TWI689890B - Noise equalization method and de-noise method - Google Patents

Abstract

The present disclosure provides a noise equalization method and a de-noise method. The noise equalization method simulates degrees of noise magnitudes corresponding to different signal magnitudes by a smooth curve, and implements noise equalization corresponding to different signal magnitudes by an equalization curve, such that the noise magnitudes corresponding to different signal magnitudes become the same or close. In other words, the signal-dependent noise (SDN) is converted to the signal-independent noise (SIN), and SDN characteristic of the real noise-free signal can be estimated more precisely based on noisy signals. In addition, using the equalization curve to perform the de-noise method can lower the computational complexity and calculate better noise parameters, thereby improving noise removal for each input pixel value in an input image and generating an output image with lower noise.

Description

Noise equalization method and noise removal method

This case provides a noise equalization method and a noise removal method, and particularly relates to a noise equalization method and a noise removal method based on the pixel values after the noise equalization.

Electronic devices (such as smart phones, cameras, cameras, etc.) generate complex noise during the process of capturing images. This kind of noise is not a simple additive Gaussian White Noise (AWGN), but a signal-related noise (Signal-Dependent Noise, SDN), that is, its noise size is related to the signal strength . Therefore, different pixel values will have noise with different intensities.

For example, FIG. 1 shows a schematic diagram of standard deviations of multiple regions of an input image. The input image Fr0 has multiple image blocks EN1 and EN2. Both image blocks EN1 and EN2 have multiple pixel values, and the overall pixel value of image block EN1 is lower than the overall pixel value of image block EN2. As shown in FIG. 1, due to the SDN phenomenon in the input image, the electronic device will calculate the regional standard deviation Std of the pixel values in the image block EN1 to 1.15, and calculate the pixel values in the image block EN2 The regional standard deviation of Std is 2.95, which means that the same input image will have different intensity of noise.

Common Noise Estimation is designed for Gaussian white noise and can be roughly divided into two types. One is real-time noise estimation, the other One is offline (Offline) noise estimation.

The real-time noise estimation is to estimate the noise based on the image converted by the high-pass filter. However, the converted image only exhibits noise characteristics in areas with small variances (ie, smooth areas in the image), and the real-time noise estimation is designed for Gaussian white noise and has a high-pass filter The reference range of the image will also affect the accuracy of noise estimation, so the conventional real-time noise estimation cannot estimate the SDN characteristics close to the real signal.

Offline noise estimation is to pre-establish an SDN module to calculate the distribution characteristics of noise under different pixel values. Then the input image with noise is generated according to the SDN module to generate an output image with no noise. However, the conventional SDN module uses a linear method to describe the distribution characteristics of the real SDN, but the distribution characteristics of the real SDN at different pixel values are nonlinear. In addition, the SDN module is used to simulate the different real signal strengths (Noise -Free Signal), the distribution characteristics of the noise size, but when actually using the SDN module, you can only use the SDN module to query based on the pixel value (Noisy Signal) affected by the noise, and you cannot get the true signal of the pixel (Noise-Free Signal). Therefore, the conventional offline noise estimation also cannot accurately estimate the SDN characteristics of real signals from pixels affected by noise.

If the electronic device cannot estimate the SDN characteristics of the real signal, it will not be able to estimate better noise parameters (such as standard deviation and variance), which will affect the noise reduction effect of the input image (Denoise). Therefore, if the electronic device can estimate the characteristics close to the real SDN, the noise removal effect can be improved, thereby improving the input image quality.

In order to improve the noise removal effect, and thus improve the input image quality, this case provides a noise equalization method, suitable for an electronic device, used to equalize the noise at different values. The noise equalization method includes: obtaining a plurality of representative values corresponding to a plurality of representative values in a numerical interval and the representative values; according to the representative values and The corresponding representative standard deviations calculate an intermediate standard deviation corresponding to each intermediate value between the representative pixel values, and the representative pixel values, the representative standard deviations, the intermediate values and the intermediate standards The difference is mapped to multiple values and multiple standard deviations in a smooth curve; and an equalization curve is calculated according to each value in the smooth curve and the corresponding standard deviation to correspond each value to an equalization Value. In this noise equalization method, a current value of these values plus the corresponding standard deviation is an adjusted value, and the equalization curve satisfies the equalized value corresponding to the adjusted value minus the equalized value corresponding to the current value The latter value is 1 unit.

This case provides a noise removal method suitable for an electronic device. Such a noise removal method includes: receiving an input image; extracting a current position block from the input image, wherein the current position block corresponds to one or more input pixel values; and extracting multiple blocks adjacent to the current position block Adjacent blocks; mapping each input pixel value in the current position block and the adjacent blocks through an equalization curve to obtain the equalized value corresponding to each input pixel value, wherein the equalization curve includes multiple Values and corresponding multiple equalized values, a current value of these values plus a corresponding standard deviation is an adjusted value, and the equalized curve satisfies the equalized value corresponding to the adjusted value minus the current value The equalized value of is 1 unit; based on these equalized values, a degree of difference between each neighboring block and the current location block is calculated; according to each degree of difference, a weight value corresponding to the neighboring block is determined; According to the weight values, a main pixel value in the corresponding neighboring blocks is weighted and averaged to generate a modified pixel value.

This case provides a noise removal method suitable for an electronic device. Such a noise removal method includes: receiving an input image; capturing a current position block in the input image, wherein the current position block corresponds to a plurality of input pixel values; and determining a main pixel value according to the input pixel values; The input pixel values are mapped through an equalization curve to obtain corresponding multiple equalized values, wherein the equalized curve includes multiple values and corresponding multiple equalized values. The value plus a corresponding standard deviation is an adjustment value, and the equalization curve satisfies the adjustment value The equalized value minus the equalized value corresponding to the current value is 1 unit; the equalized values are averaged to produce a low-frequency component value; the equalized values and low-frequency components corresponding to the input pixel values are calculated A variation between the values; a high-frequency component value of the main pixel value in the current location block is calculated from the low-frequency component value; a high-frequency component value of the high-frequency component value of the main pixel value is calculated from the variation number; and the low frequency The component value is added to the high-frequency component value of the high-frequency ratio to generate an output pixel value corresponding to the main pixel value.

