Correction method for reducing non-uniformity of CMOS based on SVD

Abstract

The invention relates to the technical field of image sensor correction, solves the technical problems of non-uniformity correction calibration parameter quantity and large non-uniformity noise existing in a traditional correction method, and particularly relates to a correction method for reducing CMOS non-uniformity based on SVD (Singular Value Decomposition). A least square method is introduced to carry out non-uniformity linear correction on each pixel point to obtain correction parameters, then an SVD decomposition method is introduced, main characteristics of the parameters are intercepted to carry out dimensionality reduction, data after dimensionality reduction are stored, and approximate original correction parameter data are reconstructed according to singular vectors after dimensionality reduction in actual application. According to the non-uniformity correction method, the non-uniformity correction effect of the CMOS image sensor is kept, and meanwhile, millions of correction parameters which are possibly needed originally are reduced to thousands of correction parameters, so that hardware resources are greatly saved, hardware implementation is easy, and the real-time processing requirement is met.

Description

Technical Field

[0001] The present invention relates to the technical field of image sensor correction, and in particular to a correction method for reducing CMOS non-uniformity based on SVD. Background Art

[0002] CMOS image sensors have obvious advantages in power consumption, integration, cost performance, high speed, etc., and have been widely used in many fields in recent years. However, due to the unevenness of manufacturing materials, limitations of process conditions, differences in readout circuits, etc., CMOS generally has varying degrees of non-uniform noise.

[0003] Several patents provide methods for correcting the non-uniformity of image sensors. For example, the invention with publication number CN115144010A proposes a unified correction model, which realizes the test calibration and correction of the non-uniformity and linearity of the image sensor response. The invention with publication number CN112291446B relates to a method for correcting the non-uniformity of a large-array CMOS image sensor. First, the original image of the CMOS image sensor is collected under uniform illumination of an integrating sphere, and the non-uniformity of the collected image is calculated to obtain an image corresponding to the maximum non-uniformity. Then, according to the structural characteristics of the CMOS image sensor, the columns of the large-array CMOS image are grouped and processed, and the parameters are obtained by grouping, and finally a correction model for correcting the non-uniformity of the CMOS image sensor is established. The invention with publication number CN115665579A calculates the offset of the image at different saturations offline. When calculating online, only a standard correction matrix needs to be known. The correction matrix parameters at the saturation can be obtained through the offset at different saturations. In the process of correction calculation, linear interpolation, nonlinear interpolation, and the nearest neighbor principle are introduced to calibrate the standard parameter matrix to increase the correction accuracy. When selecting a non-uniformity correction method, factors such as the specific application scenario, processing speed, and hardware cost need to be considered to achieve the best correction effect.

[0004] However, due to the unevenness of manufacturing materials, limitations of process conditions, differences in readout circuits, etc., CMOS generally has varying degrees of non-uniform noise. Therefore, in order to improve imaging quality, non-uniformity correction of CMOS has become an inevitable issue. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention provides a correction method for reducing CMOS non-uniformity based on SVD, which solves the technical problems of large non-uniformity correction calibration parameter quantity and non-uniformity noise in traditional correction methods.

[0006] To solve the above technical problems, the present invention provides the following technical solution: a correction method for reducing CMOS non-uniformity based on SVD, the method comprising the following steps:

[0007] S1. Obtain s original images I with different grayscales from the CMOS image sensor under uniform illumination of an integrating sphere s , and record each original image I s The exposure time exp s And the grayscale change value gray corresponding to the M×N pixels of the CMOS image sensor s ;

[0008] S2, according to the exposure time exp s and grayscale change value gray s For the original image I s The mean linear response fitting is performed on several pixel values ​​in the image to obtain the correction parameters (K c ,B c );

[0009] S3, calibration parameters (K c ,B c ) to perform SVD decomposition and set the pass rate Rate interception correction parameter (K c ,B c ) is used to reduce the dimension of the main features to obtain the correction parameters after dimension reduction (U k , S k ′、V k ′;U b , S b ′、V b ′);

[0010] S4, calibration parameters (U k , S k ′、V k ′;U b , S b ′、V b ′) to reconstruct the parameters and obtain the reconstructed correction slope matrix K cp And the correction bias matrix B cp ;

[0011] S5. Using the correction slope matrix K cp and the correction bias matrix B cp For the original image I s Perform non-uniformity correction to obtain the corrected image img c ;

[0012] S6, correcting the image img according to the non-uniformity value NU of the CMOS image sensor c Conduct evaluation;

[0013] If the non-uniformity value NU is less than the threshold T, the correction slope matrix K is output cp And the correction bias matrix B cp and ends;

[0014] On the contrary, the process returns to step S3 to increase the passing rate Rate.

