JP3143627B2 - Edge-preserving noise reduction method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image noise reduction system for restoring a digital image, and more particularly, to smoothing an input image signal accompanied by deterioration due to noise, thereby quickly and optimally converting the original image signal. The present invention relates to an image noise reduction method to be estimated.

[0002]

2. Description of the Related Art Normally, an input image signal obtained by digitally converting an analog image signal or an image scanning signal by a scanner (scanner) contains noise. Conventionally, in order to remove noise contained in this image, the bright surface of the image is gradually changed and smoothed, and at the same time, sharper luminance change portions such as edges, lines, and contours are saved. Is going. Residual distortion in the image that has been smoothed compared to the noisy original image is identified as an error in the image signal estimation process.

[0003] For example, when a TV camera used for process control in a factory, a portable TV camera for collecting commercial information such as an advertisement, an image camera for simultaneous broadcasting, and the like are often used. Face low lighting conditions. In that case, the signal / noise ratio (S /
N ratio) as a whole, and as a result, the image quality is reduced. In such a case, the image signal is generally observed as a clear point in a displayed image in which electronic noise and noise due to unevenness of the CCD detector are emphasized. This problem also appears in still video cameras. With a still video camera, one snapshot is taken for each scene, in which case the fixed pattern and the random noise are combined so that the noise appearing in the video becomes more clearly visible and the viewer There was a problem of bothering.

[0004] Further, in the case of television cameras used for in-store monitoring, in-chamber monitoring, and the like, the image of which is enhanced is often observed with high noise caused by the low level of the light source. was there.

[0005] Also, image noise is a medical imaging system,
For example, it also appears in an ultrasonic contrast device that shows a state inside a human body in a tomographic manner, a photon counting image system that is a special electro-optical device used under extremely low illumination conditions, and the like.

[0006] In the case of image scanning from a fixed pattern by a CCD scanner detector, the image quality of the obtained image depends on the conditions of the light source and the scanned data (image or text on paper or film).

[0007] There is also noise caused by the non-uniformity of the sensitivity of the CCD scanner detector. And in these cases as well, it is often necessary to reduce noise.

[0008]

As described above, the problem of image restoration for removing noise mixed in a video (image restored for display) is caused by electro-optical scanning contrast, electronic television, graphic display, or the like. Is widely applied.

Incidentally, this noise can be corrected by using a digital image restoration processing scheme. Conventional digital image restoration processing schemes can generally be divided into two types. That is, the first is a spatial smoothing operator using only spatial image information, and the second is a time integration operator for sequentially integrating a digitized sequence of images (digital series) by a finite response time. This extends the effective exposure duration and improves the signal / noise ratio.

The image processing result by the spatial smoothing operator is
Normally, compared to the real image, that is, the original image, residual distortion such as "blur" or "bleeding" occurs more frequently in image contents such as edges, lines, outlines, etc. It becomes an image. However, in the case of a TV image, for example, while there is “movement” in the image, the image is continuously updated, so that the blurring of the image becomes inconspicuous together with the human vision correction ability.

When the image processing by the temporal image integration operator is applied to a still image, since the real image signal is constant but the noise signal is random, it is easy to remove noise, and the restored image is The image quality is sufficiently improved. However, for a moving subject, the effective exposure of the video camera is extended by integrating continuous video signals (video frames), causing inevitable image bleeding. Therefore, it turns out that it is not suitable for image processing of a moving image.

As described above, each method has advantages and disadvantages, and there is no method capable of optimally removing noise from a restored image regardless of a moving image or a still image. A first object of the present invention is to realize an image noise reduction method for smoothing an input image signal accompanied by deterioration due to noise and for quickly and optimally estimating an original image signal.

A second object is that, in an image noise reduction system, each pixel signal of an image is automatically
By selecting the most appropriate signal model, either an edge model, a line or contour model, or a smooth surface model,
Combining spatial smoothing with the noise reduction benefits of time integration.