Fr0‧‧‧ input image

EN1‧‧‧Image block

EN2‧‧‧Image block

ZN1‧‧‧Image block

ZN2‧‧‧Image block

100‧‧‧Electronic device

110‧‧‧Image capture device

120‧‧‧Image processor

Fr1‧‧‧ input image

P0-Pn‧‧‧Input pixel value

P0’-Pn’‧‧‧ output pixel value

Re0, Re1, Re2, Re3, Re4, Re5, Re6, Re7, Re8, Re9, Re10, Re11‧‧‧ Grayscale block

Fd0‧‧‧Target block

I0, I1, I2, I3, I4, I5, I6, I7, I8, I9, I10, I11‧‧‧ represent pixel values

σ0, σ1, σ2, σ3, σ4, σ5, σ6, σ7, σ8, σ9, σ10, σ11‧‧‧ represents the standard deviation

NTC‧‧‧ test image

Crv1‧‧‧Smooth curve

Crv2‧‧‧Equalization curve

Fr1’‧‧‧ input image after equalization curve mapping

S310, S320, S330 ‧‧‧ steps

S410, S420, S430 ‧‧‧ steps

S460, S470, S480, S490

S910~S980‧‧‧Step

Pch‧‧‧Current location block

Qch1, Qch2 ‧‧‧ neighboring blocks

Fr‧‧‧ input image

P0, P17, P18, P19, P25, P26, P27, P28, P33, P34, P35, P36, P37, P42, P43, P44, P45, P51, P52, P53, P63

S1110, S1120, S1130, S1140, S1150, S1160, S1170, S1180

FIG. 1 is a schematic diagram of standard deviations of multiple regions of an input image.

FIG. 2 is a schematic diagram of an electronic device according to an embodiment of this case.

3 is a flowchart of a noise equalization method according to an embodiment of the present invention.

4A is a detailed flowchart of step S310 according to an embodiment of the present invention.

FIG. 4B is a schematic diagram of a test image according to an embodiment of the present case.

4C is a detailed flowchart of step S310 according to another embodiment of the present case.

FIG. 5 illustrates the representative pixel value and the representative standard deviation according to an embodiment of the present invention corresponding to the pixel value and the standard deviation in the dynamic range.

6 is a schematic diagram of a smooth curve drawn according to an embodiment of the present case.

7 is a schematic diagram of an equalization curve according to an embodiment of the present invention.

8 is a schematic diagram of signal distribution of an input image according to an embodiment of the present invention after being mapped by an equalization curve.

9 is a flowchart of a noise removal method according to an embodiment of the present case.

FIG. 10 is a schematic diagram of the current location block and adjacent blocks of the input image according to an embodiment of the present invention.

11 is a flowchart of a noise removal method according to another embodiment of the present invention.

In the following, the present case will be described in detail by illustrating various exemplary embodiments of the case. However, the concept of the present case can be embodied in different forms and should not be interpreted as being limited to the exemplary embodiments set forth herein. In addition, the same reference numerals in the drawings may indicate similar elements.

The noise equalization method provided by the embodiment of the present invention uses a smooth curve to simulate the nonlinear relationship of the real SDN under different pixel values. Then, through an equalization curve (consisting of multiple pixel values and corresponding equalized values), the noise under different pixel values is converted to the same or similar values, that is, the equalization curve satisfies a current pixel value to move one The distance of the standard deviation corresponds to the value shifted by one unit after equalization. In this way, the noise equalization method can convert signal-related noise (Signal-Dependent Noise, SDN) into signal-independent noise (Signal-Dependent Noise, SIN), and then from pixels affected by noise ( noisy signal) more accurately estimates the SDN characteristics of a noise-free signal.

In addition, the noise removal method provided by the embodiment of the present invention uses the equalization curve calculated by the above noise equalization method to obtain the SIN of each input pixel value in an input image to simplify the noise removal method and calculate a better Noise parameters (such as parameters such as standard deviation and variance), which in turn produces a lower noise output image.

First, FIG. 2 is a schematic diagram of an electronic device 100 according to an embodiment of the present case. As shown in FIG. 2, in the noise equalization method, the electronic device 100 uses the mapping of an equalization curve (composed of multiple pixel values and corresponding equalized values) to apply noise under different signal intensities Equalization, and then convert the noise related to the signal into noise independent of the signal. In the noise removal method, the electronic device 100 will obtain the equalized value corresponding to each input pixel value P0-Pn in an input image Fr1 according to the above equalization curve to generate a lower noise output pixel value P0 '-Pn'. In this embodiment, the electronic device 100 may be a smart phone, a camera, a video camera, a tablet computer, a notebook computer, or other electronic devices with image capturing functions, which is not limited in this case.

<An embodiment of a noise equalization method>

Referring to FIG. 2, the electronic device 100 includes an image capturer 110 and an image processor 120. The image processor 120 is coupled to the image capturer 110 and is used to perform the following steps of the noise equalization method to generate a desired equalization curve (consisting of multiple pixel values and corresponding equalized values). Referring also to FIG. 3, which shows a flowchart of a noise equalization method according to an embodiment of the present case. First, the image processor 120 of the electronic device 100 will obtain a plurality of representative pixel values in a pixel value interval and a representative standard deviation corresponding to each representative pixel value (step S310).

In this embodiment, the pixel value interval can be designed according to the dynamic range of the pixel value. For example, the dynamic range of pixel values is 0-255, which represents the possible values of pixel values 0-255. 0 represents the darkest and 255 represents the brightest. Therefore, the pixel value interval will be set to 0-255. The electronic device 100 then selects a plurality of suitable values from the pixel value interval as representative pixel values (for example, selects 5 representative pixel values with the values of 0, 50, 150, 100, and 255, respectively) and calculates the corresponding Stands for standard deviation.

