simulation to improve denoiser robustness. This creates a more challenging yet realistic training environment.
  • Dual-UNet for Multi-Scale Feature Fusion: The stacked UNet architecture allows for hierarchical feature extraction and fusion at different scales, which is critical for balancing noise removal with detail preservation.
  • MRDB and ASPP for Enhanced Detail Preservation: The MRDB, incorporating ASPP, is key to capturing fine image details by effectively utilizing multi-scale receptive fields, a common challenge in denoising.
  • Generator as a "Noise Forger": The generator's role is not just to create noise but to "forge" it in a way that is indistinguishable from real noise, thereby pushing the denoiser to its limits and improving its generalization.
  • Early Stopping for Efficiency: The implementation of an early-stopping mechanism helps prevent overfitting and control training time, making the approach more practical.

Key Results

The proposed dual-path adversarial denoising network achieved state-of-the-art performance on the SIDD, DND, and PolyU datasets.

  • Quantitative Improvements: On the SIDD dataset, the model achieved a PSNR of 39.29 dB and an SSIM of 91.50%, outperforming other methods like Trident GAN and CS-PCN. Similar superior results were observed on the DND dataset.
  • Qualitative Improvements: Visual comparisons showed that the proposed method effectively removed noise while preserving fine textures and structural integrity, producing cleaner and more visually appealing images compared to other methods.
  • Computational Efficiency: The model is designed to be computationally efficient, with an end-to-end latency of 55 ms, which is suitable for real-time applications.

Limitations and Future Work

  • Noise Stability: While robust to common noise patterns (Gaussian, salt-and-pepper), the model's stability may degrade with structured or non-stationary noise (e.g., Perlin noise).
  • Parameter Sensitivity: The model's performance can be sensitive to noise intensity and distribution, potentially affecting the adversarial equilibrium.
  • Fixed Dilation Rates: The fixed dilation rates in the ASPP modules might limit adaptability to noise scales outside the trained range.
  • Future Directions: The authors suggest exploring dynamic dilation mechanisms or noise-specific attention modules to enhance generalization. They also highlight the need for noise-aware training protocols or hybrid methods to address complex real-world scenarios and the trade-off between generality and specialization.

Summary updated

Abstract

Digital image quality is crucial for reliable analysis in applications such as medical imaging, satellite remote sensing, and video surveillance. However, traditional denoising methods struggle to balance noise removal with detail preservation and lack adaptability to various types of noise. We propose a novel three-module architecture for image denoising, comprising a generator, a dual-path-UNet-based denoiser, and a discriminator. The generator creates synthetic noise patterns to augment training data, while the dual-path-UNet denoiser uses multiple receptive field modules to preserve fine details and dense feature fusion to maintain global structural integrity. The discriminator provides adversarial feedback to enhance denoising performance. This dual-path adversarial training mechanism addresses the limitations of traditional methods by simultaneously capturing both local details and global structures. Experiments on the SIDD, DND, and PolyU datasets demonstrate superior performance. We compare our architecture with the latest state-of-the-art GAN variants through comprehensive qualitative and quantitative evaluations. These results confirm the effectiveness of noise removal with minimal loss of critical image details. The proposed architecture enhances image denoising capabilities in complex noise scenarios, providing a robust solution for applications that require high image fidelity. By enhancing adaptability to various types of noise while maintaining structural integrity, this method provides a versatile tool for image processing tasks that require preserving detail.
BibTeX
@inproceedings{yu2025dual,
  author = {Jinchi Yu and Yu Zhou and Mingchen Sun and Dadong Wang},
  title = {{Dual-Path Adversarial Denoising Network Based on UNet}},
  booktitle = {Sensors},
  year = {2025},
  doi = {10.3390/s25154751},
}
Figure