[0014]

This spatial smoothing method involves a Kalman filter gain factor and employs, for example, a Wiener filter as an estimation filter for predicting and filtering a stationary time series. For example, a median filter is used as a filter for edge-preserving smoothing.

The repetitive spatial smoothing is achieved by repeating the processing using the same type of two-dimensional arithmetic unit having a cascade configuration. The spatial smoothing operator for this purpose is a cyclic two-dimensional operator in two directions, and corresponds to real-time (real-time) processing without an appropriate video line memory for estimating calculation of an image luminance signal value. A suitable pipeline structure for

This operator provides a local signal / noise ratio that is a simple and accurate estimate of the local image statistics for additive noise in the image. This statistical estimate is selectively calculated depending on whether the current pixel (the pixel currently being processed) is an edge pixel, a line or contour pixel, or a smooth luminous surface pixel, thereby selecting an appropriate Kalman gain factor. I do. The Kalman gain factor for estimating the complete range of the signal / noise ratio is cyclically calculated and aggregated into one point to be in a constant state, and is stored in the index table of the direct access memory.

The combination of the cyclic two-dimensional calculation using the pixel line smoothed before the current pixel (the pixel row immediately above the row of pixels where the current pixel is located) is based on the edge real image value and the line or contour An input value that is not confused with a real image value and is efficiently provided to an operator.

In this spatial smoothing method, in order to estimate a luminance value of a current pixel, a smoothed image (for example, a video image, immediately before a frame or a field where a current pixel exists in a video image) is estimated. (3), a cyclic three-dimensional calculation is performed using the digital sequence constituting the frame or field, and related pixel data are combined in the time domain.

As described above, the advantage of the integration method for a still image is realized by applying a minimum cyclic operation to a moving image which tends to cause “bleeding” by a simple spatial smoothing method.

According to the method of the present invention, the image enhancement, encoding and compression, interpolating, and electronic zooming are performed by changing some of the control variables of the operator and the prepared Kalman filter gain factors. Enlargement) can be directly and easily applied to a wide range of image reproduction processing.

[0021]

Next, the structure and operation of an embodiment of the present invention will be described sequentially in detail. FIG. 1 is a conceptual block diagram showing a function arrangement and a processing procedure of a calculation process in which an edge-preserving image noise reduction method of the present invention is executed.

In FIG. 1, in the unit 10, first,
The input analog video signal or image scanning signal is digitized. Further, all means for automatically performing gain level control for optimal digitization of an input signal are included.

The unit 11 is composed of a complete noise reduction calculation processing spatial algorithm, and has a pipeline-type parallel arrangement of operators having the same configuration from the first to the n-th. Therefore, iterative calculations can be performed without reducing data flow throughput and scan rates.

Each unit 11 has a number 7 from number 12 to number 18
It consists of sub units. The sub-unit 12 is composed of a broadly-defined stationary Markov image model described later, that is, a basic one-dimensional Kalman estimator based on piecewise smoothing.

The sub-unit 13 is an intermediate storage with a delay device, that is, a delay line FIFO (First In First Out) for storing an input signal, and the sub-units 14 and 15 are for storing intermediate results of the calculation processing.

The subunit 16 constitutes a two-way mechanism for automatically selecting the appropriate signal mode of the current pixel in response to a set of logic rules, ie, edge signals, fine line or contour signals, or smooth density signals. Then calculate the smoothed result for both scan directions.

The sub-unit 17 has a fixed state look-up table, ie, a full range of expected signal / noise ratio estimates and fixed (assumed and preset) spatial correlation values. It has a pre-calculated Kalman filter gain factor.

The sub-unit 18 comprises a two-dimensional smoothing calculation process which combines the current two-way smoothing result with the previously scanned and smoothed image line. Unit 19 comprises a manual or automatic noise level entered in the form of an image noise variation estimate for the noise reduction scheme. Then, this value is used as a scale factor for the estimated signal variation, and is also used as an input signal for noise measurement of the Kalman gain factor look-up table of the subunit 17.

