Image Denoising Using a Library of Functions

[0001] The invention relates generally to image processing, and more particularly to processing images to reduce noise and enhance image content.

[0002] The goal of image denoising is to reconstruct a “noise free” enhanced image from an image corrupted with noise. Many image processing methods exist for image denoising.

[0003] The first group of methods uses local methods where a noise free estimate of a target pixel is reconstructed as a weighted average of pixel values within a small spatial neighborhood of the target pixel. One such method convolves the noisy image with a smoothing filter such as a Gaussian kernel function where weights decrease with distance from the target pixel according to the function. Due to the linear filtering operation, which is insensitive to the local image structure (pixel values) within the neighborhood, the resulting image usually has blurry artifacts.

[0004] To overcome the blurring artifact, other local methods attempt to make the filters “edge-aware” by using non-linear filters. A median filter reconstructs a pixel value as a median value of the pixel values within the neighborhood.

[0005] Another method uses anisotropic diffusion, where a shape of the filter can be adapted to local image structure as a result of a diffusion process, which is sensitive to image discontinuities.

[0006] A sigma filter identifies pixels that are similar to the target pixel within its neighborhood by thresholding the absolute pixel value difference between the target pixel and other pixels within the neighborhood. The target pixel is then reconstructed by an average of only the similar pixel values (pixels with difference within the similarity threshold) within the neighborhood. The weights in that model can be considered as being 0 or 1 according to similarity criteria.

[0007] Several extensions to sigma filter are known. For example, one can use a bilateral filter where, instead of a hard thresholding operator (0 or 1 weights), the weights are continuously varied according to a multiplication of two kernel functions, one in spatial domain, and one in range (pixel value) domain.

[0008] Another method examines different local image areas around the pixel of interest, and determines a noise free estimate for each local area. The estimate for the pixel of interest is then determined as some weighted average of the local area estimates.

[0009] Another method determines a noise variance for each pixel value, which is stored in a lookup table. A similarity threshold and weights are then adjusted based on this noise model.

[0010] Local denoising methods benefit from spatial locality, which allows fast computation. However, those methods fail to use global statistics of an image, such as repeating patterns, which is an important cue for image denoising.

[0011] The second group of methods uses non-local denoising methods. The non-local method retrieves image patches similar to the target image patch by searching all patches within the image. Then, the target patch is replaced by a weighted average of the similar patches.

[0012] Alternative non-local methods include sparse reconstruction of image patches using a learned dictionary from the same image, and transform domain non-local filtering. The non-local search step of those methods makes them difficult to use in systems with computational and memory constraints.

[0013] Alternatively, non-local denoising can be achieved by performing inference on a Markov random field (MRF) using graph cuts or belief propagation. To achieve fast performance, the number of iterations is small, leaving considerable noise in the result.

[0014] The embodiments of the invention provide a method for denoising a noisy image to generate a noise free enhanced image. The method uses a library of local denoising functions. For each pixel of the image, a key is constructed using a local neighborhood (patch) around the pixel. The key maps each pixel to a denoising function in the function library. The noise free reconstruction of the pixel is determined by applying the selected function to the patch of pixels.

[0015] The library of denoising functions can be learned, for example, using training images in an offline process. The learning process minimizes a reconstruction error. The method combines non-linear mapping through the keys with the functions. The functions are optimized for various patch configurations to provide better reconstruction than existing hand tuned local denoising methods, while still allowing fast processing.

[0023] FIG. 1 is a flow diagram of a method for denoising 110 a noisy image 101 using a library 501 of local denoising functions to produce a denoised image 102 according to embodiments of the invention. The method can be performed in a processor connected to memory and input/output interfaces by buses as known in the art.

[0024] FIG. 2 is a detailed flow diagram of the denoising method. A pixel x in the noisy image has a value I(x). A local neighborhood or “patch” around the pixel is P(x). The patch can have arbitrary shapes such as a rectangle, ellipse, or an irregular arrangement of pixels adapted to the image content.

[0025] Noise estimator 210 determines a noise variance σ 211 for each pixel using the local neighborhood or patch of pixels P(x) in the noisy image 101. The function mapping m(x) 510 maps each pixel to a function 230 using the noise variance estimation and a local neighborhood of the pixel, see FIG. 5. The selected function ƒ 230 is applied 240 to the patch P(x) in the noisy image, to generate a corresponding denoised pixel for the denoised image 102.

[0026] During the noise estimation 210, the method estimates the variance of the noise for each pixel. In one embodiment the variance a is assumed to be identical for all pixels having the same value v

if I(x)=v and I(y)=v, then σ(x)=σ(y)=σ_v.

[0027] The variance is estimated using local patches around each pixel. For each patch in the image with mean value v, the variance of the intensity values of the patch is determined, one variance for each patch. If the patch originates from a constant color area, then the variance of this patch is equal to an empirical estimation of the noise.

