Denoising with Noise2Void

Training deep-learning models for denoising usually relies on either pairing high-noise input images with low-noise output images or using independent pairs of noisy images in an approach known as Noise2Noise (N2N). These approaches can be limited if the acquisition of low-noise or noisy training targets is not possible, as is often the case for biomedical image studies.

As an additional option for training deep-learning models for denoising, an approach known as Noise2Void (N2V), in which training is done directly on the data to be denoised, is available. Below is an example of the results for applying a trained N2V model.

Original image (on left) compared with denoised version (on right)

Advantages of Noise2Void Denoising

• Other classic deep-learning models for denoising, such as noise-to-low noise and noise-to-noise, may not be usable if the required outputs cannot be acquired. For Noise2Void models, each pixel is replaced with another pixel in the same image.

• Training N2V models usually does not require many epochs.

• N2V models are very good for reducing salt and pepper noise.