FIG. 2 shows the Kalman gain factor K (i) and the signal /
FIG. 3 is a characteristic diagram showing a relationship with a noise ratio snr. FIG.
2 shows a basic one-dimensional estimation calculation mechanism and a data flow of the present invention.

In FIG. 3, the image signal containing noise is obtained from an adjacent region modeled as a first (broadly defined) Markov process by means of the mean value, variation, and correlation coefficient related to each region. It is modeled with the possibility as a piecewise steady process. That is, FIG. 3 shows a calculation method of the function “S (i) = S (i−1) + K (i) [X (i) −S (i)]”, where “i” is X (i) is an input pixel value, S (i) is an estimated value having directionality, K
(I) is the Kalman gain factor. This Kalman gain factor is updated for all pixels as a function of the local signal to signal-to-noise ratio (SNR). This signal-to-noise ratio (S
NR) is estimated as a function of a pre-estimated value (predicted estimated value) having directionality of adjacent pixels. In the figure, a unit 301 includes X (i), S (i-1)
And the signal-to-noise ratio SNR as a function of
(I) is calculated. Therefore, unit 301 loads a look-up table into unit 302 along with the K (i) value. Unit 302 receives the output of unit 305 (the difference between X (i) and S (i-1)) and multiplies this value by K (i). Unit 303 is an adder. This adder outputs the output of the unit 30 to the unit 30.
4 are added. Unit 304 is a delay.
The delay receives the estimate S (i) and delays it by one pixel.

In general, image signals are degraded by normally scattered uncorrelated additive noise. The purpose of this configuration is to estimate the input noise-degraded image signal as the most desirable original image signal by smoothing.

In general, at the same time, on the one hand, the edge,
That is to say, while preserving the inherently high amplitude signal content, such as lines and contours, on the other hand, successively smoothing the varying luminance (density) signal. The residual distortion in the smoothed image compared to the original image signal is defined as an estimation error.

The least mean square error (MMSE) cyclic unbiased estimate of the Markov process in a broad sense with non-zero mean is

[0034]

(Equation 1)

Is given by Here, outside 1 is MMSE
Represents a cyclic estimate,

[0036]

[Outside 1]

S (i-1) represents the previous Markov processing output value, ρ _S represents the correlation coefficient of the processing, and μ _S represents the average value of the processing.

The utility of Markov model processing, which is also Gaussian distribution processing, is that Kalman cyclic estimate theory applies directly to the problem in the most convenient sense and in a closed form as a surely valid solution. Image pixel estimates are represented as state estimates by the well-known separate (discontinuous) Kalman filter equation. The system model is
_S (i) = ρ _S s (i-1) + ψ (i-1) which is a Markov model of a sequence of image pixels; ψ to N (0, б _S ²
(1−ρ _S ² )). Where the notation N (μ, б ² )
Represents a decorrelated normally distributed (Gaussian) process. The density measurement of the noise-degraded pixel is represented by x (i), and the measurement model symbolic method in this one-dimensional decomposition is: x (i) = s (i) + v (i); v〜N (0, σ _n ²⁾ ).

At each sample point i in the estimation process,
State estimation represented by

[0040]

[Outside 2]

The value extrapolation equation is

[0042]

(Equation 2)

Is given by Where outside 3
Is the state estimate in the previous processed sample

[0044]

[Outside 3]

Represents an update. Error covariance extrapolation at the sampling point i is P _- represented by (i), P _- is given by ^{_{(i) = ρ S 2 P}} + (i-1) + σ S 2 (1-ρ S 2). Here, P ₊ (i−1) represents the error covariance update in the previous processed sample. The update of the state estimate at pixel i is

[0046]

(Equation 3)

Is given by Here, K (i) represents the Kalman gain factor of the pixel i, K (i) = P - (i) / (P - (i) + ρ n 2) = (ρ S 2 P + (i-1) + Σ _S ² (1−ρ _S ² )) / (ρ _S ² P ₊ (i−1) + σ _S ² (1−ρ _S ² ) + σ _n ² ). The error covariance update _{is, P + (i) = [} 1-K (i)] P - is given by _{(i) = K (i)} σ n 2.

