Image denoising techniques
AssigneeMicrosoft Corporation
InventorDenis Demandolx
Life of this patent
Estimating the term…
Abstract
Image denoising techniques include determining wavelet-domain noise model and a non-parametric multivariate wavelet description from the image signal for raw image data. A noise corrected image may then be determined from the image signal, the wavelet-domain noise model and the non-parametric, multivariate wavelet description and the image signal.
Claims12
An image denoising method comprising:
- decomposing an image signal into a wavelet domain using a discrete wavelet transform (DWT), wherein coefficients are grouped in sub-bands of different scales and orientations, including determining wavelet domain noise model and a non-parametric multivariate wavelet-description of an image;
- and
- determining a noise corrected image in the wavelet domain from the image signal, the wavelet-domain noise model and the non-parametric multivariate wavelet description including Gaussian filtering of probability density function applying a Bayesian multivariate minimum mean square error (MMSE) approximation using two-dimensional marginal projections of each given coefficient, the four nearest neighbors in a same sub-band of the given coefficient and a parent at the same pixel location and orientation in the next coarsest scale.
One or more non-transitory computing device readable media including instructions which when executed cause a computing device to implement an image denoising method comprising:
- determining a wavelet domain noise model and a non-parametric multivariate wavelet-domain description of a signal probability density function for an image signal;
- and
- determining a noise corrected image in the wavelet domain from the image signal, wavelet-domain noise model and the non-parametric multivariate wavelet-domain description of the signal probability density function applying a Bayesian multivariate minimum mean square error (MMSE) estimation using two-dimensional marginal projections of each given coefficient, the four nearest neighbors in a same sub-band of the given coefficient and a parent at the same pixel location and orientation in the next coarsest scale.
One or more non-transitory computing device readable media including instructions which when executed cause a computing device to implement an image denoising system comprising:
- a noise analyzer for receiving an image and generating a wavelet domain noise model and a non-parametric, multivariate wavelet-description of the image using a discrete wavelet transform (DWT);
- a noise correction engine for receiving the image, the wavelet-domain noise model and the non-parametric, multivariate wavelet description and generating a noise corrected image in the wavelet domain by applying a multivariate minimum mean square error (MMSE) approximation using two-dimensional marginal projections of each given coefficient, the four nearest neighbors in a same sub-band of the given coefficient and a parent at the same pixel location and orientation in the next coarsest scale;
- and
- transforming the noise corrected image in the wavelet domain back into the image domain using a transfer function including approximating with a one-dimensional estimator of the given wavelet coefficient, adding the one-dimensional approximation of an independent contribution of neighbor coefficients, and adding the approximation of a separable product of one-dimensional functions.
Dependencies are inferred from the claim text.