FIELD OF THE INVENTION

[0001]
The present invention relates generally to video image processing, and more particularly to noise reduction in video images.
BACKGROUND OF THE INVENTION

[0002]
Video images are often corrupted by noise during the video image acquisition or transmission process. In order to improve the visual appearance of such images, an effective filtering technique is often required to reduce the noise level therein. Traditional noise reduction techniques mainly involve applying a linear filter such as an averaging filter to all of the pixels in a video frame (“image”). While this reduces noise level in the image, such a linear filtering technique also indiscriminately blurs edges in the image.

[0003]
In order to prevent image edge blurring, a noise reduction filter must be adaptive to local structures, such as edges, in the image. One such adaptive technique is known as directional filtering. Directional filtering attempts to avoid image blurring by adapting linear filtering to image edge directions in such a way that the filter utilized is always applied along the edge direction not across the edge direction.

[0004]
FIG. 1 shows a block diagram of an example directional filter 100. At each image pixel, first the 2D local variance is computed by a local variance calculator 120 for a small window. Then, the 1D local variances are computed along the horizontal, vertical, diagonal from upper left to lower right, and diagonal from upper right to lower left directions within the same window of pixels. To determine the edge direction, the 2D variance is compared with a predetermined threshold in an edge direction detector block 140. If the 2D variance is less than the threshold, then no edge is present at the pixel, and the pixel is considered having “no direction”. If the 2D variance is greater than the threshold, then an edge is present at that pixel, and the direction with the smallest 1D variance is considered as the edge direction of the pixel. Utilizing a filter 160, at “no direction” pixels, a 2D average filter is applied. At a pixel with a detected edge direction, a 1D average filter is applied along the detected direction. By filtering along image edge directions, the directional filter 100 is able to retain most of the image structures while reducing the noise level of the input image.

[0005]
There are two major shortcomings to the directional filtering technique. The first is that the threshold value must be manually tuned and usually it is difficult to select the right value. An improperly selected threshold value will cause either image blurring or insufficient noise reduction. The second shortcoming of the directional filter is that the filter strength is fixed. That means a relatively clean image is processed the same way as a highly noisy image. This causes the relatively clean image to unnecessarily lose some fine structures and be degraded.
BRIEF SUMMARY OF THE INVENTION

[0006]
The present invention addresses the above shortcomings. As such, in one embodiment the present invention provides a global and local statistics controlled noise reduction system wherein the video image noise reduction processing is effectively adaptive to both image local structure and global noise level. And, a noise estimation method according to the present invention provides reliable global noise statistics to the noise reduction system.

[0007]
Such a global and local statistics controlled noise reduction system dynamically/adaptively configures a local filter for processing each image pixel, and processes the pixel with that local filter. The filtering process of the noise reduction system is controlled by both global and local image statistics that are also computed by the system. In one example, the local statistics computed by the system are 1D and 2D local variances, and the global statistics computed by the system is the global noise standard deviation. The local filter configured by the system for each image pixel has different filtering directions and variable strength for different pixels. The direction of the local filter is determined by 1D local variances. The strength of the local filter is computed directly from the local variances and the global noise standard deviation.

[0008]
According to a further aspect of the present invention, the global noise standard deviation is estimated by a noise estimation method. First, the image is divided into overlapping or nonoverlapping blocks, and the mean and the standard deviation of each block are calculated. Then, the smallest standard deviation is found together with the corresponding block mean. After the smallest standard deviation and its corresponding mean have been found, a “saturation checking” process is applied to determine whether the block with the smallest standard deviation has saturated. This determination is based on the relation between the smallest standard deviation and its corresponding mean.

[0009]
If saturation is not detected, the calculated block standard deviations that are within a neighboring interval centered at the smallest standard deviation are averaged, and the average value is taken as the estimated global noise standard deviation. The radius of the neighboring interval depends on the value of the smallest standard deviation.

[0010]
If saturation is detected, first a saturation compensation term is added to the smallest standard deviation to generate a compensated smallest standard deviation. The saturation compensation term is computed from the smallest standard deviation and its corresponding mean. Then, the calculated block standard deviations that are within a neighboring interval centered at the compensated smallest standard deviation are averaged to obtain the estimated global noise standard deviation. The noise standard deviation is used in the noise reduction system.