From the above symbolic formula, the Kalman gain factor K
(i) is, depending on K (i-1), K (i) = (ρ S 2 K (i-1) σ n 2 + σ S 2 (1-ρ S 2)) / (ρ S 2 K (i-1) σ _n ² + σ _S ² (1−ρ _S ² ) + σ _n ² ). Then, with the signal / noise ratio given by snr = σ _S ² / σ _n ² , the following simple expression for K (i) is obtained.

K (i) = (ρ _S ² K (i−1) + snr (1−ρ _S ² )) / (ρ _S ² K (i−1) + snr (1−ρ _S ² ) +1) Estimation processing In order to calculate K (i) for a given point, a piecewise (local) mean, variance, and correlation estimate must be calculated. Given these statistics, a constant state value for the Kalman gain factor can be cyclically pre-computed as a function of the expected values of snr and ρ _S and stored in a look-up table. . The cyclic operation is initialized by setting an initial value K (0) to the following equation: K (0) = snr / (snr + 1).

An additional simplification is introduced into the estimation process. The value for ρ _S is previously set to one constant. And the noise fluctuation σ _n ² of the input signal is well known,
It estimates snr piecewise. snr is related to the degree of uncertainty in the partition estimation. S for each pixel
The self-heuristic approximate estimate of nr is given by the difference equation

[0051]

(Equation 4)

[0052] Further compound sn
The r estimate is calculated using a window that fits near the current pixel location. The section average μ _S is expected to correct between adjacent regions. To avoid iterative calculation of the segmented area average, update the state estimation update to

[0053]

[Outside 4]

[0054]

[0055]

(Equation 5)

The Kalman gain factor K (i) in this equation is

[0057]

(Equation 6)

With snr given by The simplification by this correction prevents the estimation process from flowing from the region average value. The one-dimensional Kalman estimation filter (subunit 12 in FIG. 1) is a basic calculation configuration of the method of the present invention, and is as shown in FIG.

Regarding the above-described one-dimensionally approximated Kalman estimation procedure, two points should be noted. First, the fitting Kalman gain factor is compatible with the intuition where the Kalman gain factor is proportional to the degree of uncertainty in the estimation and inversely proportional to the measurement noise. It is. This makes small changes in the current estimation when the measurement noise is large compared to the estimation uncertainty (low snr).

On the other hand, small measurement noise and large uncertainties (high snr) in state estimation suggest that the current sample measurement contains information to consider. That is, when there is a high degree of uncertainty in the estimation, and at the same time, a small degree of measurement noise, the current measure is considered to be highly reliable and thus contains information about possible changes in signal amplitude. Can be Therefore, the difference between the actual and the predicted measurement is used as a basis for a strong correction to the current estimate.

The second point relates to the main disadvantage of the one-dimensional calculation method. That is, the degree of the noise signal stored in the image as the calculation result increases with the amplitude of the noise. That is, the intensity noise modulation and fluctuations remain unchanged.