[0015] Furthermore, in step S2, the specific process includes the following steps:

[0016] S21, exposure time exp s is the independent variable x i As the horizontal axis, the degree change value gray s is the true value y i Draw the light response curve y of each pixel of the CMOS image sensor as the ordinate i =kx i +b;

[0017] S22, introduce the least squares method, and calculate the light response curve y according to the principle of minimum residual square sum i =kx i +b’s linear fitting obtains the fitting slope k and intercept b of each pixel;

[0018] S23, calculate each original image I s The grayscale mean avg i ;

[0019] S24, exposure time exp s is the horizontal axis, grayscale mean avg i Draw the mean linear response curve y for the ordinate i =KMx i +BM;

[0020] S25, linear response curve y to the mean i =KMx i +BM for linear fitting to obtain the fitting slope KM and intercept BM, where:

[0021]

[0022] Where S represents the number of observation points; exp s Indicates exposure time;

[0023] S26, with mean linear response curve y i =KMx i +BM is the target response to light response curve y i =kx i+b is corrected to obtain the correction slope matrix K consistent with the resolution of the CMOS image sensor c And the corrected intercept matrix B c .

[0024] Furthermore, in step S22, the calculation formulas for the fitting slope k and intercept b are:

[0025]

[0026] Where S represents the number of observation points; exp s Indicates exposure time; gray s Indicates the grayscale change value of the pixel.

[0027] Furthermore, in step S23, the grayscale mean avg i The calculation formula is:

[0028]

[0029] In the formula, I i (m,n) represents the grayscale value of the pixel in the mth row and nth column of the i-th original image; M represents the row height of the original image; and N represents the column of the original image.

[0030] Further, in step S26, the slope matrix K is corrected c And the corrected intercept matrix B c The calculation formula is:

[0031]

[0032] B c (i,j)=bk c *BM

[0033] In the formula, bk represents the uniformity correction slope matrix of the pixel in the i-th row and j-th column; c *BM represents the uniformity correction intercept matrix of the pixel in the i-th row and j-th column.

[0034] Furthermore, in step S3, the specific process includes the following steps:

[0035] S31, setting the correction slope matrix and the correction intercept matrix K c =B c =U*S′*V′, where U is an M-order orthogonal matrix, V′ is an N-order orthogonal matrix, and S′ is an M×N rectangular diagonal matrix composed of non-negative diagonal elements arranged in descending order;

[0036] S32, determine the correction slope matrix K c And the corrected intercept matrix Bc The singular values ​​of

[0037] S33, sort the singular values ​​by size and keep the first m, and calculate the proportion of singular values ​​m i ,Right now:

[0038]

[0039] In the formula, S i is the singular value after decomposition;

[0040] S34, set the proportion m of singular values i Select the pass rate Rate for retaining the correction parameter characteristics, specifically:

[0041]

[0042] In the formula, p is the selected dimension feature;

[0043] S35, select features of 1 to p, p ≥ 1 dimensions according to the passing rate Rate to obtain the correction parameters after dimensionality reduction (U k , S k ′、V k ′;U b , S b ′、V b ′).

[0044] Furthermore, in step S4, specifically:

[0045] Set the first p columns of the orthogonal matrix U to U k and the first p values ​​of the diagonal matrix S′ k ′, we get the left singular vector U weighted by the singular value. k *S k ′, and then with the first p rows of the orthogonal matrix V′ k ′, we can get U k *S k ′*V k ',Right now:

[0046] K cp =U k *S k ′*V k '

[0047] Similarly, the correction bias matrix B cp for:

[0048] B cp =U b *S b ′*V b '

[0049] Where U k , S k ′、V k ′、U b , S b ′、V b ′ are the correction parameters after dimensionality reduction.