Furthermore, multiple representative pixel values and corresponding representative standard deviations can be obtained through a test image. As shown in FIGS. 4A and 4B, the image processor 120 first receives a test image NTC (here, a grayscale image is used as an example, but this case is not limited to this), and the test image NTC has a plurality of different grayscale blocks Re0, Re1, Re2, Re3, Re4, Re5, Re6, Re7, Re8, Re9, Re10, Re11 (step S410), the color corresponding to one gray-scale block is white, and the color corresponding to the other gray-scale block is black, And the other gray scale blocks correspond to the colors between white and black. In this embodiment, the color corresponding to the gray-scale block Re0 is white, the color corresponding to the gray-scale block Re11 is black, and the other gray-scale blocks Re1-Re10 correspond to the colors between white and black.

In some embodiments, after receiving the test image (step S410), the image processor 120 will capture multiple pixel values in each grayscale block (step S420). Between steps S410 and S420, if the test image NTC has lens vignetting (shading) problem, the image processor 120 first performs shading compensation on the test image NTC, and then executes step S420. If the test image NTC has no lens vignetting problem, the image processor 120 will directly execute step S420.

In some embodiments, the edges of the gray-scale blocks Re0-Re11 are easily affected by noise to form jagged blocks. In order to obtain more stable multiple pixel values in the gray-scale blocks Re0-Re11 (such as pixel values with similar values in the corresponding gray-scale block, in other words, more flat areas in the corresponding gray-scale block are obtained Pixel values), the image processor 120 may also shrink a predetermined distance (for example, a distance of 2 pixels) in each gray-scale block Re0-Re11 to form a target block, and then separately at each target Retrieve pixel values within the block. Taking the gray-scale block Re0 as an example for illustration, the image processor 120 extracts each pixel value in a target block Fd0 smaller than the gray-scale block Re0 to serve as more stable pixels in the gray-scale block Re0 value. Of course, the image processor 120 can also capture each pixel value or pixel value with a similar value in each gray-scale block Re0-Re11, which is not limited in this case.

After retrieving a plurality of pixel values in each gray-scale block Re0-Re11 (step S420), the image processor 120 will calculate a correspondence according to the retrieved pixel values in each gray-level block Re0-Re11 The representative pixel value and the representative standard deviation of the gray scale block (step S430). Taking the gray-scale block Re0 as an example for illustration, if the image processor 120 captures a total of 81 pixel values, the image processor 120 averages 81 pixel values to generate an average pixel value as the gray-scale block Re0 It represents the pixel value I0, and calculates a standard deviation of 81 pixel values according to the above average pixel value as the representative standard deviation σ0 of the gray scale block Re0. The representative pixel values and standard deviations corresponding to the gray-scale blocks Re0-Re11 are organized in the following table <1>.

In other embodiments of step S310 in FIG. 3, the image processor 120 can also obtain multiple representative pixel values and corresponding representative standard deviations through multiple test images NTC. As shown in FIG. 4C, the image processor 120 will first receive a test image at multiple time points, and the test image has multiple gray-scale blocks (step S460). For example, the image processor 120 captures the test image NTC at 80 different time points, and the test image NTC has a plurality of different gray-scale blocks Re0-Re11.

Next, the image processor 120 will capture at least one pixel value in each grayscale block of each test image (step S470). The implementation of the image processor 120 for extracting multiple pixel values in the gray-scale block can be generally derived from step S420 of FIG. 4A, and will not be described here. In addition, since there are test images at multiple time points, in some embodiments, one pixel value may be captured in each grayscale block of each test image.

Next, in each test image, the image processor 120 calculates a block pixel value and a block standard deviation of each gray-scale block according to the captured pixel values (step S480). The implementation of the image processor 120 calculating the block pixel value and the block standard deviation of each grayscale block in each comparison image is roughly the same as the representative pixel value of step S430 in FIG. 4A (corresponding to the block pixel Value) is the same as the implementation of the representative standard deviation (corresponding to the block standard deviation), so it will not be repeated here.

Finally, in each test image, the image processor 120 averages the block pixel values of the gray-scale blocks at the same position as the corresponding representative pixel values, and averages the block standard deviation of the gray-scale blocks at the same position As the corresponding representative standard deviation (step S490). Taking the grayscale block Re0 of the test image NTC as an example for illustration, the image processor 120 averages the block pixel values of the grayscale block Re0 at the same position in each test image NTC as the corresponding representative pixel value I0 , And the average block standard deviation of the gray-scale block Re0 at the same position is taken as the corresponding representative standard deviation σ0. The representative pixel values and standard deviations corresponding to the gray-scale blocks Re0-Re11 are shown in the above table <1>.

As shown in FIG. 5, in this embodiment, the pixel value interval is set to be the same as the dynamic range, that is, 0-255. The image processor 120 corresponds the representative pixel values I0-I11 of the gray-scale blocks Re0-Re11 and the corresponding representative standard deviations σ0-σ11 to the partial pixel values in the pixel value interval and the corresponding standard deviations.

3 and 5-6 at the same time, in order to find the standard deviation corresponding to each pixel value in the pixel value interval, the image processor 120 will calculate the intermediate values corresponding to the multiple intermediate pixel values between these representative pixel values I0-I11 Standard deviation to form a smooth curve Crv1. Furthermore, the image processor 120 will calculate the intermediate standard deviation corresponding to each intermediate pixel value between the representative pixel values according to these representative pixel values and the corresponding representative standard deviation, and will represent the representative pixel value and the corresponding representative standard deviation 3. The intermediate pixel value and the intermediate standard deviation are respectively mapped to the multiple pixel values and the multiple standard deviations in the smooth curve Crv1, thereby finding the standard deviation corresponding to each pixel value in the pixel value interval (step S320).

It is worth noting that two adjacent pixel values will have the characteristics of close noise size. Therefore, the standard deviation corresponding to each pixel value in the smooth curve Crv1 should be as close as possible to the representative standard deviation σ0-σ11 corresponding to the representative pixel values I0-I11 (hereinafter referred to as data fitness), and each pixel The change between the standard deviations corresponding to the values should be as smooth as possible (hereinafter referred to as smoothness). Therefore, in this embodiment, the electronic device 100 will estimate each intermediate image through a smooth curve function The median standard deviation corresponding to the prime value. The smooth curve function of this embodiment is as shown in equation (1).