[0011]
Other objects, features and advantages of the present invention will be apparent from the following specification taken in conjunction with the following drawings.
BRIEF DESCRIPTION OF THE DRAWINGS

[0012]
FIG. 1 shows a block diagram of a prior art directional filter;

[0013]
FIG. 2 shows a functional block diagram of an embodiment of a global and local statistics controlled noise reduction system according to the present invention;

[0014]
FIG. 3 illustrates example directions for computing the 1D variances in the system of FIG. 2;

[0015]
FIG. 4 shows an example curve representing dependency of the filter strength on local variance and global noise standard deviation;

[0016]
FIG. 5 shows a function block diagram of an embodiment of a noise estimation system utilized for the global statistics computing unit of FIG. 2; and

[0017]
FIGS. 6A and 6B show example diagrams illustrating the effect of pixel values driven into saturation (0 or 255) by noise.
DETAILED DESCRIPTION OF THE INVENTION

[0018]
As noted above, in one embodiment the present invention provides a global and local statistics controlled noise reduction system wherein the video image noise reduction processing is effectively adaptive to both image local structure and global noise level. And, a noise estimation method according to the present invention provides reliable global noise statistics to a noise reduction system. The system dynamically configures a local filter for processing each image pixel, and processes the pixel with that local filter. The filtering process is controlled by both global and local image statistics. In one example, the local statistics computed by the system are 1D and 2D local variances, and the global statistics computed by the system is the global noise standard deviation. The dynamically configured local filter has different directions and variable strength for different pixels. The direction of the local filter is determined by 1D local variances. The strength of the local filter is computed directly from the local variances and the global noise standard deviation.

[0019]
The global noise standard deviation is estimated by a preferred noise estimation method that comprises the following steps. First, the image is divided into overlapping or nonoverlapping blocks, and the mean and the standard deviation of each block are calculated. Then, the smallest standard deviation is found together with the corresponding block mean. After the smallest standard deviation and its corresponding mean have been found, a “saturation checking” process is applied to determine whether the block with the smallest standard deviation has saturated. The process is based on the relation between the smallest standard deviation and its corresponding mean.

[0020]
If no saturation is detected, the calculated block standard deviations that are within a neighboring interval centered at the smallest standard deviation are averaged, and the average value is taken as the estimated global noise standard deviation. The radius of the neighboring interval depends on the value of the smallest standard deviation. If saturation is detected, first a saturation compensation term is added to the smallest standard deviation to generate a compensated smallest standard deviation. The saturation compensation term is computed from the smallest standard deviation and its corresponding mean. Then, the calculated block standard deviations that are within a neighboring interval centered at the compensated smallest standard deviation are averaged to obtain the estimated global noise standard deviation (the radius of the neighboring interval depends on the value of the compensated smallest standard deviation). The global noise standard deviation is used in the noise reduction system.

[0021]
An example of the noise reduction system according to the present invention is now described. FIG. 2 shows a functional block diagram of the example global and local statistics controlled noise reduction system 200 according to an embodiment of the present invention. The system 200 comprises a Global Statistics unit 210, a Local Statistics unit 220, a Direction Detector 230, a Filter Generator 240, and a Pixel Filtering unit 250. A digital video input image is first supplied to both the Global Statistics unit 210 and the Local Statistics unit 220. The Global Statistics unit 210 estimates the global noise statistics using said noise estimation method (also described further below). The output of the Global Statistics unit 210 is the global noise standard deviation σ, which is supplied to the Filter Generator 240. The Local Statistics unit 220 computes the 2D local variance within a small window centered at the current pixel and the 1D local variances along four directions within the same window.

[0022]
FIG. 3 illustrates an example twodimensional window 300 including nine image pixels 310, and example directions for computing the 1D local variances for pixel (i,j), where i and j are the indices for pixel row and column, respectively. The designations L_{1}, L_{2}, L_{3 }and L_{4 }denote the horizontal, vertical, diagonal from upper left to lower right, and diagonal from upper right to lower left directions, respectively. Further, the designation σ_{k} ^{2 }denotes the 1D local variance computed for direction L_{k}, wherein k=1, 2, 3 or 4. And, the designation σ_{0} ^{2 }denotes the 2D local variance.

[0023]
In the example of FIG. 3, P(i,j) denotes e.g. the image grayscale value of the pixel at position (i,j), wherein the local variances σ_{k} ^{2 }(k=0, 1, 2, 3, 4) at pixel (i,j) are computed within the 3×3 window 300 by example as follows.