[0050] Further, in step S5, the corrected image img c for:

[0051] img c =img*K cp +B cp

[0052] Where img is the original image I s .

[0053] Furthermore, in step S6, the calculation formula of the non-uniformity value NU is:

[0054]

[0055] Where V mean represents the average grayscale of the original image; M represents the number of pixels of the CMOS image sensor; V i Represents the gray value of the i-th pixel.

[0056] By means of the above technical solution, the present invention provides a correction method for reducing CMOS non-uniformity based on SVD, which has at least the following beneficial effects:

[0057] 1. While retaining the non-uniformity correction effect of the CMOS image sensor, the present invention reduces the number of correction parameters that may be originally required from millions to thousands, greatly saving hardware resources, facilitating hardware implementation, and meeting real-time processing requirements.

[0058] 2. The present invention introduces an SVD decomposition method for dimensionality reduction of correction parameters. In dimensionality reduction, SVD is used to find the main features in the data and project the data into a lower dimensional space while retaining as much information as possible, thereby saving a lot of hardware resources and ensuring the effect of non-uniformity correction. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0060] Figure 1 is a flow chart of the correction method in the present invention;

[0061] Figure 2is a curve chart of the proportion of singular values ​​in the present invention;

[0062] Figure 3 It is a comparison diagram before and after correction in the present invention. DETAILED DESCRIPTION

[0063] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific implementation methods, so that the implementation process of how the present application uses technical means to solve technical problems and achieve technical effects can be fully understood and implemented accordingly.

[0064] CMOS (Complementary Metal Oxise Semiconductor) image sensors have obvious advantages in power consumption, integration, cost performance, high speed, etc., and have been widely used in many fields in recent years. However, due to the unevenness of manufacturing materials, limitations of process conditions, differences in readout circuits, etc., CMOS generally has different degrees of non-uniform noise. Therefore, in order to improve the imaging quality, non-uniformity correction of CMOS has become an inevitable problem.

[0065] Common non-uniformity correction methods for image sensors include two-dimensional lookup table (2D LUT), multiple exposure correction, etc. The number of parameters in conventional calibration methods is determined by the CMOS resolution or structural characteristics. For example, a CMOS sensor with a resolution of 1280x1024 has a total of 1280*1024=1310720 pixels, which requires 2*1310720 linear correction parameters. The number of parameters is in the millions, which consumes a lot of hardware resources and is not conducive to hardware implementation.

[0066] In order to reduce the amount of CMOS non-uniformity correction calibration parameters and keep the CMOS non-uniformity correction effect as much as possible, please refer to Figure 1-Figure 3 This embodiment proposes a correction method based on SVD to reduce CMOS non-uniformity. By introducing the least squares method, the multi-exposure integrating sphere image is used to obtain the CMOS uniformity correction parameters, combined with the SVD decomposition method, and the main features are selected according to the proportion of the parameters to achieve parameter dimension reduction, which is much smaller than the original correction parameter amount, and then the matrix after dimension reduction is reconstructed to obtain the CMOS pixel-by-pixel correction parameters to achieve CMOS image sensor non-uniformity correction. This can save hardware resources, is easy to implement, and has been proven to significantly improve image uniformity. The method includes the following steps:

[0067] S1. Obtain s original images I with different grayscales from the CMOS image sensor under uniform illumination of an integrating sphere s , and record each original image I s The exposure time exp sAnd the grayscale change value gray corresponding to the M×N pixels of the CMOS image sensor s In this embodiment, it is necessary to collect the original image I of the CMOS image sensor under uniform illumination of the integrating sphere and adjust the exposure time. s Until the image reaches saturation, we get s original images of different grayscales I1, I2, ..., I s , the original image I s The resolution is M×N, and each original image I is recorded at the same time s The exposure time exp1, exp2, …, exp s And the grayscale change values ​​of M×N pixels gray1, gray2, …, gray s .

