A constrained correction method for non-uniformly illuminated multispectral imaging data

Abstract

The application provides a constraint correction method for non-uniform illumination multispectral imaging data, comprising the following steps: step one, imaging data introduction; step two, dark noise removal; step three, region adaptive non-uniform illumination image correction; and step four, global correction based on spectral prior constraint: the corrected multi-channel spectral image obtained in step three is taken as an integral input to increase the illumination non-uniform global correction model with spectral prior constraint, global correction is performed, and corrected multispectral imaging data is obtained. The constraint correction method for non-uniform illumination multispectral imaging data is best in subsequent cultural relic spectral image classification, and can provide reliable multispectral data for cultural relic research and protection analysis. The constraint correction method for non-uniform illumination multispectral imaging data can effectively correct the spectral drift problem caused by non-uniform illumination, and constrain the spectral correlation and consistency among multiple imaging channels.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of digital image processing, and relates to multispectral imaging data, in particular to a constraint correction method for multispectral imaging data under non-uniform illumination. BACKGROUND

[0002] Due to the complex on-site situation of cultural relics protection, in the actual spectral imaging process, the quality of multispectral imaging data is affected by imaging system equipment and image acquisition environment. The non-uniform photoelectric response of the imaging system sensor, the influence of the dark current of the equipment under long-time work, and the changeable and non-uniform illumination of the shooting environment can all cause the reliability of the multispectral imaging data to decrease, cause the spectral characteristics of the same substance in the multispectral imaging data of the multi-component lens to shift, and finally affect the accuracy of pigment detection and classification recognition.

[0003] At present, histogram equalization, function adjustment, frequency domain filtering and Retinex theory method are mostly used for image correction, but they ignore the spatial spectral correlation information and have different degrees of loss of gray information.

[0004] In recent years, deep learning models have also been introduced into the field of image illumination enhancement. Due to the uniqueness of cultural relics, different painted cultural relics are unique individuals, and the multispectral imaging data of painted cultural relics do not have a universal data set, and the data acquisition is very limited, which makes it difficult for deep learning models to fully train and generalize. SUMMARY

[0005] In view of the deficiencies in the prior art, the purpose of the present application is to provide a constraint correction method for multispectral imaging data under non-uniform illumination, which solves the technical problem that the reliability of the multispectral imaging data in the prior art needs to be further improved.

[0006] In order to solve the above technical problems, the present application adopts the following technical solutions:

[0007] A constraint correction method for multispectral imaging data under non-uniform illumination, the method comprising the following steps:

[0008] Step 1, introduction of imaging data.

[0009] Step 2, dark noise removal.

[0010] According to the environmental dark noise, the environmental dark noise is removed from the image to obtain an original multi-channel spectral image I(x, y).

[0011] Step 3, region adaptive non-uniform illumination image correction:

[0012] Step 301: Based on the original multi-channel spectral image I(x,y) obtained in Step 2, the original illumination component L(x,y) estimated by the guided filter is calculated using a multi-scale Retinex model based on the guided filter of local pixel linear relationship.

[0013] Step 302: Based on the original illumination component L(x,y) estimated by the guided filter obtained in step 301, the image is segmented into regions. For the brightness characteristics of each region in the image, a dynamically adjusted gamma correction method is used to achieve a balance between overly bright and overly dark regions, resulting in the adaptively gamma-corrected illumination component L. g (x,y).

[0014] The dynamically adjusted gamma correction method uses the gamma correction parameter γ(x,y).

[0015] The method for calculating the gamma correction parameter γ(x,y) is as follows:

[0016]

[0017] In the formula:

[0018] μ i The mean value of the illumination component in the i-th region;

[0019] σ i Let be the variance of the illumination component in the i-th region.

[0020] Step 303: Use the CLAHE algorithm to enhance the adaptive gamma-corrected illumination component L obtained in step 302. g The contrast and detail sharpness of (x,y) are obtained by enhancing the illumination component L after adaptive gamma correction. c (x,y).

[0021] Step 304: The enhanced adaptive gamma-corrected illumination component L obtained in step 303 is... c (x,y) and the original illumination component L estimated by the guided filter. g (x,y) are weighted and fused to obtain the final illumination correction component L. enhancement :

[0022] Step 305, based on illumination correction component L enhancement A corrected multichannel spectral image is obtained.