However, this problem is solved by the present invention.
It is solved by combining with the two-way processing method. There, the Kalman gain factor is selected using a two-way estimation corresponding to a section-selective theory of signal enhancement between edge signals, line or contour signals, or smooth density signals. FIG.
FIG. 3 is a diagram illustrating a two-direction calculation method for calculating an improved SNR (i) in two directions as a function of a value estimated with two directions in advance. In the figure, for example, a dimension is defined as a horizontal dimension, a vertical dimension, a space diagonal dimension, or a time dimension. Here, the dimension refers to the arrangement direction of pixels to be calculated. For convenience of explanation, it is assumed here that the first dimension is horizontal. However,
The same processing as described below can be repeated by a method of circulating through other dimensions from the second dimension on down. When the first dimension is horizontal as described above in the figure, the first direction is from left to right, and all pixels in the horizontal dimension are:
In the example shown in the first direction indicated by the arrow 409, pixels 405, 406, 407, 408,... Are estimated up to the end of the arrangement of pixels along the horizontal dimension. That is, the direction referred to here refers to a direction in which calculations are sequentially performed. .., 403,..., And 410, starting from the right end of the pixel array, from the right to the left in the figure, Then, the same processing as described above is performed through the pixel 405 at the left end of the dimension. It should be noted here that the other dimensions also have two directions and can be mutually opposite. In the example of FIG. 4, the processing may be performed first in the direction indicated by the arrow 410, and then performed in the direction indicated by the arrow 409. The important thing is that the two directions are opposite to each other. In the figure, a unit 411 includes an absolute estimator S1, that is, S _LR (i) −S
The absolute value of the difference of _RL (i + 1) is calculated. Unit 412
Is the absolute estimator S2, that is, S _LR (i−1) −S _RL (i
The absolute value of the difference of +1) is calculated. And the unit 41
7 executes a logical operation of a comparison function using S1 and S2 to calculate a two-way SNR (i) value for pixel i. And here, the two-way SNR (i) values are used to recalculate the directional estimates described above for the two directions. What should be noted here.
What is achieved here is that when S1> S2, we estimate a smooth luminance signal (also referred to in this description as a smooth density signal in the same sense) and a new bi-directional SNR (i ) Use the smaller value of S2 in the values. Therefore, unit 302 is more stable,
A stronger, smoother gain factor K (i) can be used. Then, the result of the sharp part removal is obtained here. That is, this removes sharp noise from the image. In other situations, ie when S2> S1, it is assumed that an edge signal is present, in which case the value of S1 in the new two-way SNR (i) value of unit 301 is used.
Therefore, the unit 302 can use a more stable and appropriate smoothing gain variable K (i). And
Here, the result of preserving the sharpness of the edge is obtained. The processing shown in FIG. 4 can be realized by using a line buffer for one line as a pixel memory for processing.

As shown in FIG. 1, the two-way smoothing mechanism (sub-unit 16) is calculated based on the one-dimensional smoothing process, the input signal of the temporary storage area, and the smoothing result in the related delay line FIFO group. Due to the high sensitivity of the snr estimate to image noise, the two-way smoothing logic uses the piecewise Kalman gain snr
In order to estimate a more robust measure for each pixel position (see FIG. 4), a left-to-right smoothing estimator
(Hereafter, ⌒s → (i)

[0064]

[Outside 5]

) And a right-to-left smoothing estimator ６ 6 (hereinafter denoted as ⌒s ← (i))

[0066]

[Outside 6]

The two one-dimensional smoothing results are used to combine the two. Assuming that the one-dimensional smoothed result of the complete left-to-right smoothing estimator ⌒s → (i) has been pre-calculated and stored in one FIFO 13, the two-way smoother starts. And the following equation

[0068]

(Equation 7)

[0069]

(Equation 8)

By comparing the two absolute differences s ₁ (i) and s ₂ (i) given by, the smoothed result of the right-to-left smoothed estimated amount ⌒s ← (i) at each pixel is calculated. . If s ₁ (i)> s ₂ (i), the smooth density signal is identified and the associated s
The nr measure is given by snr = s ₂ ² / σ _n ² . The snr measure then selects the Kalman gain factor K (i) and selects the selected Kalman gain factor K
(i) is

[0071]

(Equation 9)

Accordingly, the right to left smoothing estimation amount 推定 s ←
Used to calculate (i). Additionally the same
The nr measure is:

[0073]

(Equation 10)

Is used to correct the value of the left-to-right smoothing estimation amount ⌒s → (i) given by Alternatively, if s ₁ (i) <s ₂ (i), the edge signal is hypothesized and the right-to-left smoothing estimator ⌒s ← (i) is calculated as You. That is, the expression

[0075]

[Equation 11]

According to the above. Snr measure related to this is given by ^{_{snr = s 1 2 / σ n}} 2. Now, using the right-to-left smoothed estimator ⌒s ← (i) smoothed for the current pixel, the snr measure for the left-to-right smoothed estimator ⌒s → (i) is Re-estimate according to the formula.

