A color image denoising method under extremely low illumination

Abstract

The application discloses a color image denoising method under extremely low illumination, which is suitable for fast denoising of color images under extremely low illumination and comprises the following steps: converting a source color image from an RGB space into an image in a YCbCr space, and saving a converted blue chroma component Cb image and a red chroma component Cr image; extracting the converted luminance component Y; measuring the image noise intensity and image smoothness of the luminance component Y; obtaining filter parameters of a non-local mean filter based on the image noise intensity and image smoothness of the luminance component Y; and performing denoising on the image of the luminance component Y by using the non-local mean filter based on the obtained filter parameters to obtain a denoised luminance component Y until the denoising meets the requirements; and merging the denoised luminance component Y with the saved blue chroma component Cb image and red chroma component Cr image in step S1, and converting the merged image into an image in an RGB color space.

Description

Technical Field

[0001] This invention relates to the field of image processing noise reduction, and in particular to a method for noise reduction of color images under extremely low illumination. Background Technology

[0002] Currently, the poor image quality of extremely low-light color images is a major concern for various military and civilian low-light cameras. A key challenge lies in noise reduction during this process. While typical image denoising problems primarily target grayscale images contaminated with Gaussian noise, denoising color images under extremely low light conditions is far more complex. The difficulties stem from two main factors: firstly, the unique nature of the noise model in extremely low-light imaging, which uses a Poisson distribution instead of a Gaussian distribution; and secondly, the multi-channel nature of color images. Compared to the single-channel nature of grayscale images, color images require complex vector processing instead of simple scalar processing, leading to excessive computational demands and time consumption.

[0003] Image denoising algorithms can generally be divided into two main categories based on their processing domain: spatial domain denoising algorithms and frequency domain denoising algorithms. Generally speaking, frequency domain methods require a process of forward frequency transformation, frequency processing, and inverse frequency transformation, which is time-consuming and not conducive to fast processing. Spatial domain image denoising algorithms are a type of denoising algorithm that directly processes the grayscale values ​​on the original image. They can be divided into two main categories: one is point-by-point operation, which processes the image point by point; the other is local operation, which processes the image in the neighborhood space of each pixel. Traditional spatial domain denoising algorithms include mean filtering, partial differential filtering, and total variation filtering. Essentially, mean filtering avoids the variation calculation process of partial differential filtering and total variation filtering algorithms, and its fast computation speed meets application requirements. However, mean filtering loses detail information, causing a certain degree of image blurring. Therefore, for denoising under extremely low illumination, various deformation filtering methods are mostly used.

[0004] Non-local means (NLM) image denoising is a spatial domain denoising algorithm that originates from the mean filtering algorithm and is a further extension of it. Compared to neighborhood filtering algorithms, which use the gray values ​​of individual pixels to calculate weights, NLM compares the gray value distribution of the entire area surrounding a single pixel and calculates weights based on the similarity of the gray value distributions. However, its drawbacks include: 1) the choice of local window size has a significant impact on the denoising effect, making it difficult to select an appropriate window size; 2) the computational cost of local similarity is high, resulting in low practical application value; 3) under extremely low illumination, because it does not consider the noise intensity and smoothness of the original image, the amplification level of pixels near the edges cannot be controlled, leading to overly bright and wide edges.

[0005] Existing image denoising algorithms, such as median filtering and Gaussian filtering, all have their own limitations and are not suitable for noise filtering in extremely low light conditions. For example, median filtering uses the median of the pixel values ​​within the filter window to obtain the pixel value at the center of the window, and it is a non-linear smoothing filter. Its advantage is that it can suppress noise while preserving image details as much as possible, and it is most effective in filtering out impulse interference and image scanning noise. However, in extremely low light conditions, details such as points or lines in the image become more blurred.

[0006] Gaussian filtering is a type of linear filtering. Its principle involves convolving each point in the image with a Gaussian kernel, summing the results, and outputting the sum to the target image. Its advantages include consistent smoothness across all directions, preserving the original image's edge characteristics, and being unaffected by high-frequency signals during the filtering process. It also performs well in reducing noise that conforms to a Gaussian distribution. However, it struggles to reduce random noise with a Poisson distribution under extremely low illumination.