[0068] S2, according to the exposure time exp s and grayscale change value gray s For the original image I s The mean linear response fitting is performed on several pixel values ​​in the image to obtain the correction parameters (K c ,B c ); In step S2, the specific process includes the following steps:

[0069] S21, exposure time exp s is the independent variable x i As the horizontal axis, the degree change value gray s is the true value y i Draw the light response curve y of each pixel of the CMOS image sensor as the ordinate i =kx i Specifically, in this embodiment, a light response curve is drawn for each pixel of the CMOS image sensor, and the horizontal axis x is the exposure time exp1, exp2, ..., exp s , the vertical coordinate y is the grayscale change value of the pixel gray1, gray2, ..., gray s .

[0070] S22, introduce the least squares method, and calculate the light response curve y according to the principle of minimum residual square sum i =kx i +b, the linear fitting of each pixel obtains the fitting slope k and intercept b; among them, the principle of minimizing the residual sum of squares is the optimal parameter principle in linear fitting, and the calculation formula of the residual sum of squares P is:

[0071]

[0072] In the formula, P represents the residual sum of squares of all observation points S, which is the objective function to be minimized in the least squares method and is used to measure the fitting effect of the model. The smaller the residual sum of squares, the closer the predicted value of the model is to the actual value; y i represents the actual value (true value) of the i-th observation point, which is the real data obtained through experiments or observations, that is, the grayscale change values ​​gray1, gray2, ..., gray s ;x i represents the independent variable, i.e., the exposure time exp1, exp2, …, exp s ; a represents the linear regression coefficient, and b represents the intercept.

[0073] The CMOS sensor has M×N pixels (M represents the original image I s The row height of N represents the original image I s The column width of each pixel is 1000, and the linear fitting operation is completed for each pixel. Then the slope matrix K and intercept matrix B can be obtained, with a size of M×N. Among them, K(i,j) and B(i,j) respectively represent the fitting slope k and intercept b of the linear fitting of the pixel in the i-th row and j-th column of the CMOS image sensor. The calculation formula is:

[0074]

[0075] Where S represents the number of observation points; exp s Indicates exposure time; gray s Indicates the grayscale change value of the pixel.

[0076] S23, calculate each original image I s The grayscale mean avg i , the calculation formula is:

[0077]

[0078] In the formula, I i (m,n) represents the grayscale value of the pixel in the mth row and nth column of the i-th original image; M represents the row height of the original image; N represents the column width of the original image, that is, the total number of pixels of the CMOS image sensor is M×N;

[0079] S24, exposure time exp s is the horizontal axis, grayscale mean avg i Draw the mean linear response curve y for the ordinate i =KMx i +BM;

[0080] S25, linear response curve y to the mean i =KMx i+BM for linear fitting to obtain the fitting slope KM and intercept BM, where:

[0081]

[0082]

[0083] Where S represents the number of observation points; exp s Indicates exposure time;

[0084] S26, with mean linear response curve y i =KMx i +BM is the target response to light response curve y i =kx i +b is corrected to obtain the correction slope matrix K consistent with the resolution of the CMOS image sensor c And the corrected intercept matrix B c , which is K c (i,j),B c (i,j).

[0085] In this embodiment, it is assumed that the light response curve of the pixel in the i-th row and j-th column is y i =kx i +b, with mean linear response curve y i =KMx i +BM is the target response and the uniformity correction slope k of the pixel can be obtained. c and the corrected intercept b c ,Right now:

[0086]

[0087] b c =bk c *BM

[0088] Where k is the fitting slope of the linear fit; b is the intercept of the linear fit.

[0089] Furthermore, by completing the above processing for each pixel, a correction slope matrix K consistent with the resolution of the CMOS image sensor can be obtained. c And the corrected intercept matrix B c , the size is M×N. K c (i,j),B c (i, j), respectively represents the correction slope matrix k of the pixel in the i-th row and j-th column of the CMOS image sensor c And the correction intercept matrix b c ,Right now:

[0090]

[0091] B c (i,j)=bk c *BM

[0092] In the formula, bk represents the uniformity correction slope matrix of the pixel in the i-th row and j-th column; c *BM represents the uniformity correction intercept matrix of the pixel in the i-th row and j-th column.

[0093] CMOS non-uniformity correction full parameter correction processing, namely:

[0094] Ic=I*K c +B c

[0095] Where I is the original image and Ic is the corrected image.