[0077]

(Equation 12)

And the snr measure related to this is

[0079]

(Equation 13)

Is as follows. In this way, the bidirectional smoother allows the directional smoothing estimator of the right-to-left estimator ⌒s ← (i) and the left-to-right estimator ⌒s → (i) at each pixel to be a Kalman gain factor. Is calculated from the most likely pixel for, using an estimate of that pixel.

The main disadvantage of having a two-way smoother is that
It is often the case that image thin lines and contours, which are essential image information, are smoothed. The present invention compensates for this drawback by combining a two-dimensional scheme that utilizes the previous pixel line smoothed pixels in estimating the current pixel. FIG.
FIG. 4 is a diagram showing a preceding pixel line and a current pixel line.
Pixels 503, 504, and 505 are connected to the preceding pixel line 501.
Shows three pixels according to the pixel processing, S _1-1 (i-1) as the predictive estimated pixel values of the two-dimensional direction of the adjacent pixels in the adjacent one line, S _1-1 (I), _S1-1
It is indicated by the value of (i + 1). The preceding pixel line is a pixel line above the current pixel line when the pixel line advances downward from the top of the figure, and the current pixel line when the pixel line advances upward from the bottom. This is the pixel line below the line. Also, pixels 506, 507, and 508 in the figure represent three pixels involved in processing of the current pixel line 502. Pixel 506 is S
_LR (i-1), which is indicated by estimated pixel values in two directions of the current line. The pixel 508 is S _RL (i + 1), which is also indicated by estimated pixel values in the two directions of the current line.
Pixel 507 is then indicated by the current estimated pixel value S ₁ (i), which is a value equal to the weighted sum of the aforementioned pixel and all the predicted estimates of the input pixel value X (i).
Note that the vertical dimension described in FIG. 4 indicates the arrangement direction of the pixels 507 and 504 in FIG. 5, and the spatial diagonal dimension indicates the arrangement direction of the pixels 507 and 505 or the pixels 507 and 503 in FIG. I have. And the time dimension is
Indicates the relationship between the line 502 and the line 501. The pixel value S _1-1 (i-
1), _S1-1 (i) and _S1-1 (i + 1) are respectively
The pixel values S _{l, i-1} (i), S _{1, i} (i), S of the current pixel
_{1, i + 1} (i) is used for estimating with directionality. These three intermediate results are then summed with the current line direction estimates S _LR (i) and S _RL (i) to obtain the following S ₁ , (i) results estimated in the two-dimensional direction. Averaged. S ₁ , (i) = 1/5 (S _l , _i−1 (i) + S ₁ , _i (i) + S ₁ , _{i + 1} (i) + S _LR (i) + S _RL (i)) It should be noted that the above two-dimensional estimation is applied not only from top to bottom but also from bottom to top, as well as to the progression of the pixel lines in order to achieve vertical adjustment. Accordingly, it is repeatedly applied in both directions. Note that the processing shown in FIG. 5 can be realized using a line buffer for two lines as a pixel memory for processing. To explain this further, the currently processed out-of-pixel
Of the previous line adjacent to

[0082]

[Outside 7]

A one-dimensional Kalman smoothing operator is applied to each of the two-dimensionally smoothed pixels using a gain factor index value given by the following difference equation.