[0007] Therefore, existing noise reduction methods cannot be directly applied to noise reduction of color images under extremely low illumination, and their direct application results are poor. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for denoising color images under extremely low illumination, so as to achieve rapid denoising of color images under extremely low illumination.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows: a method for denoising color images under extremely low illumination, comprising the following steps:

[0010] S1. Convert the source color image from RGB space to YCbCr space, and save the converted blue chromaticity component Cb image and red chromaticity component Cr image respectively; extract the converted luminance component Y.

[0011] S2. Measure the image noise intensity and image smoothness of the luminance component Y;

[0012] S3. Obtain the filtering parameters for nonlocal mean filtering based on the image noise intensity and image smoothness of the luminance component Y; and use nonlocal mean filtering to denoise the image of the luminance component Y based on the obtained filtering parameters to obtain the denoised luminance component Y.

[0013] S4. Determine whether noise reduction is complete. If not, return to step S2; otherwise, proceed to step S5.

[0014] S5. After merging the image of the luminance component Y after noise reduction with the blue chromaticity component Cb image and the red chromaticity component Cr image saved in step S1, the merged image is converted to an image in the RGB color space.

[0015] Step S2 includes:

[0016] S21. Perform three-dimensional randomization on the brightness component Y image to generate a random noise image;

[0017] S22. Calculate the noise intensity of the noisy luminance component image;

[0018] S23. Calculate the image smoothness of the image containing the noisy luminance component.

[0019] In step S3, the method for obtaining the filter parameters of the nonlocal mean filtering algorithm includes:

[0020] The filter parameter h is calculated using the following formula:

[0021] Where α is the weighting coefficient. Let η be the noise intensity and η be the image smoothness.

[0022] Step S4 includes:

[0023] S41. Calculate the noise intensity based on the luminance component Y after noise reduction processing in step S3. and image smoothness η;

[0024] S42. The noise intensity calculated in step S41 is... The image smoothness η is compared with the pre-set image noise intensity threshold σ0 and image smoothness threshold η0;

[0025] when If any condition is not met, it is determined that the noise reduction is not finished and feedback is sent to step S2; otherwise, it is determined that the noise reduction is finished and step S5 is initiated.

[0026] Step S5 includes:

[0027] The denoised luminance component Y image is merged with the original blue and red chrominance component images, and the image is converted to the RGB color space. After obtaining the denoised luminance component, the blue chrominance component Cb and the red chrominance component Cr of the original low-light color image are merged, and then the image is converted from the YCbCr color space to the RGB color space. The conversion method uses the following formula to calculate the converted image information:

[0028] R=1.164*(Y-16)+1.596*(Cr-128)

[0029] G=1.164*(Y-16)-0.392*(Cb-128)-0.813*(Cr-128)

[0030] B=1.164*(Y-16)+2.017*(Cb-128)

[0031] Step S1 includes:

[0032] S11. Obtain the source color image A to be denoised;

[0033] S12. Convert the source color image A from the RGB color space to the YCbCr color space:

[0034] Normalize the RGB image values ​​to the [0,1] interval, and then substitute the normalized RGB image values ​​into the following formula to obtain the corresponding luminance component Y, blue chromaticity component Cb, and red chromaticity component Cr: Y = 0.257*R + 0.564*G + 0.098*B + 16

[0035] Cb=-0.148*R-0.291*G+0.439*B+128

[0036] Cr=0.439*R-0.368*G-0.071*B+128.

[0037] Step S22, measuring the image noise intensity of the luminance component Y, includes:

[0038] To obtain the noisy luminance component Y image after processing in step S21, the image is first divided into overlapping blocks, and the covariance matrix of L blocks is calculated. Then, the local variance of each image block is calculated. Ten image blocks are randomly selected, and the block with the smallest local variance is designated as the optimal smoothed image block. A covariance matrix ∑ is constructed using the selected smoothed image block, and then the minimum eigenvalue λ of this matrix is ​​calculated. min Finally, the estimated noise intensity is calculated using the following formula:

[0039]

[0040] Where L is the number of smoothing blocks used to construct the covariance matrix ∑.

