Probabilistic Noise2Void: Unsupervised Content-Aware Denoising (1906.00651v2)

Abstract: Today, Convolutional Neural Networks (CNNs) are the leading method for image denoising. They are traditionally trained on pairs of images, which are often hard to obtain for practical applications. This motivates self-supervised training methods such as Noise2Void~(N2V) that operate on single noisy images. Self-supervised methods are, unfortunately, not competitive with models trained on image pairs. Here, we present 'Probabilistic Noise2Void' (PN2V), a method to train CNNs to predict per-pixel intensity distributions. Combining these with a suitable description of the noise, we obtain a complete probabilistic model for the noisy observations and true signal in every pixel. We evaluate PN2V on publicly available microscopy datasets, under a broad range of noise regimes, and achieve competitive results with respect to supervised state-of-the-art methods.

Probabilistic Noise2Void: Unsupervised Content-Aware Denoising

The paper presents Probabilistic Noise2Void (PN2V), an innovative approach to unsupervised content-aware denoising utilizing Convolutional Neural Networks (CNNs). Image denoising is a fundamental task in image restoration, particularly crucial for microscopy data, where acquiring pairs of clean and noisy images for supervised models is often impractical. The Noise2Void (N2V) framework previously addressed this by enabling self-supervised training of CNNs on single noisy images, but its efficacy was limited compared to supervised methods that use paired image data. PN2V significantly advances this paradigm by offering a probabilistic model that predicts per-pixel intensity distributions and incorporates a generalized noise description.

Methodology

PN2V extends the N2V approach by constructing a complete probabilistic model for denoising that is not confined to Gaussian noise assumptions or intensity predictions. It accomplishes this by:

1. Noise Model Integration: Unlike traditional self-supervised approaches, PN2V integrates an adaptable noise model that can be represented as a histogram. This encompasses the observation likelihood based on the noise description.
2. Per-Pixel Intensity Distribution: CNNs trained with PN2V predict a distribution of possible true pixel intensities, allowing for a prior estimate constructed from predicted samples.
3. MMSE Inference: PN2V utilizes Minimum Mean Squared Error (MMSE) estimation to derive the final prediction, outperforming other self-supervised methods by effectively utilizing the probabilistic model.

Experimental Results

PN2V was tested on various publicly available microscopy datasets, using noise regimes spanning from raw data to progressively denoised setups. The results demonstrated that PN2V often provides denoising performance competitive with supervised CARE networks, highlighted by quantitative analyses through Peak Signal-to-Noise Ratio (PSNR) measurements. Notable PSNR improvements were observed across different noise levels on diverse datasets, underscoring PN2V's robustness and effectiveness.

Implications and Future Directions

PN2V establishes a new benchmark for self-supervised denoising methods, enabling applications that previously relied on the acquisition of clean image pairs. The probabilistic framework offers flexibility in terms of noise modeling, potentially expanding the applicability of CNN-based denoising to challenging environments such as low-light imaging scenarios.

From a theoretical standpoint, PN2V enriches the understanding of denoising algorithms by demonstrating the power of probabilistic modeling in self-supervised neural networks. It opens avenues for further research on probabilistic representations in deep learning architectures, particularly in areas requiring robust handling of noise variance and complexity.

Future developments may focus on refining the integration with diverse noise models and exploring its application beyond microscopy, potentially benefiting fields such as astronomy and medical imaging where noise reduction is pivotal. Furthermore, PN2V could inspire novel architectures that leverage probabilistic insights to enhance neural network training and inference across broader machine learning tasks.