[0096] S3, calibration parameters (K c ,B c ) to perform SVD decomposition and set the pass rate Rate interception correction parameter (K c ,B c ) is used to reduce the dimension of the main features to obtain the correction parameters after dimension reduction (U k , S′ k , V′ k ; U b , S′ b , V′ b ); In step S3, the specific process includes the following steps:

[0097] S31, setting the correction slope matrix and the correction intercept matrix K c =B c =U*S′*V′, where U is an M-order orthogonal matrix, V′ is an N-order orthogonal matrix, and S′ is an M×N rectangular diagonal matrix composed of non-negative diagonal elements arranged in descending order;

[0098] S32, determine the correction slope matrix K c And the corrected intercept matrix B c The singular values ​​of , where the elements on the diagonal of the M×N rectangular diagonal matrix are the singular values. The size of the singular value indicates the importance of the feature in the original data. The larger the singular value, the more significant the impact of the corresponding feature in the data.

[0099] S33, sort the singular values ​​by size and keep the first m, and calculate the proportion of singular values ​​m i ; In this step, in order to retain the correction parameters (K c ,B c), the basic idea of ​​dimensionality reduction of correction parameters is to retain features with larger singular values ​​and ignore features with smaller singular values. Generally speaking, singular values ​​are arranged in order of size, so we can choose to retain the first m singular values, where m is a small integer representing the desired dimension after dimensionality reduction. Assume that K c The first 10 singular values ​​after decomposition (S1, S2, ..., S 10 ), then the proportion of the i-th singular value is:

[0100]

[0101] The singular value ratio curve drawn in this example is as follows Figure 2 shown.

[0102] S34, set the proportion m of singular values i Select the pass rate of the retained correction parameter features, select the retained features according to the singular value ratio, and perform truncation. In order to ensure the uniformity correction effect as much as possible, it is necessary to retain the correction parameter features as much as possible. For example, the singular values ​​after decomposition are S1, S2, …, S m , which accounts for:

[0103]

[0104] The pass rate is set as Rate (adjustable, such as 90%).

[0105]

[0106] In the formula, p is the selected dimension feature.

[0107] S35, select features of 1 to p, p ≥ 1 dimensions according to the passing rate Rate to obtain the correction parameters after dimensionality reduction (U k , S k ′、V k ′;U b , S b ′、V b ′), specifically, extract the first p components of the original data, namely U k =U(1:M,1:p),S k ′=S′(1:p,1:p),V k ′=V′(1:N,1:p), M and N are the row height and column width of the image respectively. Correction parameter K c The parameter is changed to U k , S k ′、V k ′, the number of parameters is p*(M+N+1), which is an order of magnitude lower than the previous M×N number of parameters. At the same time, when the correction parameter characteristics are more obvious, the lower the p value, the fewer the parameters. Similarly, for the correction intercept matrix Bc Select the first p eigencomponents and correct the intercept matrix B c The parameter is changed to U b , S b ′、V b ′, there are p*(M+N+1) parameters in total.

[0108] Among them, the left singular vectors (columns of the U matrix) represent the orthogonal basis of the row space of the original matrix, which reflects the distribution of the original data in different dimensions and can be regarded as the projection direction of the data in the new coordinate system after dimensionality reduction. The right singular vectors (rows of the V′ matrix) represent the orthogonal basis of the column space of the original matrix, reflecting the correlation between the features of the original data, and are often used for feature extraction after dimensionality reduction.

[0109] S4, calibration parameters (U k , S k ′、V k ′;U b , S b ′、V b ′) to reconstruct the parameters and obtain the reconstructed correction slope matrix K cp And the correction bias matrix B cp , specifically matrix multiplication reconstruction, first calculate U k *S k ′, set the first p columns of the orthogonal matrix U to U k and the first p values ​​of the diagonal matrix S′ k ′, we get the left singular vector U weighted by the singular value. k *S k ′, and then with the first p rows of the orthogonal matrix V′ k ′, we can get U k *S k ′*V k ',Right now:

[0110] K cp =U k *S k ′*V k '

[0111] Similarly, the correction bias matrix B cp for:

[0112] B cp =U b *S b ′*V b '

[0113] Where U k , S k ′、V k ′、U b , Sb ′、V b ′ are the correction parameters after dimensionality reduction.