[0041] Step S23, measuring the image smoothness of the luminance component Y, includes:

[0042] Obtain the noisy luminance component Y image after processing in step S21, select a Gaussian filter kernel k, convolve the Gaussian filter kernel with the luminance component and take the absolute value to obtain the filtered image Res: Res=|k*Y|, where * means convolution;

[0043] Calculate image smoothness using the following formula:

[0044]

[0045] Step S21 includes:

[0046] Randomized pixel coordinates:

[0047] Suppose the resolution supported by the luminance component Y image is M rows and N columns, and randomized pixel coordinates are generated using the linear congruential method:

[0048] x0 = (c1*Seed) x +disp1)mod N

[0049] x i =(c1*x i-1 +disp1)mod N

[0050] y0=(c2*Seed y +disp2)mod M

[0051] y j =(c2*y j-1 +disp2)mod M

[0052] Where c1 and c2 are coefficients, which can be determined by the values ​​of M and N; Seed x and Seed y , , are the two-dimensional Cartesian coordinates of the image seed point Seed; disp1 and disp2 are the minimum intervals between random numbers along the x-axis and y-axis, respectively;

[0053] Randomized pixel values: After randomizing the pixel coordinates, the randomized pixel coordinates of the luminance component Y are obtained as (x, y). Then, the randomized pixel values ​​are generated using the linear congruential method.

[0054] f(x0,y0)=(c3*Seed f(x,y) +disp3)mod 256

[0055] f p (x,y)=(c3*f p-1 (x,y)+disp3)mod 256

[0056] Where c3 is a coefficient, Seed f(x,y) The seed point values ​​are the image amplitude values ​​within the range, and disp3 is the minimum interval between the random amplitude numbers.

[0057] The advantages of this invention are:

[0058] 1. Suitable for noise reduction of color images under extremely low illumination, with good noise reduction effect;

[0059] 2. Convert the RGB image to the YCbCr image space, retaining the blue chromaticity components Cb and Cr unchanged, and only reducing noise in the luminance component Y, which is sensitive to the human eye. Therefore, the process is faster and can maintain less color distortion in the enhanced image.

[0060] 3. The self-feedback controllable nonlocal mean filter can adaptively feed back filtering parameters based on the image's noise intensity and smoothness, and automatically select the number of filtering iterations according to the control threshold until the filtering requirements are met. Fewer filtering iterations result in faster operation; more filtering iterations result in better filtering effects, which depends entirely on the user's needs.

[0061] 4. By adopting fast nonlocal mean filtering as the core, the image filtering speed is accelerated, which effectively improves the image processing speed in extremely low noise environments.

[0062] 5. Compared with other methods, this patent maintains good performance indicators in terms of peak signal-to-noise ratio and structural similarity. Attached Figure Description

[0063] The following is a brief explanation of the contents of each of the accompanying drawings and the markings in the drawings:

[0064] Figure 1 This is an example diagram of the three-dimensional randomization of the present invention;

[0065] Figure 2 This is a schematic diagram illustrating the filtering principle of the present invention; Detailed Implementation

[0066] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and the description of the preferred embodiments.

[0067] This application designs a method for denoising color images under extremely low illumination (referred to as self-feedback controllable nonlocal mean filtering for simplicity). For extremely low illumination color images, the image is first converted from RGB space to YCbCr space. Then, the separated blue chromaticity component (Cb) and red chromaticity component (Cr) are saved separately. Next, the noise intensity and smoothness of the luminance component (Y) are measured. Based on this, fast nonlocal mean filtering parameters are selected and denoising is performed. Afterward, the noise intensity and smoothness of the denoised image are measured, and the processing method is automatically selected according to a set control threshold. If the noise intensity exceeds the set control threshold, a second round of filtering is performed, with parameters automatically selected based on the noise intensity and smoothness of the image after the first round of denoising. This process is repeated until both indicators fall below the set control threshold. Finally, the denoised luminance component is merged with the previously saved blue chromaticity component (Cb) and red chromaticity component (Cr). The result is then converted back to RGB space, and the extremely low illumination denoised color image can be output.

[0068] like Figure 2 The flowchart shown is a noise reduction method of this application, and its steps include:

[0069] (1) Input a noisy color image with extremely low illumination;

[0070] (2) Convert the image from the RGB color space to the YCbCr color space;

[0071] (3) Separate and store the Cb and Cr components from the image corresponding to the YCbCr color space after conversion in step (2);

[0072] (4) Separate the Y component image from step (2);

[0073] (5) Set the image noise intensity control threshold σ0 and the image smoothness control threshold η0;

[0074] (6) Set the upsampling template filtering number Θ and the upsampling template filtering scale (2n+1)*(2n+1) for the nonlocal mean filtering and initialize them;

[0075] (7) Measure the noise intensity, smoothness, and image smoothness threshold of the noisy color image.