其中，n為代表像素值的個數(本實施例的代表像素值I0-I11的個數為12)，m為像素值區間內像素值的個數(本實施例的個數為256)，λ為一平滑程度的控制參數，x為描述動態範圍內每一訊號之對應標準差(本實施的像素值0-255分別對應的標準差)的一標準差向量，A₁為描述每一個代表像素值的一第一矩陣(本實施例的代表像素值I0-I11)，b1為描述每一個代表標準差的一第一向量(本實施例的代表標準差為σ0-σ11)，A₂為描述每一個標準差對應的前一個標準差與下一個標準差為一平滑關係的一第二矩陣，第二向量b2為0向量。

Where n is the number of representative pixel values (the number of representative pixel values I0-I11 in this embodiment is 12), and m is the number of pixel values in the pixel value interval (the number of pixel values in this embodiment is 256), λ is a control parameter for the degree of smoothness, x is a standard deviation vector describing the corresponding standard deviation of each signal in the dynamic range (the standard deviation corresponding to pixel values 0-255 in this implementation), and A ₁ is a description of each representative A first matrix of pixel values (representing pixel values I0-I11 in this embodiment), b1 is a first vector describing each representative standard deviation (representing standard deviation in this embodiment is σ0-σ11), A ₂ is A second matrix describing a smooth relationship between the previous standard deviation and the next standard deviation corresponding to each standard deviation, the second vector b2 is a zero vector.

Furthermore, the first matrix A ₁ and the corresponding first vector b1 are used to describe data matching. Therefore, in this embodiment, the first matrix A ₁ and the first vector b1 are as shown in equations (2) and (3), respectively.

於式(2)中，當第i個取樣像素值屬於像素區間內第j個像素值時，V₁(i,j)=1；其他情形，V₁(i,j)=0。於式(3)中，n為代表像素值的個數，且σ0-σn-1為代表標準差的數值。第一矩陣A₁的矩陣大小為(代表像素值的個數*像素區間內像素值的個數)，且第 一向量b1的大小為(代表像素值的個數*1)。然而，第一矩陣A₁與第一向量b1也可以其他描述方式來作設計，本案對此不作限制。

In equation (2), when the i-th sampled pixel value belongs to the j-th pixel value in the pixel interval, V ₁ (i,j)=1; otherwise, V ₁ (i,j)=0. In equation (3), n is the number representing the pixel value, and σ0-σn-1 is the value representing the standard deviation. The matrix size of the first matrix A ₁ is (representing the number of pixel values*the number of pixel values in the pixel interval), and the size of the first vector b1 is (representing the number of pixel values*1). However, the first matrix A ₁ and the first vector b1 can also be designed in other description manners, which is not limited in this case.

In addition, the second matrix A ₂ (indicating that the previous standard deviation and the next standard deviation corresponding to each standard deviation are in a smooth relationship) and the corresponding second vector b2 are used to describe data smoothing. Therefore, in this embodiment, the second matrix A ₂ and the second vector can be represented by the following equations (4) and (5), respectively.

其中，V₂(i,j)為第二矩陣A₂中的第i行第j列的數值，m為像素值區間內像素值的個數。故第二矩陣A₂的矩陣大小為(像素區間內像素值的個數*像素區間內像素值的個數)，且第二向量b2的大小為(像素區間內像素值的個數*1)。而第二矩陣A₂與第二向量b2也可以其他描述方式來作設計，本案對此不作限制。

Where V ₂ (i, j) is the value of the i-th row and j-th column in the second matrix A ₂ , and m is the number of pixel values in the pixel value interval. Therefore, the matrix size of the second matrix A ₂ is (the number of pixel values in the pixel interval*the number of pixel values in the pixel interval), and the size of the second vector b2 is (the number of pixel values in the pixel interval*1) . The second matrix A ₂ and the second vector b2 can also be designed by other description methods, which is not limited in this case.

Therefore, the image processor 120 estimates each pixel value (0-255 in this embodiment) and the corresponding standard deviation (ie vector x) in the pixel interval according to the smooth curve function of formula (1) to form Smooth the curve Crv1, as shown in Figure 6. At this time, the smooth curve Crv1 represents the non-linear relationship of the simulated real SDN under different pixel values.

3 and 7 at the same time, in order to convert signal-related noise (SDN) into signal-independent noise (SIN), the image processor 120 will then follow the smoothing curve Crv1 for each pixel value and the corresponding standard The difference calculates an equalization curve Crv2 to correspond each pixel value to an equalized value (step S330). Go further That is to say, the image processor 120 converts the noise under different pixel values into the same or similar values through the equalization curve Crv2. Therefore, the equalization curve Crv2 needs to satisfy a distance that the current pixel value moves by one standard deviation, and the corresponding value after the equalization moves by one unit.

According to this, in this embodiment, the equalization curve Crv2 needs to satisfy an equalized value corresponding to an adjusted pixel value minus the equalized value corresponding to the current pixel value to 1 unit, where the adjusted pixel value represents a current pixel Value plus corresponding standard deviation. And this satisfying condition can be sorted into formula (6), as shown below.

其中，Teq為等化後數值，a為等化曲線Crv2中的一目前像素值，Xa為目前像素值對應的標準差，且m為像素區間內像素值的個數。

Where Teq is the value after equalization, a is a current pixel value in the equalization curve Crv2, Xa is the standard deviation corresponding to the current pixel value, and m is the number of pixel values in the pixel interval.

Furthermore, the equalization curve Crv2 is also satisfied. The equalized value corresponding to the next pixel value and the equalized value corresponding to the previous pixel value are subtracted and divided by 2 (can be regarded as the curve at the current pixel value a Is the reciprocal of the standard deviation corresponding to the current pixel value. And this satisfying condition can be sorted into equation (7), as shown below.

另外，當a=0時，Teq(1)-Teq(0)=1/X0；當a=m-1時，Teq(m-1)-Teq(m-2)=1/Xm-1，其中，Teq為等化後數值，a為等化曲線Crv2中的一目前像素值，Xa為目前像素值對應的標準差，且m為動態範圍內訊號的個數。

In addition, when a=0, Teq(1)-Teq(0)=1/X0; when a=m-1, Teq(m-1)-Teq(m-2)=1/Xm-1, Where Teq is the value after equalization, a is a current pixel value in the equalization curve Crv2, Xa is the standard deviation corresponding to the current pixel value, and m is the number of signals in the dynamic range.

