Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
119 tokens/sec
GPT-4o
56 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Noise2Void - Learning Denoising from Single Noisy Images (1811.10980v2)

Published 27 Nov 2018 in cs.CV

Abstract: The field of image denoising is currently dominated by discriminative deep learning methods that are trained on pairs of noisy input and clean target images. Recently it has been shown that such methods can also be trained without clean targets. Instead, independent pairs of noisy images can be used, in an approach known as Noise2Noise (N2N). Here, we introduce Noise2Void (N2V), a training scheme that takes this idea one step further. It does not require noisy image pairs, nor clean target images. Consequently, N2V allows us to train directly on the body of data to be denoised and can therefore be applied when other methods cannot. Especially interesting is the application to biomedical image data, where the acquisition of training targets, clean or noisy, is frequently not possible. We compare the performance of N2V to approaches that have either clean target images and/or noisy image pairs available. Intuitively, N2V cannot be expected to outperform methods that have more information available during training. Still, we observe that the denoising performance of Noise2Void drops in moderation and compares favorably to training-free denoising methods.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (3)
  1. Alexander Krull (19 papers)
  2. Tim-Oliver Buchholz (7 papers)
  3. Florian Jug (29 papers)
Citations (979)

Summary

Summary of "Noise2Void - Learning Denoising from Single Noisy Images"

The paper "Noise2Void - Learning Denoising from Single Noisy Images" introduces an innovative approach for training denoising convolutional neural networks (CNNs) using solely noisy images without the need for clean targets or pairs of noisy images. This method, coined as Noise2Void (N2V), extends the previously established Noise2Noise (N2N) framework, which required pairs of noisy images. N2V offers significant practical advantages, particularly for fields such as biomedical imaging, where acquiring clean or paired noisy images is often infeasible.

Theoretical Framework and Implementation

N2V is built on the assumption that signal pixels in an image are not statistically independent, while the noise is conditionally pixel-wise independent given the signal. This forms the basis for the blind-spot network architecture, a key feature in N2V. The blind-spot network excludes the central pixel from its receptive field, thus preventing the network from simply learning an identity mapping.

To implement this, the authors propose a masking scheme during the training phase, randomly selecting and replacing the center pixel within training patches, ensuring that no direct information of the pixel value is included in the receptive field. The network is thus trained to predict the masked pixel value based on its surrounding context, using the noisy image itself as both input and target.

Experimental Evaluation

Extensive experiments were conducted to validate the efficacy of N2V:

  1. BSD68 Dataset: Natural images with added Gaussian noise were used to train and evaluate the N2V model, with results indicating competitive performance compared to traditional and N2N-trained networks. Despite the theoretical constraints of having less information, the performance drop was only moderate relative to state-of-the-art methods such as BM3D and DnCNN.
  2. Simulated Microscopy Data: Simulated fluorescence microscopy data showed that N2V achieved results nearly on par with traditional and N2N training methods, demonstrating its robustness and applicability to different noise models and domains.
  3. Real Microscopy Data: The application of N2V to cryo-Transmission Electron Microscopy (cryo-TEM) and fluorescence microscopy datasets from the Cell Tracking Challenge showcased its practical utility. N2V produced visually appealing denoised images and exposed underlying structures in scenarios where other training methods could not be applied due to the lack of clean or paired noisy images.

Analysis of Limitations

The authors also address the limitations of N2V:

  • Complex Signal Patterns: N2V struggles with highly irregular and unpredictable signal patterns as it relies heavily on surrounding pixel context.
  • Structured Noise: If the noise is not pixel-wise independent, but has a structured pattern, N2V may leave these structures intact while removing pixel-wise independent noise, potentially revealing said patterns (e.g., striped artifacts in microscopy images).

Future Implications

The practical applications of N2V are extensive, especially in fields where acquiring clean training data is challenging. By enabling the training of denoising networks on single noisy images, N2V broadens the scope of image restoration tasks that can be tackled using deep learning-based methods. Furthermore, future research can delve into addressing the identified limitations, possibly by integrating domain-specific knowledge to handle structured noise more effectively.

Conclusion

Noise2Void represents a significant step forward in the field of image denoising by alleviating the dependency on clean or paired noisy training data. Its robust performance across various datasets and noise models demonstrates its potential for widespread use, especially in biomedical imaging applications where clean data acquisition poses substantial challenges. While not entirely without limitations, the proposed method opens new avenues for self-supervised learning in image restoration tasks.

Youtube Logo Streamline Icon: https://streamlinehq.com