[0076] (8) Calculate the corresponding filtering parameters based on the noise intensity and smoothness of the Y image under the current filtering and the image smoothness threshold; call the fast nonlocal mean filtering algorithm to perform noise reduction filtering on the Y image;

[0077] (9) Measure the noise intensity and smoothness of the noisy color image after denoising in step (8);

[0078] (10) Determine if σ≥σ0 or η≥η0; if yes, proceed to (11); otherwise, proceed to step (13).

[0079] (11) Save the currently filtered image;

[0080] (12) Update the upsampling template filtering iteration Θ and the noise reduction upsampling template filtering scale, in the following way: In the second round of denoising, the upsampling template filtering count Θ in the fast nonlocal mean filtering algorithm can be increased by 1 and the denoising upsampling template filtering scale can be increased to (2n+3)*(2n+3). Then, the return value is used to perform cyclic denoising processing in step (8). The process continues until the judgment conditions in step (9) are not met, and then the process proceeds to step (13).

[0081] (13) Determine if the upsampling template filtering count Θ is 0 at this time; if it is, it means that only one filtering operation has been performed at this time, and the denoised Y component is directly output; otherwise, it means that more than one denoising operation has been performed, and the denoised Y component is output after accumulating and averaging all the saved filtered images. The specific scheme for accumulating and averaging all filtered images is to add and average the pixel values, and obtain a better result by accumulating and averaging.

[0082] (14) After merging the Y component, Cb and Cr components and converting them into the RGB color space, the noise-reduced image is obtained.

[0083] The specific implementation schemes for each step will be described in detail below:

[0084] A method for denoising color images under extremely low illumination includes the following steps:

[0085] S1. Convert the source color image A from the RGB color space to the YCbCr space, retaining the blue chromaticity component Cb and the red chromaticity component Cr.

[0086] S2. Extract the converted luminance component Y;

[0087] S3. Perform three-dimensional randomization on the brightness component image to generate a random (Poisson) noise image;

[0088] S4. Calculate the noise intensity of the noisy luminance component image.

[0089] S5. Calculate the image smoothness of the image containing the noisy luminance component;

[0090] S6. Calculate the filtering parameters and perform noise reduction using fast nonlocal mean filtering;

[0091] S7. Select noise reduction output or continue noise reduction according to the control strategy; determine whether noise reduction has ended. If yes, proceed to step S8; otherwise, return to S4 to continue execution.

[0092] S8. Merge the noise-reduced luminance component with the original blue and red chrominance components, and convert the image to the RGB color space.

[0093] Step S1 specifically includes:

[0094] S11. Obtain the source color image A;

[0095] S12. Convert the source color image A from the RGB color space to the YCbCr color space. First, normalize the RGB image values ​​to the [0,1] interval.

[0096] R'=R / 255; G'=G / 255; B'=B / 255 (1)

[0097] Where R, G, and B represent the red component, green component, and pixel value of the original color image, respectively. Then, by substituting into formulas (2)-(4), the luminance component Y, the blue chromaticity component Cb, and the red chromaticity component Cr are obtained, respectively.

[0098] Y=0.257*R+0.564*G+0.098*B+16 (2)

[0099] Cb=-0.148*R-0.291*G+0.439*B+128 (3)

[0100] Cr=0.439*R-0.368*G-0.071*B+128 (4)

[0101] like Figure 1 The diagram shows the principle of three-dimensional randomization. Step S3 of three-dimensional randomization specifically includes:

[0102] S31. Randomized Pixel Coordinates. Let the resolution supported by the luminance component Y image be M rows and N columns, where x0 and y0 are the row and column coordinates of a random initial point in the image, and x... i and y i Given the row and column coordinates of the i-th random point generated iteratively from a random initial point, randomized pixel coordinates are generated using the linear congruential method: x0 = (c1 * Seed) x +disp1)mod N

[0103] x i =(c1*x i-1 +disp1)mod N (5)

[0104]

[0105] Where c1 and c2 are coefficients, which can be determined by the values ​​of M and N, and are usually taken as 1 / 2. and Seed x and Seed y These are the two-dimensional Cartesian coordinates of the image seed point Seed. disp1 and disp2 are the minimum intervals between random numbers along the x-axis and y-axis, respectively, and are generally set to 1.