[0114] The reconstructed correction slope matrix K cp And the correction bias matrix B cp , compared to the correction slope matrix K c And the corrected intercept matrix B c In terms of , the number of parameters that need to be stored is reduced from M×N to p*(M+N+1).

[0115] S5. Using the correction slope matrix K cp and the correction bias matrix B cp For the original image I s Perform non-uniformity correction to obtain the corrected image img c ,Right now:

[0116] img c =img*K cp +B cp

[0117] Where img is the original image I s ;

[0118] S6, correcting the image img according to the non-uniformity value NU of the CMOS image sensor c Conduct evaluation;

[0119] If the non-uniformity value NU is less than the threshold T, the correction slope matrix K is output cp And the correction bias matrix B cp and ends;

[0120] On the contrary, the process returns to step S3 to increase the passing rate Rate.

[0121] Specifically, it is also necessary to evaluate the images before and after the non-uniformity correction to verify that the scheme is effective. The commonly used standard root mean square deviation represents the non-uniformity NU of the CMOS image sensor, and the expression is:

[0122]

[0123] Where V mean represents the average grayscale of the original image; M represents the number of pixels of the CMOS image sensor; V i Represents the grayscale value of the i-th pixel. The lower the NU value, the lower the non-uniformity of the CMOS image sensor and the higher the image quality.

[0124] In this embodiment, the non-uniformity values ​​NU of the corrected images under different exposures are analyzed for the integrating sphere uniform light source of the CMOS image sensor. When the NU values ​​are all <2% (the threshold can be set by yourself), the parameters meet the non-uniformity correction requirements. If the NU value does not meet the requirements, it returns to step S3 to increase the pass rate Rate and use more correction parameters to improve image uniformity.

[0125] The scheme proposed by the present invention is used to obtain the correction parameters of the CMOS image sensor after SVD dimension reduction. Figure 3 a in is the original image, Figure 3 b in the figure is the 3D image corresponding to the original image. Figure 3 After the non-uniformity correction in a) in the figure, we get Figure 3 In the image c, the NU value is reduced from the original 3.652% to 1.688%, and the corresponding three-dimensional image is Figure 3 The d in the image is also flatter, and the image quality is significantly improved.

[0126] The present invention first collects different exposure images of the CMOS image sensor under uniform illumination of an integrating sphere, introduces the least squares method to perform linear non-uniformity correction on each pixel, obtains correction parameters, and then introduces the SVD decomposition method to intercept the main features of the parameters for dimensionality reduction, stores the reduced-dimensional data, and reconstructs the approximate original correction parameter data according to the reduced-dimensional singular vectors in actual application. While retaining the CMOS non-uniformity correction effect, this solution reduces the number of correction parameters that may be required from millions to thousands, greatly saving hardware resources, being easy to implement in hardware, and meeting real-time processing requirements.

[0127] Those skilled in the art can understand that all or part of the steps in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a program, so the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0128] Each embodiment in this specification is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the above embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments.

[0129] The above implementation methods have been described in detail. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

Claims

1. A correction method for reducing CMOS non-uniformity based on SVD, characterized in that: The method comprises the following steps: S1. Obtain s original images I with different grayscales from the CMOS image sensor under uniform illumination of an integrating sphere s , and record each original image I s The exposure time exp s And the grayscale change value gray corresponding to the M×N pixels of the CMOS image sensor s ; S2, according to the exposure time exp s and grayscale change value gray s For the original image I s The mean linear response fitting is performed on several pixel values ​​in the image to obtain the correction parameters (K c ,B c ); S3, calibration parameters (K c ,B c ) to perform SVD decomposition and set the pass rate Rate interception correction parameter (K c ,B c ) is used to reduce the dimension of the main features to obtain the correction parameters after dimension reduction (U k , S k ′、V k ′;U b , S b ′、V b ′); S4, calibration parameters (U k , S k ′、V k ′;U b , S b ′、V b ′) to reconstruct the parameters and obtain the reconstructed correction slope matrix K cp And the correction bias matrix B cp ; S5. Using the correction slope matrix K cp and the correction bias matrix B cp For the original image I s Perform non-uniformity correction to obtain the corrected image img c ; S6, correcting the image img according to the non-uniformity value NU of the CMOS image sensor c Conduct evaluation; If the non-uniformity value NU is less than the threshold T, the correction slope matrix K is output cp And the correction bias matrix B cp and ends; On the contrary, the process returns to step S3 to increase the passing rate Rate.