[0084]

[Equation 14]

This extension is easily shown to preserve the image lines in the smoothed output image. Thus, five intermediate smoothing results are obtained. That is, two two-way estimates (⌒s → (i), ⌒s ← (i)) and three one-dimensional values obtained from the previous out-of-line 8

[0086]

[Outside 8]

This is an estimated value. These five estimates are then averaged to equal weights, resulting in a two-dimensional smooth pixel estimate.

[0088]

(Equation 15)

Is obtained. The two-dimensional smoothing method is shown in FIG.
It can be easily repeated by using a similar computational mechanism as shown in. That is, the output signal smoothed in the first iteration is used as an input to the second iteration, and is sequentially repeated.

This pipeline structure does not affect the data processing capacity (slope rate) of the entire system. Also, this structure is a convenient design solution for performing various estimation processes that require the sum of the differences of the iterative operations.

Only the required control variables are changed according to the corresponding iterative scheme, resulting in a noise-smooth estimate of the input signal. Also, it is calculated in advance from a smoothing simulation based on the smoothing and noise density signals.

Each repetition is performed directionally and variably in another direction. For example, the first iteration is top to bottom, the second is bottom to top, the third is left to right, and the fourth is right to left. When performing such an iteration, the image resulting from the first iteration is scanned in the second iteration. The resulting image is scanned in the third iteration. The resulting image of the third iteration is scanned in the fourth iteration. Thus, each iteration uses the image obtained as a result of the immediately preceding iteration. Therefore, each iteration occurs after the completion of the previous iteration.

Each iteration performed in a directional manner is based on the premise that the original image including the original noise and the smoothed image obtained from each iteration are stored in the digital image RAM. Therefore, line-sequential scanning in both the horizontal direction and the vertical direction is possible, and any digital image scanning shape can be taken.

The above scheme uses one image RAM buffer for directional repetition for still images. On the other hand, for a moving image, a pipeline structure using four image RAM buffers is required. In any case, between the input original image including noise and the output image after the smoothing processing, as a result, four scanning execution periods (one scanning is one frame or one field line sequential scanning) are performed. Although a delay occurs, it does not become a problematic delay in practical use.

The above two-dimensional smoothing method can be applied to a color image. There, for example, in a red, green, and blue (RGB) image, the image for each color is individually smoothed using the same calculation method. As a result, a noise-reduced RGB color image is obtained.

The two-dimensional smoothing method uses a similar combination of Kalman estimations, for example, applied to the case of a digital sequence of a moving image, so that a previously smoothed image (frame or field) can be used for the current pixel. Can be estimated, and smoothing can be extended to the time domain, whereby a three-dimensional region consisting of spatiotemporal space can be estimated (see FIG. 6).

FIG. 6 shows a state in which the schematic diagram of FIG. 5 is expanded in a three-dimensional direction. Pixels 604 to 60 in FIG.
Reference numeral 9 denotes pixels 503 to 508 in FIG. Then, the three pixels which are hidden under the pixels 601 to 603 in FIG.
The pixel of the preceding image frame. Accordingly, in estimating the current pixel, in addition to including the predicted estimated pixel value adjacent to the current line and the preceding line described above, the estimation of the current pixel is performed within the preceding image frame spatially (three-dimensionally) adjacent to the current image frame. Is used to estimate the current pixel. This three-dimensional method is extremely effective for iterative calculation of the time axis in a stationary sequence, and is also effective for area estimation in which the motion of an image affects the result of the iterative calculation and degrades the image quality. The processing shown in FIG. 6 can be easily realized by using a frame buffer for one frame as the processing pixel memory and a line buffer for two lines used in FIG. In this case, the spatial diagonal dimension is the direction in which the pixels 608 and 603 or the pixels 608 and 601 are arranged.

[0098]

As described above, the optimum pixel estimation amount can be easily calculated and output by the operator having the same configuration regardless of whether the image is a still image or a moving image. It is applicable to a wide range of applications such as electro-optical scanning contrast, electronic television, or graphic display.