Therefore, in this embodiment, the electronic device 100 can calculate the equalized value corresponding to each pixel value through Equation (7), and the arrangement is as in Equation (8).

The image processor 120 performs matrix operation of equation (8) to obtain the equalized values Teq(0), Teq(1)..., Teq(255) corresponding to each pixel value (ie 0-255), so that the equalization Curve Crv2 can satisfy the condition that a current pixel value moves by one standard deviation, and the corresponding equalized value is the condition of moving one unit, and then convert signal-related noise (SDN) into signal-independent noise (SIN) .

Take the image processor 120 receiving a current pixel value of 30 as an example. 6-7, the electronic device 100 obtains a standard deviation corresponding to the current pixel value 30 of 2.37 according to the smoothing curve Crv1 (see FIG. 6 and table <1>), and the equalized value Teq(31) corresponding to the next pixel value 31 =22.31, and the equalized value Teq(29)=21.47 corresponding to the previous pixel value 29. The image processor 120 calculates [Teq(31)-Teq(29)]/2=(22.31-21.47)/2=0.42, and 1/2.37=0.42 according to equation (7), that is, the equalization curve Crv2 meets the equation ( 7) Conditions.

Accordingly, as shown in FIG. 8, after receiving each input pixel value P0-Pn of the input image Fr1 of FIG. 2, the image processor 120 obtains the corresponding value of each input pixel value P0-Pn through the equalization curve Crv2. The equalized value is used to generate the equalized input image Fr1′, so that the noise value under different pixel values is the same or close. As shown in FIG. 8, in the equalized input image Fr1′, the image processor 120 calculates the regional standard deviation Std of these pixel values in the image block ZN1 to be 4.32, and calculates the The regional standard deviation Std of these pixel values is 4.33, which means that the equalized input image Fr1' has noise of the same or close intensity. Therefore, the image processor 120 has better noise removal effect on the equalized input image Fr1' to remove noise.

After obtaining the equalization curve, the image processor 120 stores the memory (not shown in the drawings) or an external storage (not shown in the drawings) in the electronic device 100. When the image processor 120 receives each input pixel value in an input image, it can obtain the equalized value corresponding to each input pixel value according to the above equalization curve, and generate lower noise by the noise removal method The output pixel value of the signal is P0'-Pn'.

<An embodiment of the noise removal method>

Reference is also made to FIGS. 2 and 9, which are flowcharts of the noise removal method according to an embodiment of the present invention. First, the image processor 120 of the electronic device 100 will receive an input pixel value at each pixel position in an input image (step S910). For example, as shown in FIG. 10, the size of the input image Fr is 8*8 and has 64 input pixel values. The image processor 120 will receive the input pixel values P0-P63 at each pixel position in the input image Fr.

Next, the image processor 120 will sequentially extract a current position block among the pixel positions of the input image (step S920). The current location block can correspond to one input pixel value or multiple input pixel values. Furthermore, the shape of the current location block may be a square. The size of the input image is M*M, and the size of the current location block is K*K, where 1<=K<M and K and M are positive integers. Therefore, the image processor 120 will sequentially retrieve the corresponding input pixel values in the current location block according to the shape and size of the current location block. For example, if the size of the current location block is 1*1, the image processor 120 will sequentially capture the corresponding input pixel values in units of one pixel location size. For another example, if the size of the current location block is 3*3, the image processor 120 will sequentially capture the corresponding input pixel values in units of 9 pixel location sizes.

In other embodiments, the current location block may also have other shapes. The image processor 120 may also have repeated input pixel values each time the current location block is captured, which is not limited in this case.

After obtaining the current location block (step S920), the image processor 120 will capture multiple adjacent blocks adjacent to the current location block (step S930). The shape and size of the current location block and the neighboring block are the same. Furthermore, during the process of capturing neighboring blocks, the image processor 120 will determine that the current block has an input pixel value or multiple input pixel values. If the image processor 120 determines that the current location block has only one input pixel value, the image processor 120 will capture the input pixel value of the neighboring current location block according to the number of neighboring blocks, and separately capture the captured input images Prime values are used as neighboring blocks.

If the image processor 120 determines that the current location block has a plurality of input pixel values, the image processor 120 will search for blocks of the same size in the neighboring area according to the size of the current location area to generate corresponding neighboring blocks.

The current location block Pch corresponds to 9 input pixel values P26-P28, P34-P36, P42-P44 and the number of neighboring blocks is 2 as an example. As shown in FIG. 10, the image processor 120 diffuses the distance of 1 pixel position outward based on the input pixel values P26 and P44 to generate corresponding adjacent blocks Qch1 and Qch2. The neighboring block Qch1 corresponds to nine input pixel values P17-P19, P25-P27, P33-P35. The neighboring block Qch2 corresponds to 9 input pixel values P35-P37, P43-P45, P51-P53. At this time, the shape and size of the current location block Pch and the neighboring blocks Qch1 and Qch2 are the same.

Returning to FIG. 9, after obtaining a plurality of neighboring blocks (step S930), the image processor 120 then maps the input pixel values of the current position block and each of the neighboring blocks via an equalization curve to obtain each A certain equalized value corresponding to the pixel value is input (step S940). In this embodiment, referring to FIG. 7 at the same time, the image processor 120 inputs each pixel value P17-P19, P25-P28, P33-P37, P42-P45 in the current position block Pch and the adjacent blocks Qch1-Qch2 , P51-P53 are mapped via the equalization curve Crv2 to obtain the corresponding equalized values Teq(P17)-Teq(P19), Teq(P25)-Teq(P28), Teq(P33)-Teq(P37), Teq (P42)-Teq(P45), Teq(P51)-Teq(P53).