[0024]
If k=0, then σ_{0} ^{2 }is the 2D variance, and is computed by:
${\sigma}_{0}^{2}=\left(\sum _{s=1}^{1}\sum _{t=1}^{1}{\left(P\left(i+s,j+t\right){\mu}_{0}\right)}^{2}\right)/9,$

[0025]
where μ_{0 }is the corresponding 2D mean, defined as:
${\mu}_{0}=\left(\sum _{s=1}^{1}\sum _{t=1}^{1}P\left(i+s,j+t\right)\right)/9.$

[0026]
If k>0, then σ_{k} ^{2 }(k=1, 2, 3, 4) are 1D variances along the direction L_{k }(k=1, 2, 3, 4), and are computed by:
σ_{1} ^{2}=((P(i, j−1)−μ_{1})^{2}+(P(i, j)−μ_{1})^{2}+(P(i, j+1) μ_{1})^{2})3
σ_{2} ^{2}=((P(i−1, j)−μ_{2})^{2}+(P(i, j)−μ_{2})^{2}+(P(i+1, j) μ_{2})^{2})3
σ_{3} ^{2}=((P(i−1, j−1)−μ_{3})^{2}+(P(i, j)−μ_{3})^{2}+(P(i+1, j+1)−μ_{3})^{2})/3;
σ_{4} ^{2}=((P(i−1, j+1)−μ_{4})^{2}+(P(i, j)−μ_{4})^{2}+(P(i+1, j−1)−μ_{4})^{2})/3;

[0027]
where μ_{k }(k=1, 2, 3, 4) are the means along the direction L_{k }(k=1, 2, 3, 4), and are computed by:
μ_{1}=(P(i, j−1)+P(i, j)+P(i, j+1))/3;
μ_{2}=(P(i−1, j)+P(i, j)+P(i+1, j))/3;
μ_{3}=(P(i−1, j−1)+P(i, j)+P(i+1, j+1))/3;
μ_{4}=(P(i−1, j+1)+P(i, j)+P(i+1, j−1))/3;

[0028]
After computing the local variances, the Local Statistics unit 220 (FIG. 2) provides the 1D local variances to the Direction Detector 230 to determine the local edge direction. The Direction Detector 230 then selects the direction that has the smallest 1D variance as the local edge direction, and provides it to the Filter Generator 240. The Local Statistics unit 220 also provides the computed 1D and 2D local variances to the Filter Generator 240, to generate/configure a local filter based on the statistics quantities provided by both the Local Statistics unit 220 and the Global Statistics unit 210.

[0029]
The Filter Generator 240 generates a local filter for the pixel to be filtered. The direction of the local filter is the local edge direction detected by the Direction Detector 230. The strength of the local filter is computed by using the global noise standard deviation σ provided by the Global Statistics unit 210 and the local variances σ_{k} ^{2 }(k=0, 1, 2, 3, 4) provided by the Local Statistics unit 220. For edge direction L_{k }(k=1, 2, 3, 4), the designation α_{k }(k=1, 2, 3, 4) denotes the corresponding filter strength along those directions. Further, the designation α_{0 }denotes the filter strength for nonedge area filtering. While α_{k }(k=1, 2, 3, 4) controls the strength for filtering along the edge direction, α_{0 }controls the strength for nonedge area filtering.

[0030]
The filter strengths α_{k }(k=1, 2, 3, 4) for edge direction L_{k }(k=1, 2, 3, 4) are functions of the global noise standard deviation and the local variance, and are computed in one example as:
α_{k}=min(2σ, max(3σ−σ_{k},0))/(2σ);

[0031]
wherein min(a,b) is the minimal function that returns the smaller one of the two values a and b, and max(a,b) is the maximal function that returns the larger one of the two values a and b.

[0032]
FIG. 4 shows an example curve/plot 400 of the filter strength function α_{k}. When σ_{k }(i.e., the square root of the local variance) is small in comparison with the global noise standard deviation σ (indicating that the local change is caused by noise), the local filter has full strength. When σ_{k }is large in comparison to σ (indicating that the local change along the detected direction is caused by image structure), the local filter has zero strength. In between, the local filter strength continuously varies with σ_{k}. Further, the filter strength α_{k }varies with the global noise standard deviation a (i.e., it increases as the global noise standard deviation increases).

[0033]
The filter strength α_{0 }for nonedge area is computed similarly as following:
α_{0}=min(2σ, max(3σ−σ_{0},0))/(2σ).

[0034]
The curve for α_{0 }is similar to that for α_{k }(k=1, 2, 3, or 4) shown in FIG. 4.