[0106] S32. Randomizing Pixel Values. Assume that after step S31, the randomized pixel coordinates (x, y) of the luminance component Y are obtained, where f(x0, y0) is the initial amplitude of the image at point (x0, y0), and f... p (x, y) represents the p-th random amplitude generated iteratively from the initial amplitude. Then, randomized pixel values ​​are generated using the linear congruential method:

[0107]

[0108] Where c3 is a coefficient, usually taken as 25. Seed f(x,y) `disp3` represents the seed point value within the range of image amplitude values. `disp3` is the minimum interval between random amplitude numbers, typically set to 1. Three-dimensional randomization is used to theoretically generate color images that more closely resemble real Poisson noise. Three-dimensional randomization is performed only before the first round of filtering; subsequent rounds of noisy Y-image processing only require multiple rounds of denoising.

[0109] Step S4 specifically includes:

[0110] S41. Calculate the noise intensity of the noisy luminance component image. Assuming the luminance component Y is obtained after step S31, first divide the image into overlapping blocks, and calculate the covariance matrix of the L blocks.

[0111]

[0112] Among them, P i Represents the i-th image patch. It is the mean of the L image patch vectors.

[0113] S42. Calculate the local variance of image patches. The local variance of image patch P is

[0114]

[0115] Among them, y j This represents the value of the j-th pixel in image block P. It is the average of all pixel values ​​in image block P.

[0116] S43. Find the best smooth image patch. Randomly select 10 image patches from the noisy luminance component Y, and set the image patch with the smallest local variance as the best smooth image patch.

[0117] S44. Construct the covariance matrix ∑ using the selected smooth image patches. For details, please refer to formula (8). Then calculate the minimum eigenvalue λ of this matrix. min Finally, the noise intensity is estimated using equation (10).

[0118]

[0119] Where L is the number of smoothing blocks used to construct the covariance matrix ∑.

[0120] Step S5 specifically includes:

[0121] S51. Calculate the image smoothness of the noisy luminance component image. Assuming the luminance component Y is obtained after step S31, then select the Gaussian filter kernel:

[0122]

[0123] S52. Then, convolve the Gaussian filter kernel with the luminance component and take the absolute value to obtain the filtered image Res:

[0124] Res=|k*Y| (12)

[0125] Where * represents convolution. The image smoothness can be calculated using formula (13):

[0126]

[0127] Step S6 specifically includes:

[0128] S61. Calculate the filtering parameters. Since both the noise intensity and image smoothness of the noisy image have a significant impact on the filtering parameters, a weighted approach is used to design the filtering parameters. Assume that the noise intensity is obtained through steps S43 and S52 respectively. Given the image smoothness η, the filtering parameter h is set by equation (14):

[0129]

[0130] Where α is the weighting coefficient, which is usually set to 0.5.

[0131] S62. Create a small search window. Assume the row is x1 and the column is x2, and the coordinates of a point within the window are z = (z1, z2). Then the region of the small window can be represented as Ω = {z = (z1, z2) ∈ N}. 2 Let v be the neighborhood of the region, and construct the integral image S of the difference between the pixel values. t ,and

[0132]

[0133] Where s t (z)=||v(z)-v(z+t)|| 2 S t (i)=S t (x1,x2)+S t (x1-1,x2)-S t (x1,x2-1)-S t (x1-1,x2-1).

[0134] S63. The formula for calculating the distance between neighbors v(i) and v(j) is as follows:

[0135]

[0136] Where dh Let d be the radius of the neighborhood window, d = 2 × d h The diameter of the neighborhood window.

[0137] S64. Calculate the weighted average between neighborhoods v(i) and v(j):

[0138]

[0139] Where Gen(i) is the normalization constant, and

[0140]

[0141] In the formula Let w(i,j) be the Gaussian weighted Euclidean distance based on gray level between the two neighborhoods of pixels i and j, and h be the filtering parameters calculated by step S61, and ∑w(i,j)=1.

