An unsupervised deep denoising method.

[Paper] [Code] [Data]


Noise2Sim is the first similarity-based unsupervised deep denoising approach using noisy images only, suppressing not only independent but also correlated noises. Theoretically, Noise2Sim is equivalent to supervised denoising under mild conditions. On common benchmarks and practical CT datasets, Nosie2Sim recovers intrinsic structures from noisy low-dose and photon-counting CT images as effectively as or even better than the supervised learning methods. Read Paper for more details.


Noise2Sim package can be easily installed using pip install noise2sim. Note that Pytorch is required. Source codes are available at GitHub.


Noise2Sim has the potential in various applications. Here Noise2Sim was applied to denoising natural images with independent noise, low-dose CT (LDCT) and photon-counting CT (PCCT) images with correlated noises. All datasets can be freely accessed as follows.

Natural Images (grayscale): BSD400 and BSD68.

Natural Images (color): BSD500 and Kodak.

LDCT Data (Mayo): Mayo.

LDCT Data (FDA): LDCT-b40f, NDCT-b40f, LDCT-b60f, NDCT-b60f-1, NDCT-b60f-2.

PCCT Data (chicken drumstick): LDPCCT, NDPCCT, and NDPCCT-reference.

PCCT Data (live mouse): NDPCCT.

PCCT Data (died mouse): NDPCCT.


Instructions of training and testing denoising models with Noise2Sim package are here.