Next, the image processor 120 separately calculates a degree of difference between each neighboring block and the current location block (step S950). Furthermore, the image processor 120 calculates the difference between the equalized value of each signal in the current location block and the equalized value of each signal in each neighboring block as the size of several standard deviations to generate each correspondingly The degree of difference between neighboring blocks. Following the above example, the image processor 120 will calculate the difference between the equalized value in the current location block Pch and the equalized value in the neighboring block Qch1 by several standard deviations to correspondingly generate the neighboring block Qch1 The degree of difference Diff(Qch1). The image processor 120 will also calculate the difference between the equalized value in the current location block Pch and the equalized value in the neighboring block Qch2 as several standard deviations to correspond to the difference degree Diff generated in the neighboring block Qch2 (Qch2).

In this embodiment, the degree of difference between neighboring blocks is calculated by a degree of difference function, and is shown in the following formula (9).

其中S為目前位置區塊Pch中的像素個數，Teq(Pch)為目前位置區塊的等化後數值，Teq(Qchn)為第n個鄰近區塊中的等化後數值，|Teq(Pch)-Teq(Qchn)|為目前位置區塊中的等化後數值與第n個鄰近區塊中的等化後數值之間的幾個標準差的差距，Diff(Qchn)為第n個鄰近區塊與目前位置區塊Pch的差異程度。

Where S is the number of pixels in the current location block Pch, Teq(Pch) is the equalized value of the current location block, Teq(Qchn) is the equalized value in the nth neighboring block, Teq( Pch)-Teq(Qchn)| is the difference of several standard deviations between the equalized value in the current location block and the equalized value in the nth neighboring block, Diff(Qchn) is the nth The degree of difference between the adjacent block and the current location block Pch.

The image processor 120 calculates the difference degree Diff(Qch1) of the adjacent block Qch1 as an example for illustration. The number S of pixels in the current position block Pch is 9. |The total of Teq(Pch)-Teq(Qchn)| is |Teq(P26)-Teq(P17)|+|Teq(P27)-Teq(P18)|+|Teq(P28)-Teq(P19)| +|Teq(P34)-Teq(P25)|+|Teq(P35)-Teq(P26)|+|Teq(P36)-Teq(P27)|+|Teq(P42)-Teq(P33)|+| Teq(P43)-Teq(P34)|+|Teq(P44)-Teq(P35)|. In this way, the image processor 120 can calculate the difference Diff(Qch1) of the neighboring block Qch1 by dividing the total of |Teq(Pch)-Teq(Qchn)| by the number S.

After obtaining the difference degree of each neighboring block, the image processor 120 will determine a weight value of the corresponding neighboring block according to each difference degree (step S960). Furthermore, if the difference is smaller, it means that the neighboring block Qch1 is more similar to the current block Pch. At this time, the image processor 120 will determine a larger weight value according to the degree of difference. Conversely, if the difference is larger, it means that the neighboring block Qch1 is less like the current block Pch. At this time, the image processor 120 will determine the smaller weight value according to the degree of difference.

In this embodiment, the image processor 120 calculates the weight value through a weight value function, as shown in the following formula (10).

其中，ω(Qchn)為第n個鄰近區塊的權重值，Diff(Qchn)為第n個鄰近區塊的差異程度，k為相似與不相似的基準常數，物理意義上為k個標準差。

Among them, ω(Qchn) is the weight value of the nth neighboring block, Diff(Qchn) is the degree of difference of the nth neighboring block, k is the reference constant of similarity and dissimilarity, and is k standard deviations in the physical sense .

For example, the image processor 120 determines the weight value ω (Qch1) of the neighboring block Qch1 and the weight value ω (Qch2) of the neighboring block Qch2. In this example, k is 2, the degree of difference Diff(Qch1) is 2 and the degree of difference Diff(Qch2) is 3. Therefore, the weight value ω(Qch1)=exp(-2/4)=0.6 of the neighboring block Qch1. The weight value ω(Qch2)=exp(-7/4)=0.17 of the neighboring block Qch2. According to the above example, the smaller the difference between the adjacent block Qch1 and the current location block Pch, the more similar the adjacent block Qch1 and the current location block Pch are, so the image processor 120 will determine the larger Weight, this larger weight represents a higher reference value.

After determining the weight value of each neighboring block (step S960), the image processor 120 weights and averages a main pixel value in the corresponding neighboring block according to these weight values to generate a modified pixel value (step S970) ). In this embodiment, the main pixel value in the neighboring block may be the input pixel value at an intermediate pixel position, or the average value after averaging each input pixel value in the neighboring block. Of course, the main pixel value in the neighboring block can also be obtained by other means, and this case does not limit it.

In addition, the image processor 120 of this embodiment calculates the corrected pixel value through a correction function, as shown in the following formula (11).

其中，COR為修正像素值，ω(Qchn)為第n個鄰近區塊的權重值，M(Qchn)為第n個鄰近區塊中的一主像素值。

Where COR is the corrected pixel value, ω(Qchn) is the weight value of the nth neighboring block, and M(Qchn) is a main pixel value in the nth neighboring block.

Following the above example, the weight value ω(Qch1) of the neighboring block Qch1=0.6, and the weight value ω(Qch2) of the neighboring block Qch2=0.17. The main pixel value of the adjacent block Qch1 is, for example, an input pixel value P26 at an intermediate pixel position, and the main pixel value of the adjacent block Qch1 is, for example, an input pixel value P44 of an intermediate pixel position. Therefore, the corrected pixel value COR=[(0.6*P26)+(0.17*P44)]/(0.6+0.17).

In some embodiments, the current position block Pch itself can also be regarded as one of the neighboring blocks, and its corresponding weight value can be set by itself, such as 0.8, therefore, the corrected pixel value COR=[(0.8*P35)+(0.6* P26)+(0.17*P44)]/(0.8+0.6+0.17).

Next, the image processor 120 generates each output pixel value in the current location block corresponding to the corrected pixel value (step S980). In some embodiments, if the corrected pixel value is only generated for the specific pixel of the current location block, step S980 can also be ignored, and this case is not limited to this.

Therefore, after the image processor 120 executes the above noise removal method (ie steps S910-S970) on each input pixel value P0-Pn in the input image Fr1, the image processor 120 will generate a lower noise output pixel value P0'-Pn'.