[0142] S65. The noise reduction grayscale value of pixel i is calculated using the nonlocal mean algorithm according to formula (19):

[0143]

[0144] Step S7 specifically includes:

[0145] S71. Set the initial parameters of the algorithm. The image noise intensity threshold σ0 and the image smoothness threshold η0 reflect the control requirements. Based on statistical data in extremely low illumination environments, σ0 = η0 = 1 is generally chosen. The upsampling template filtering times Θ and the upsampling template filtering scale n (which can form a (2n+1)×(2n+1) neighborhood) realize the controllability of the algorithm. During initialization, Θ = 0 and n = 1, so the minimum window scale is 3×3.

[0146] S72. Measure the denoised image obtained in step S65 using steps S4 and S5 to obtain the denoised noise intensity. And the image smoothness index η, then make a judgment.

[0147]

[0148] Obviously, if any inequality in equation (20) holds, it means that the filtering has failed to achieve the final effect. Therefore, it is necessary to save the current noise reduction result and let

[0149]

[0150] Since the upsampling template filtering scale has become n+1, a neighborhood of (2n+3)×(2n+3) can be formed to prepare for the next filtering. At the same time, the algorithm will jump to step S6 to continue execution.

[0151] S73. After judgment, it was found that none of the inequalities in equation (20) were valid, indicating that the filtering had met the control requirements. At this time, it was determined whether the number of upsampling template filtering times was 0. If it was 0, it meant that the requirements were met after only one filtering, so there were no previously retained filtered images, and the current denoised image could be output directly. If it was not 0, it meant that more than one filtering was performed, so the current denoised image and the previously retained denoised image needed to be accumulated and averaged.

[0152] Step S8 specifically includes:

[0153] S81. Merge the denoised luminance component with the original blue and red chrominance components, and convert the image to the RGB color space. After obtaining the denoised luminance component, merge the blue chrominance component Cb and the red chrominance component Cr of the original low-light color image, and then convert it from the YCbCr color space to the RGB color space. The conversion method is as shown in equations (21)-(23):

[0154] R=1.164*(Y-16)+1.596*(Cr-128) (22)

[0155] G=1.164*(Y-16)-0.392*(Cb-128)-0.813*(Cr-128) (23)

[0156] B=1.164*(Y-16)+2.017*(Cb-128) (24)

[0157] The denoised Y component is then merged with the saved Cr and Cb components and converted into an image in the RGB color space, which is the output denoised image.

[0158] The key and protected aspects of this application are that, under extremely low illumination, it automatically selects filtering parameters for rapid non-local mean filtering and noise reduction based on the noise intensity and smoothness of the luminance component that is sensitive to the human eye. Then, based on a control threshold, it can select output or further filtering to ultimately complete the filtering and noise reduction task by meeting the control threshold requirements. This application has the following advantages:

[0159] 1. Converting an RGB image to a YCbCr image retains the blue chromaticity components Cb and Cr unchanged, and only reduces noise in the luminance component Y, which is sensitive to the human eye. Therefore, the process is faster and maintains less color distortion in the enhanced image.

[0160] 2. The self-feedback controllable nonlocal mean filter can adaptively feed back filtering parameters based on the image's noise intensity and smoothness, and automatically select the number of filtering iterations according to the control threshold until the filtering requirements are met. Fewer filtering iterations result in faster operation; more filtering iterations result in better filtering effects, which entirely depends on the user's needs.

[0161] 3. By adopting fast nonlocal mean filtering as the core, the image filtering speed is accelerated, which effectively improves the image processing speed in extremely low noise environments.

[0162] 4. Compared with other methods, this patent maintains good performance indicators in terms of peak signal-to-noise ratio and structural similarity of images, as detailed in Tables 1 and 2.

[0163] To demonstrate that the noise reduction effect of the method described in this application is better, this application verifies the effect through experimental comparison, and the results are shown in the following two tables:

[0164] Table 1 Comparison of objective indicators among different algorithms

[0165]

[0166]

[0167] Table 2. Objective performance test results for the self-feedback controllable nonlocal mean filtering noise reduction process (random noise, noise intensity 0.04).

[0168]

[0169] As shown in Tables 1 and 2, the denoising method of this application maintains good performance in terms of peak signal-to-noise ratio (PSNR) and structural similarity. From an objective perspective, PSNR and structural similarity are the most convincing indicators for image denoising. A comparison of PSNR and structural similarity demonstrates that the denoising method of this application performs better.