That is, T used for process control in a factory, etc.
If applied to image reproduction obtained from a V camera, a portable TV camera for collecting commercial information such as advertisements, and an image camera for broadcast, etc., the illumination is low and the signal of the image signal is low.
A system that can obtain good image quality even when the noise ratio is low can be constructed.

This is similarly effective when used for image reproduction by a television camera used for in-store monitoring, in-chamber monitoring, and the like. Further, if the present invention is applied to the reproduction of an image obtained from a still video camera, an ultrasonic scanner, a CCD scanner, or the like, a high-quality image from which noise has been removed can be reproduced. Even in the reproduction of a moving image such as a video image, an image without blurring in the contour can be obtained.

[Brief description of the drawings]

FIG. 1 is a conceptual block diagram showing a function arrangement and a processing procedure of a calculation process in which an edge-preserving image noise reduction method of the present invention is executed.

FIG. 2 Kalman gain factor K (i) and signal / noise ratio snr
FIG. 4 is a characteristic diagram showing a relationship between

FIG. 3 is a diagram showing a basic one-dimensional estimation calculation mechanism and a data flow of the present invention.

FIG. 4 illustrates the data flow, logic, and calculations of the one-dimensional estimation of the present invention combined in two directions.

FIG. 5 illustrates the data flow, logic, and calculations of the two-way estimation of the present invention combined in two dimensions.

FIG. 6 is a diagram showing data flow, logic, and calculation of three-dimensional estimation in which two-dimensional estimation according to the present invention is combined with space-time.

──────────────────────────────────────────────────続 き Continuation of front page (56) References JP-A-60-235528 (JP, A) JP-A-58-203745 (JP, A) JP-A-60-167574 (JP, A) JP-A 57-203 166663 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G06T 5/00 G06T 5/20

Claims (6)

(57) [Claims] 1. A method applied to image smoothing for removing sharp parts and preserving edges, comprising: for each individual pixel along a first dimension, a pixel group of a received image pixel group; Calculating a first sequence of estimated pixel values of a group of pixels of the received image defined along the first direction that are continuous in a first direction along one dimension; A second group of estimated pixel values of a group of pixels of the received image defined in the second direction that are continuous along a first dimension of the group in a second direction that is opposite to the first direction and defined along the second direction; Calculating a sequence of the estimated pixel values in each of the two directions to calculate an improved estimated pixel value for each pixel along the first dimension. Contiguous contiguous and improved to calculate the improved estimated pixel value for each pixel A weighted estimated pixel value. 2. The step of continuously calculating comprises: calculating a difference value between a received image pixel in a first sequence and an adjacent estimated pixel value preceding the received image pixel in a first sequence; Using the received image pixels and at least one preceding estimated pixel value in the first sequence to estimate a signal-to-noise ratio; and a signal according to a Kalman gain variable corresponding to the signal-to-noise ratio in a first direction. Generating an adjusted difference value by calculating a noise ratio; and updating the adjacent successive estimated pixel values to calculate an estimated pixel value of an individual pixel in the sequence. Using the difference value. 3. The step of calculating continuously comprises calculating a difference value between a directed and estimated neighboring pixel value and a function of the difference value in the first and second sequences. Generating an adjusted difference value to reflect the improved signal-to-noise ratio for each individual pixel in each direction; and improved estimation at each individual pixel in each direction. Using the adjusted signal-to-noise ratio to bring the pixel values up-to-date. 4. The step of weighting comprises: calculating the improved orientated estimated pixel value in a first dimension to calculate an estimated pixel value of an individual pixel in a two-dimensional direction; 2. The method according to claim 1, further comprising the step of weighting the prediction result of the adjacent pixel. 5. The step of repeating the process for at least a second dimension of the received image pixels, whereby the second dimension constitutes a spatial dimension comprising horizontal, vertical, spatial diagonal, or time. And updating the image noise for each iterative step to reflect the reduced noise level of the preceding step.
The method according to 2, 3 or 4. 6. The method according to claim 1, wherein the method is applied to a grayscale or color image.