2. The calibration method according to claim 1, characterized in that: In step S2, the specific process includes the following steps: S21, exposure time exp s is the independent variable x i As the horizontal axis, the degree change value gray s is the true value y i Draw the light response curve y of each pixel of the CMOS image sensor as the ordinate i =kx i +b; S22, introduce the least squares method, and calculate the light response curve y according to the principle of minimum residual square sum i =kx i +b’s linear fitting obtains the fitting slope k and intercept b of each pixel; S23, calculate each original image I s The grayscale mean avg i ; S24, exposure time exp s is the horizontal axis, grayscale mean avg i Draw the mean linear response curve y for the ordinate i =KMx i +BM; S25, linear response curve y to the mean i =KMx i +BM for linear fitting to obtain the fitting slope KM and intercept BM, where: Where S represents the number of observation points; exp s Indicates exposure time; S26, with mean linear response curve y i =KMx i +BM is the target response to light response curve y i =kx i +b to obtain the correction slope matrix K consistent with the resolution of the CMOS image sensor c And the corrected intercept matrix B c .

3. The calibration method according to claim 2, characterized in that: In step S22, the calculation formulas for the fitting slope k and intercept b are: Where S represents the number of observation points; exp s Indicates exposure time; gray s Indicates the grayscale change value of the pixel.

4. The calibration method according to claim 2, characterized in that: In step S23, the grayscale mean avg i The calculation formula is: In the formula, I i (m,n) represents the grayscale value of the pixel in the mth row and nth column of the i-th original image; M represents the row height of the original image; and N represents the column of the original image.

5. The calibration method according to claim 3, characterized in that: In step S26, the slope matrix K is corrected c And the corrected intercept matrix B c The calculation formula is: B c (i,j)=bk c *BM In the formula, bk represents the uniformity correction slope matrix of the pixel in the i-th row and j-th column; c *BM represents the uniformity correction intercept matrix of the pixel in the i-th row and j-th column.

6. The calibration method according to claim 1, characterized in that: In step S3, the specific process includes the following steps: S31, setting the correction slope matrix and the correction intercept matrix K c =B c =U*S′*V′, where U is an M-order orthogonal matrix, V′ is an N-order orthogonal matrix, and S′ is an M×N rectangular diagonal matrix composed of non-negative diagonal elements arranged in descending order; S32, determine the correction slope matrix K c And the corrected intercept matrix B c The singular values ​​of S33, sort the singular values ​​by size and keep the first m, and calculate the proportion of singular values ​​m i ,Right now: In the formula, S i is the singular value after decomposition; S34, set the proportion m of singular values i Select the pass rate Rate for retaining the correction parameter characteristics, specifically: In the formula, p is the selected dimension feature; S35, select features of 1 to p, p ≥ 1 dimensions according to the passing rate Rate to obtain the correction parameters (U k , S k ′、V k ′;U b , S b ′、V b ′).

7. The calibration method according to claim 1, characterized in that: In step S4, specifically: Set the first p columns of the orthogonal matrix U to U k and the first p values ​​of the diagonal matrix S′ k ′, we get the left singular vector U weighted by the singular value. k *S k ′, and then with the first p rows of the orthogonal matrix V′ k ′, we can get U k *S k ′*V k ',Right now: K cp =U k *S k ′*V k ′ Similarly, the correction bias matrix B cp for: B cp =U b *S b ′*V b ′ Where U k , S k ′、V k ′、U b , S b ′、V b ′ are the calibration parameters after dimensionality reduction.

8. The calibration method according to claim 1, characterized in that: In step S5, the corrected image img c for: img c =img*K cp +B cp Where img is the original image I s .

9. The calibration method according to claim 1, characterized in that: In step S6, the calculation formula of the non-uniformity value NU is: Where V mean represents the average grayscale of the original image; M represents the number of pixels of the CMOS image sensor; V i Represents the gray value of the i-th pixel.

Images