[0170] Obviously, the specific implementation of this invention is not limited to the above-described methods. Any non-substantial improvements made using the inventive concept and technical solution of this invention are within the protection scope of this invention.

Claims

1. A method for denoising color images under extremely low illumination, characterized in that: Includes the following steps: S1. Convert the source color image from RGB space to YCbCr space, and save the converted blue chromaticity component Cb image and red chromaticity component Cr image respectively; extract the converted luminance component Y. S2. Measure the image noise intensity and image smoothness of the luminance component Y; S3. Obtain the filtering parameters for nonlocal mean filtering based on the image noise intensity and image smoothness of the luminance component Y; and use nonlocal mean filtering to denoise the image of the luminance component Y based on the obtained filtering parameters to obtain the denoised luminance component Y. S4. Determine whether noise reduction is complete. If not, return to step S2; otherwise, proceed to step S5. S5. After merging the image of the luminance component Y after noise reduction with the blue chromaticity component Cb image and the red chromaticity component Cr image saved in step S1, the merged image is converted to the RGB color space. Step S2 includes: S21. Perform three-dimensional randomization on the Y-image of the brightness component to generate a noisy brightness component image; S22. Calculate the noise intensity of the noisy luminance component image; S23. Calculate the image smoothness of the image containing the noisy luminance component; In step S3, the method for obtaining the filter parameters of the nonlocal mean filtering algorithm includes: The filter parameter h is calculated using the following formula: ; in These are the weighting coefficients. For noise intensity and Image smoothness; Step S23, calculating the image smoothness of the noisy luminance component image, includes: Obtain the noisy luminance component Y image after processing in step S21, select a Gaussian filter kernel k, convolve the Gaussian filter kernel with the luminance component and take the absolute value to obtain the filtered image Res: ,in, For convolution; Calculate image smoothness using the following formula: Where M and N are the number of rows and columns of the resolution supported by the luminance component Y image, respectively; Res is the filtered image; Step S21, which involves three-dimensional randomization of the luminance component Y image, includes: Randomized pixel coordinates: Let the resolution supported by the Y-component image be... OK The column uses a linear congruential method to generate randomized pixel coordinates: ； ； in, and As a coefficient, it can be derived from and The value determines; and Seed points of the image Two-dimensional Cartesian coordinates; and , respectively, represent the minimum interval between random numbers along the x-axis and y-axis; x0 and y0 are the row and column coordinates of a random initial point in the image, x i and y i Let i be the row and column coordinates of the i-th random point generated through iteration from a random initial point; Randomized pixel values: After randomizing the pixel coordinates, the randomized pixel coordinates of the luminance component Y are obtained as follows: Then, the randomized pixel values ​​are generated using the linear congruential method: ； in, For coefficients, The seed point values ​​are the image amplitude values ​​within their range. The minimum interval between random numbers of amplitude; For seed point The generated initial random amplitude, and f p (x,y) is the p-th random amplitude generated by iterative generation from the initial amplitude.

2. The method for denoising color images under extremely low illumination as described in claim 1, characterized in that: Step S4 includes: S41. Calculate the noise intensity based on the luminance component Y after noise reduction processing in step S3. and image smoothness ; S42. The noise intensity calculated in step S41 is... and image smoothness With a preset image noise intensity threshold and image smoothness threshold Compare; when If any condition is not met, it is determined that the noise reduction is not finished and the process returns to step S2; otherwise, it is determined that the noise reduction is finished and the process proceeds to step S5.

3. The method for denoising color images under extremely low illumination as described in claim 1, characterized in that: Step S5 includes: After obtaining the luminance component for noise reduction, the blue chromaticity component Cb and the red chromaticity component Cr of the original extremely low-light color image are merged, and then converted from the YCbCr color space to the RGB color space. The conversion method uses the following formula to calculate the converted image information: ； ； 。 4. The method for denoising color images under extremely low illumination as described in claim 1, characterized in that: Step S1 includes: S11. Obtain the source color image A to be denoised; S12. Convert the source color image A from the RGB color space to the YCbCr color space: Normalize the RGB image values ​​to the [0,1] interval, and then substitute the normalized RGB image values ​​into the following formula to obtain the corresponding luminance component Y, blue chromaticity component Cb, and red chromaticity component Cr: ； ； 。

Images