<Another embodiment of noise removal method>

2 and 11, FIG. 11 shows a flowchart of a noise removal method according to another embodiment of the present case. First, the image processor 120 of the electronic device 100 receives an input pixel value at each pixel position in an input image (step S1110). For example, as shown in FIG. 10, the size of the input image Fr is 8*8 and has 64 input pixel values. The image processor 120 will receive the input pixel values P0-P63 at each pixel position in the input image Fr.

Next, the image processor 120 will sequentially extract a block of the current position among the pixel positions of the input image (step S1120). The current location block corresponds to multiple input pixel values, and one of the input pixel values has a main pixel value. more Further, the shape of the current location block is a square. The size of the input image is M*M, and the size of the current location block is K*K, where 1<K<M and K and M are positive integers. Therefore, the image processor 120 will sequentially retrieve the corresponding input pixel values in the current location block according to the shape and size of the current location block. For example, if the size of the current location block is 3*3, the image processor 120 will sequentially capture the corresponding input pixel values in units of 9 pixel location sizes.

In other embodiments, the current location block may also have other shapes. The image processor 120 may also have repeated input pixel values each time the current location block is captured, which is not limited in this case. In addition, in this embodiment, the main pixel value may be an input pixel value at an intermediate pixel position in the corresponding current position block, or an average value of each input pixel value in the current position block. Of course, the main pixel value in the current location block can also be obtained by other means, and this case does not limit it.

After obtaining the current position block (step S1120), the image processor 120 then maps each input pixel value in the current position block through an equalization curve to obtain a certain equalization corresponding to each input pixel value The value (step S1130). Please refer to FIG. 7 and FIG. 10 at the same time, taking the current position block Pch as an example for illustration. The image processor 120 maps each input pixel value P26-P28, P34-P36, P42-P44 in the current position block Pch through the equalization curve Crv2 to obtain the corresponding equalized value Teq(P26)-Teq( P28), Teq(P34)-Teq(P36), Teq(P42)-Teq(P44).

Next, the image processor 120 averages the equalized values corresponding to the input pixel values in the current position block to generate a low-frequency component value (step S1140). Following the above example, the low-frequency component value E[Teq(Pch)]=[(Teq(P26)+Teq(P27)+Teq(P28)+Teq(P34)+Teq(P35)+Teq( P36)+Teq(P42)+Teq(P43)+Teq(P44)]/9.

Next, the image processor 120 calculates a variation between the equalized value corresponding to the input pixel value in the current location block and the low-frequency component value (step S1150). more Further, the image processor 120 calculates the variance through a variance function, as shown in the following formula (12).

其中，Var(Pch)為目前位置區塊中的變異數，Teq(x)為某一個輸入像素值對應的等化後數值，E[Teq(Pch)]為目前位置區塊中的低頻成分值，且P為目前位置區塊中的這些輸入像素值的個數。

Among them, Var(Pch) is the number of mutations in the current location block, Teq(x) is the equalized value corresponding to a certain input pixel value, and E[Teq(Pch)] is the low-frequency component value in the current location block And P is the number of these input pixel values in the current location block.

Following the above example, the number of variations of the current location block Pch Var(Pch)=[(Teq[P26]-E[Teq(Pch)])^2+(Teq[P27]-E[Teq(Pch)])^ 2+(Teq[P28]-E[Teq(Pch)])^2+(Teq[P34]-E[Teq(Pch)])^2+(Teq[P35]-E[Teq(Pch)]) ^2+(Teq[P36]-E[Teq(Pch)])^2+(Teq[P42]-E[Teq(Pch)])^2+(Teq[P43]-E[Teq(Pch)] )^2+(Teq[P44]-E[Teq(Pch)])^2]/9.

Next, the image processor 120 calculates a high-frequency component value of the main pixel value in the current location block according to the low-frequency component value (step S1160). Furthermore, the image processor 120 subtracts the low-frequency component value to generate the high-frequency component value. Following the above example, the main pixel value in the current position block Pch is, for example, the input pixel value P35. Therefore, the high-frequency component value H(Teq(Pch))=input pixel value P35−low-frequency component value E[Teq(Pch)].

It is worth noting that the main pixel value of the current position block Pch is composed of the low-frequency component value E[Teq(Pch)] and the high-frequency component value H(Teq(Pch)). The low-frequency component value E[Teq(Pch)] describes the flat part in the main pixel value. The high-frequency component value H(Teq(Pch)) describes the details and texture in the main pixel value. However, the noise in the main pixel value may also exist in the high-frequency component value H(Teq(Pch)).

Therefore, the image processor 120 will calculate a high-frequency ratio of the high-frequency component value of the main pixel value according to the variance in the current location block to obtain the high-frequency component value for removing noise (step S1170). In this embodiment, the high frequency of the high-frequency component value The ratio is calculated by a high-frequency proportional function and is shown in the following formula (13).

其中，RTO(Pch)為目前位置區塊中高頻成分值的高頻比例，Var(Pch)為目前位置區塊中的變異數，而”1”則代表一雜訊分布常數。

Among them, RTO (Pch) is the high-frequency proportion of the high-frequency component value in the current location block, Var (Pch) is the variation in the current location block, and "1" represents a noise distribution constant.

In this way, if the difference between the variance and the noise distribution constant is smaller, it means that more noise exists in the high-frequency component value H(Teq(Pch)). At this time, the video processor 120 will calculate the lower high-frequency ratio RTO(Pch) to reduce the high-frequency component value H(Teq(Pch)). Conversely, if the difference between the variance and the noise distribution constant is larger, it means that more details and textures are present in the high-frequency component value H(Teq(Pch)). At this time, the image processor 120 will calculate a higher high-frequency ratio RTO(Pch) to increase the high-frequency component value H(Teq(Pch)).

Finally, the image processor 120 adds the low-frequency component value to the high-frequency component high-frequency component value in the current location block to generate an output pixel value corresponding to the main pixel value (step S1180). Following the above example, the main pixel value in the current position block Pch is the input pixel value P35. The image processor 120 adds the low frequency component value E[Teq(Pch)] to the high frequency component value H(Teq(Pch)) of the high frequency ratio RTO(Pch) to generate an output pixel value input pixel value P35 Value P35'.