[0035]
Using the detected local edge direction L_{k }(k=1, 2, 3, or 4), the edge direction filter strengths α_{k }(k=1, 2, 3, or 4), and the nonedge area filter strength α_{0 }computed above, the Filter Generator unit 240 (FIG. 2) generates/configures the local filter f_{k }(k=1, 2, 3, or 4) as follows.

[0036]
If the detected direction is L_{1}, then f_{1 }is a 2D local filter for horizontal direction, and is defined as:
${f}_{1}=\frac{1}{9}\left[\begin{array}{ccc}{\alpha}_{0}& {\alpha}_{0}& {\alpha}_{0}\\ {\alpha}_{0}+3{\alpha}_{1}\left(1{\alpha}_{0}\right)& {\alpha}_{0}+3\left(32{\alpha}_{1}\right)\left(1{\alpha}_{0}\right)& {\alpha}_{0}+3{\alpha}_{1}\left(1{\alpha}_{0}\right)\\ {\alpha}_{0}& {\alpha}_{0}& {\alpha}_{0}\end{array}\right];$

[0037]
If the detected direction is L_{2}, then f_{2 }is a 2D local filter for vertical direction, and is defined as:
${f}_{2}=\frac{1}{9}\left[\begin{array}{ccc}{\alpha}_{0}& {\alpha}_{0}+3{\alpha}_{2}\left(1{\alpha}_{0}\right)& {\alpha}_{0}\\ {\alpha}_{0}& {\alpha}_{0}+3\left(32{\alpha}_{2}\right)\left(1{\alpha}_{0}\right)& {\alpha}_{0}\\ {\alpha}_{0}& {\alpha}_{0}+3{\alpha}_{2}\left(1{\alpha}_{0}\right)& {\alpha}_{0}\end{array}\right];$

[0038]
If the detected direction is L_{3}, then f_{3 }is a 2D local filter for the diagonal direction from upper left to lower right, and is defined as:
${f}_{3}=\frac{1}{9}\left[\begin{array}{ccc}{\alpha}_{0}+3{\alpha}_{3}\left(1{\alpha}_{0}\right)& {\alpha}_{0}& {\alpha}_{0}\\ {\alpha}_{0}& {\alpha}_{0}+3\left(32{\alpha}_{3}\right)\left(1{\alpha}_{0}\right)& {\alpha}_{0}\\ {\alpha}_{0}& {\alpha}_{0}& {\alpha}_{0}+3{\alpha}_{3}\left(1{\alpha}_{0}\right)\end{array}\right];$
and

[0039]
If the detected direction is L_{4}, then f_{4 }is a 2D local filter for the diagonal direction from upper right to lower left, and is defined as:
${f}_{4}=\frac{1}{9}\left[\begin{array}{ccc}{\alpha}_{0}& {\alpha}_{0}& {\alpha}_{0}+3{\alpha}_{4}\left(1{\alpha}_{0}\right)\\ {\alpha}_{0}& {\alpha}_{0}+3\left(32{\alpha}_{4}\right)\left(1{\alpha}_{0}\right)& {\alpha}_{0}\\ {\alpha}_{0}+3{\alpha}_{4}\left(1{\alpha}_{0}\right)& {\alpha}_{0}& {\alpha}_{0}\end{array}\right].$

[0040]
The generated local filter f_{k }is supplied to the Pixel Filtering unit 250 (FIG. 2). The pixel is then filtered using the weighted sum of its neighboring pixels within a e.g. 3×3 window with the corresponding filter coefficients as the weights.

[0041]
As noted, the Global Statistics unit 210 estimates the global noise statistics using a preferred noise estimation method. Different methods have been proposed to estimate the noise present in the images, such as those described in the paper by S. I. Olsen: “Noise Variance Estimation in Images: An Evaluation”, Graphical Models and Image Processing, vol. 55, no. 4, pp. 319323, 1993.

[0042]
However, existing methods have not properly considered saturation effects (i.e., pixel values driven into saturation (0 or 255) by the noise, causing inaccurate estimates (in most cases underestimates) of noise). Inaccurate noise estimate can have serious impact on the performance of the noise reduction system that is controlled by global noise statistics.