[0028] However, because the patch can also originate from a non-constant color area, the patch variance estimate is approximately the lower bound of the noise variance distribution for pixels with value v. The noise variance of value v is estimated as k^th order statistics of a distribution of noise variances for central pixels of the patches with an average value v. In one embodiment, k is selected as, for example, 0.1 of the number of pixels with value v. of a distribution of variance computations of patches central pixel value

[0029] In general, the variance changes smoothly with changing pixel values. Independent noise estimation for each value results in non-smooth noise profiles. Therefore, the variances are smoothed to produce smooth noise profile.

[0030] Function Mapping

[0031] For each pixel x of the image, a key is constructed using the patch P(x) around the pixel x and the noise estimate for the pixel value. The key includes a spatial key, and a noise key.

[0032] As shown in FIG. 3 the spatial key 302 is constructed by the spatial mapping function 310 using a patch P(x) 301 of, e.g., 3×3, pixels I(x_i), x_iεP(x) around the pixel x. The size and the shape of the patch P(x) can be different than the patch that was used to estimate the variance of the noise.

[0033] In one embodiment, the spatial key is a local n-ary pattern (LnP), 320 for example a local binary pattern (LbP). For each pixel x_i within the patch one bit of information is acquired. If a difference between the pixel value I(x_i) in the patch and the pixel value I(x) in the noisy image is smaller than the variance for the pixel σ(x), then the bit is set to 0 otherwise the bit is set 321 to 1:

$s_i\left(x\right)=\{\begin{array}{cc}0& I\left(x\right)-I\left(x_i\right)\le k·\sigma \left(x\right)\\ 1& I\left(x\right)-I\left(x_i\right)>k·\sigma \left(x\right)\end{array},x_i\in P\left(x\right).$

[0034] The spatial key has bits. Examples of determined local binary patterns are shown in FIG. 3 where black pixels (0 bits) correspond to pixels within the patch having similar values to the pixel in the noisy image, i.e., the center pixel in the patch, and white pixels (1 bits) correspond to pixels within the patch having dissimilar values to the pixel in the noisy image.

[0035] FIG. 4 is a flow diagram of the noise mapping. A noise key 402 is constructed by the noise mapping function 410 using the estimation of the variance 401 for the pixel in the noisy image. In one embodiment, the noise key is given by an n bit uniform quantization of variance σ(x):

n(x)=|n| bit quantization of σ(x).

[0036] The key m for the pixel is a concatenation of the spatial key and the noise key as follows:

$m\left(x\right)=\left(\begin{array}{c}s\left(x\right)\\ n\left(x\right)\end{array}\right),$

[0037] such that the key has |s|+|n| bits.

[0038] As shown in FIG. 5, the key m 502 maps 510 a pixel of the image to a denoising function ƒ in the denoising function library 501 that is used to denoise the pixel.

[0039] Denoising Using a Library Function

[0040] As shown in FIG. 6, the key k 502 is determined from noisy image pixel x, and is the corresponding function in the function library. The denoising function uses a patch P(x) 601 around the pixel, i.e., the patch includes locally neighboring pixels around the pixel x. The size and the shape of this patch can be different than the patch that was used to determine the key. In one embodiment the denoising function is a linear function of the pixel intensity in the patch

ƒ^k(x)=W^kI(P(x)),

where W^k 602 is a weight matrix for the linear function, I(P(x)) is the matrix of pixel intensities for the patch P(x), and is an inner product operator between two matrices after vectorization by stacking columns of the matrix into a column vector.

[0041] Learning Denoising Function Library

[0042] As shown in FIG. 7, the denoising function library 501 is learned using training image samples 721. Training image samples 721 include noise free 701 and corresponding noisy 711 training image sample pairs. The training samples 721 are generated by the training samples constructor 720. In one embodiment, the noisy 711 images are acquired by adding 710 synthetic noise to the noise free 701 images.

[0043] The training or function fitting 730 optimizes the functions such that a difference between a reconstruction of the noise free image using the noisy 711 image and the noise free 701 image is minimized:

$f^k=\arg \underset{f^k}{\min }{F\left(\tilde{I}\right)-I}_2,k=1\dots l,$

[0044] where I and Ĩ are the noise free and noisy training image sample pairs 721, l is the number of denoising functions in the function library and denoising function F operates per image pixel by first mapping the pixel to the denoising function using the function mapping 510 and then uses the mapped function to denoise the pixel:

F(x)=ƒ^m(x)(x).

[0045] In one embodiment, the training is solved by grouping the pixels of the training image pairs according to the keys. Then, each group is optimized separately. When the library functions are linear functions they are learned optimally by solving a linear least squares problem. When they are non-linear functions, they are learned using non-linear optimization techniques such as gradient descent or Newton's method.

[0046] Although the invention has been described by way of examples of preferred embodiments, it is to be understood that various other adaptations and modifications can be made within the spirit and scope of the invention. Therefore, it is the object of the appended claims to cover all such variations and modifications as come within the true spirit and scope of the invention.