Therefore, when the image processor 120 executes the above noise removal method (ie steps S1110-S1180) on each input pixel value P0-Pn in the input image Fr1, the image processor 120 will generate a lower noise output pixel value P0'-Pn'.

In summary, a noise equalization method and noise removal method for different pixel values provided by the embodiments of the present invention. The noise equalization method in this case is to simulate the degree of noise under different pixel values through a smooth curve, and then to achieve the degree of noise equalization under different pixel values through an equalization curve, so that the noise under different pixel values The value of the signal is the same or close (for example, the value is 1), which means that the signal-related noise (SDN) is converted into signal-independent noise (SIN), and then received from Signals affected by noise can more accurately estimate the SDN characteristics of real signals. In addition, using the above-mentioned equalization curve to perform the noise removal method in this case can simplify the calculation complexity and calculate better noise parameters (such as standard deviation and variance), not only for pixel values under different signal strengths In addition to the consistent noise reduction capability, it can also improve the noise removal effect of each input pixel value in an input image, thereby producing a lower noise output pixel value.

100‧‧‧Electronic device

110‧‧‧Image capture device

120‧‧‧Image processor

Fr1‧‧‧ input image

P0-Pn‧‧‧Input pixel value

P0’-Pn’‧‧‧ output pixel value

Claims (10)

An image noise equalization method is suitable for an electronic device to equalize noise under different values. The noise equalization method includes: obtaining a plurality of representative values in a value interval corresponding to the representative values A plurality of representative standard deviations; calculate an intermediate standard deviation corresponding to each intermediate value between the representative pixel values according to the representative values and the corresponding representative standard deviations, and the representative pixel values and the representative Standard deviation, the intermediate values and the intermediate standard deviations are mapped to multiple values and multiple standard deviations in a smooth curve, respectively; and a first class is calculated based on each of the values in the smooth curve and the corresponding standard deviation Curve to correspond each value to an equalized value; where a current value of the values plus the corresponding standard deviation is an adjusted value, and the equalized curve satisfies the corresponding value of the adjusted value The equalized value minus the equalized value corresponding to the current value is 1 unit. As in the noise equalization method of claim 1, wherein the slope of the current value on the equalization curve is the reciprocal of the standard deviation corresponding to the current value. For example, the noise equalization method of claim 1, wherein the step of obtaining the representative values and the corresponding standard deviations further includes: receiving a test image, wherein the test image has multiple different blocks; Retrieving a plurality of first values in each of the blocks; and in each of the blocks, calculating the representative value and the standard deviation of the corresponding block based on the retrieved first values, to Obtain the representative values and the corresponding standard deviations. The noise equalization method of claim 1, wherein the step of obtaining the representative values and the corresponding standard deviations further comprises: receiving a test image at multiple time points, wherein the test image has multiple Different blocks; at least one first value is captured in each of the blocks of each of the test images; in each of the test images, each area is calculated based on the captured first values A block value of a block and a block standard deviation; and the corresponding representative value is calculated according to the block values of the blocks at the same position, and according to the blocks of the blocks at the same position The block standard deviation serves as the corresponding representative standard deviation. An image noise removal method is suitable for an electronic device, and the noise removal method includes: receiving an input image; capturing a current location block in the input image, wherein the current location block corresponds to one or Multiple input pixel values; extract multiple adjacent blocks adjacent to the current position block; map each input pixel value in the current position block and the adjacent blocks through an equalization curve to obtain each The equalized value corresponding to the input pixel value, wherein the equalization curve includes multiple values and corresponding multiple equalized values, a current value of the values plus a corresponding standard deviation is an adjusted value , And the equalization curve satisfies the equalized value corresponding to the adjusted value minus the equalized value corresponding to the current value is 1 unit; based on the equalized values, each of the neighboring blocks is calculated separately A degree of difference from the current location block; according to each degree of difference, a weight value corresponding to adjacent blocks is determined; according to the weight values, a corresponding main pixel value in the adjacent blocks is weighted and averaged To generate a modified pixel value. The noise removal method of claim 5, wherein the step of weighting and averaging the main pixel values in the corresponding neighboring blocks to generate the corrected pixel value according to the weight values further includes: setting the current A current weight value of the position block; according to the weight values and the current weight value, the weighted average of the corresponding main pixel value in the adjacent blocks and the main pixel of the current position block is generated to generate the correction Pixel values. As in the noise removal method of claim 5, in the step of determining the weight value of each of the neighboring blocks, the method further includes: separately calculating the equalized values and each of the equalized values in the current location block A standard deviation error between the equalized values in the neighboring blocks to correspondingly generate the degree of difference for each of the neighboring blocks, wherein, if the degree of difference is smaller, the weight value corresponding to the degree of difference corresponds to The larger, and if the degree of difference is greater, the weight value corresponding to the degree of difference is smaller. The noise removal method according to claim 5, wherein the main pixel value in each of the adjacent blocks is the input pixel value at an intermediate pixel position of each of the adjacent blocks, or each of the adjacent areas The average value of each of the input pixel values in the block. An image noise removal method is suitable for an electronic device, and the noise removal method includes: receiving an input image; capturing a current location block from the input image, wherein the current location block corresponds to multiple Input pixel values; determine a main pixel value according to the input pixel values; map the input pixel values through an equalization curve to obtain corresponding multiple The normalized value, where the equalization curve includes multiple values and corresponding multiple equalized values, a current value of the values plus a corresponding standard deviation is an adjusted value, and the equalization curve satisfies the Adjust the equalized value corresponding to the value minus the equalized value corresponding to the current value to 1 unit; average the equalized values to generate a low-frequency component value; calculate the corresponding values of the input pixel values A variation between the equalized value and the low-frequency component value; calculating a high-frequency component value of the main pixel value in the current location block according to the low-frequency component value; calculating the main pixel value based on the variation number A high frequency ratio of the high frequency component value; and adding the low frequency component value to the high frequency component value of the high frequency ratio to generate an output pixel value corresponding to the main pixel value. The noise removal method of claim 9, wherein the main pixel value is one of the input pixel values, or an average value of the input pixel values.

Images