[0043]
FIG. 5 shows a functional block diagram of an example noise estimation system 500 according to the present invention, for the Global Statistics Computing unit 210 of FIG. 2. The example noise estimation system 500 according to the present invention is capable of handling said saturation effect and providing accurate noise estimates. The input image is first divided into overlapping or nonoverlapping blocks B_{n }of size H×W; (n=1,2, . . . , N); where N is the total number of blocks, and the mean and the standard deviation of each block are computed by the Mean and Standard Deviation Calculator unit 510. The selected block size should not be too small to ensure a robust estimate. Preferably, the block size is e.g. 7×7 or 5×9 pixels (other block sizes can be used). The mean m_{n }and the standard deviation d_{n }of block B_{n }are computed in the unit 510, respectively, as:
${m}_{n}=\frac{\sum _{\left(i,j\right)\in {B}_{n}}P\left(i,j\right)}{H\times W};$ ${d}_{n}=\sqrt{\frac{\sum _{\left(i,j\right)\in {B}_{n}}{\left(P\left(i,j\right){m}_{n}\right)}^{2}}{H\times W}}.$

[0044]
The computed block standard deviations and means are then provided to a Minimal Finder 520. The block standard deviations are also provided to a Selective Averaging unit 530.

[0045]
The Minimal Finder 520 finds the smallest standard deviation, and records the smallest standard deviation and its corresponding block mean as d_{0 }and m_{0}, respectively, wherein the values d_{0 }and m_{0 }are then supplied to a Saturation Checker 540.

[0046]
The Saturation Checker 540 checks whether saturation has occurred in the block with the smallest standard deviation d_{0}. That is, the Saturation Block checker determines/detects if pixel values in the block are driven into saturation (e.g., 0 or 255) due to noise, which may cause inaccurate estimates of image noise. If saturation has occurred, a Saturation Compensator 550 compensates for d_{0}.

[0047]
Examples of saturation detection criteria and the compensation methods are provided below in conjunction with
FIGS. 6A6B which illustrate examples of the effect of pixel values driven into saturation by noise (e.g., pixel value is at a lower limit such as 0 or at an upper limit such as 255).

 For this example an upper limit UL=255, a lower limit LL=0 and a mid value M=128. Such that, if m_{0}<128 and d_{0}>m_{0}−0, then saturation is detected at the lower limit 0 (FIG. 6A). In this case d_{0 }is compensated as {tilde over (d)}_{0}=d_{0}+K·(d_{0}−(m_{0}−0)), wherein {tilde over (d)}_{0 }denotes the compensated smallest standard deviation; If m_{0}≧128 and d_{0}>255−m_{0}, then saturation is detected at the upper limit 255 (FIG. 6B). In this case d_{0 }is compensated as {tilde over (d)}_{0}=d_{0}+K·(d_{0}−(255−m_{0})); Otherwise, no saturation is detected. In this case no compensation is needed for d_{0}, therefore {tilde over (d)}_{0}=d_{0}.

[0049]
In the above expressions for {tilde over (d)}_{0}, the compensation parameter K is empirically determined. Preferably in this example K=5.0 is used. As those skilled in the art will recognize, the saturation detection method can be easily generalized to other situations where the images are represented by different bit values and therefore have different values for UL, LL and M.

[0050]
The compensated smallest standard deviation {tilde over (d)}_{0 }is then supplied to the Selective Averaging unit 530 (FIG. 5) to generate the final estimate of the global noise standard deviation σ, utilized in the system 200 of FIG. 2.

[0051]
The Selective Averaging unit 530 (FIG. 5) first selects those block standard deviations (provided by the Mean and Standard Deviation Calculator 510) that are within a selected range of (e.g., close to) the compensated smallest standard deviation {tilde over (d)}_{0}. In one example, a block standard deviation d_{n }(n=1,2, . . . , N) is considered within a range of {tilde over (d)}_{0 }if d_{n}−{tilde over (d)}_{0}<max({tilde over (d)}_{0},1). The selected block standard deviations are then averaged, and the average value is taken as the final estimate of the global noise standard deviation σ, for use in the system 200 (FIG. 2).

[0052]
While this invention is susceptible of embodiments in many different forms, there are shown in the drawings and will herein be described in detail, preferred embodiments of the invention with the understanding that the present disclosure is to be considered as an exemplification of the principles of the invention and is not intended to limit the broad aspects of the invention to the embodiments illustrated. The aforementioned systems 200 and 500 according to the present invention can be implemented in many ways, such as program instructions for execution by a processor, as logic circuits, as ASIC, as firmware, etc., as is known to those skilled in the art. Therefore, the present invention is not limited to the example embodiments described herein.

[0053]
The present invention has been described in considerable detail with reference to certain preferred versions thereof; however, other versions are possible. Therefore, the spirit and scope of the appended claims should not be limited to the description of the preferred versions contained herein.