[0023] Step 4, Global Correction Based on Spectral Prior Constraints:

[0024] Step 401: Based on the original multi-channel spectral image I(x,y) obtained in Step 2, the standard spectral reflectance of the sampling points is introduced to construct a linear regression model.

[0025] Step 402, fuse the structure features and the spatial domain spectral features to construct a loss function L.

[0026] The expression of the loss function L is:

[0027]

[0028] In the formula:

[0029] α1 and α2 are regularization parameters for balancing the spatial structure item and the spectral gradient item of the loss function L;

[0030] is the spatial structure item of the loss function L;

[0031] is the spectral gradient item of the loss function L;

[0032] is the 16-dimensional standard spectral data in the i-th row;

[0033] y i is the 16-dimensional spectral data output by the linear regression model in the i-th row;

[0034] ‖·‖ F is the Frobenius norm;

[0035] SSIM(·) is a structure similarity calculation;

[0036] is the spectral gradient of the 16-dimensional standard spectral data in the i-th row;

[0037] is the 16-dimensional spectral gradient output by the linear regression model in the i-th row.

[0038] Step 403, based on the minimization of the loss function L, the linear regression model is optimized to obtain an optimal linear regression model, that is, an illumination non-uniform global correction model with spectral prior constraint.

[0039] The illumination non-uniform global correction model with spectral prior constraint is:

[0040]

[0041] In the formula:

[0042] I(x, y) is an original multi-channel spectral image;

[0043] f c (·) is a CLAHE algorithm operation;

[0044] f gid(·) is a multi-scale guided filter light component estimation operation;

[0045] W is a regression coefficient matrix of size N x N;

[0046] B is an n x N column vector;

[0047] n is the total number of samples;

[0048] N is the total number of channels;

[0049] gamma is a gamma correction parameter gamma(x,y).

[0050] Step 404, input the corrected multi-channel spectral image obtained in step three as a whole into the light uneven global correction model with increased spectral prior constraint, perform global correction, and obtain corrected multi-spectral imaging data.

[0051] Compared with the prior art, the present application has the following technical effects:

[0052] (I) The constraint correction method for non-uniformly illuminated multi-spectral imaging data of the present application is best in subsequent cultural relic spectral image classification, and can provide reliable multi-spectral data for the research and protection analysis of cultural relics.

[0053] (II) The constraint correction method for non-uniformly illuminated multi-spectral imaging data of the present application can effectively correct the spectral drift problem caused by uneven illumination and constrain the spectral correlation and consistency between multiple imaging channels.

[0054] (III) In the present application, the prior-supported multi-channel global correction adds a spatial structure and spectral gradient constraint term on the basis of non-uniform illumination correction, ensuring that the correlation and consistency of the original spectral features are maintained during illumination change correction, thereby maximizing the integrity of the multi-channel spectral data during correction of uneven illumination and ensuring that the corrected spectral data accurately reflects the true reflection characteristics of the material.

[0055] (IV) In the present application, the region-adaptive two-dimensional gamma correction method can effectively improve the local brightness distortion caused by non-uniform illumination and improve the overall contrast and detail performance of the image. At the same time, aiming at the requirement of accurate classification of spectral images, a regression correction model based on spectral prior constraint is constructed, and a loss function is constructed combining spatial structure features and spectral prior features to globally constrain the original multi-channel spectral image, achieving the purpose of correcting the spectral information deviation of the multi-channel image while maintaining the spectral correlation and consistency between multiple imaging channels.

[0056] (V) The method of the present application can effectively balance the image brightness and suppress the deviation of the image spectral information, which is conducive to improving the accuracy of subsequent pigment classification and recognition and maintaining the structural information of cultural relics. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 is the simulation mural correction result.

[0058] Figure 2 is the comparison of classification results before and after the simulation mural correction.

[0059] Figure 3 is the comparison of classification results before and after the real mural correction.

[0060] The specific content of the present application is further described in detail below in combination with the drawings and examples. DETAILED DESCRIPTION

[0061] It should be noted that the algorithms and methods in the present application all use known algorithms and methods in the prior art unless otherwise specified.

[0062] Based on the current situation introduced in the background art, the present application aims to solve the problem of spectral shift caused by non-uniform illumination in the multispectral imaging process of colored relics, and needs to use an adaptive multispectral image correction method combined with prior feature constraints to obtain a spectrum-corrected image for subsequent research.

[0063] In accordance with the above technical solution, the specific embodiments of the present application are given below. It should be noted that the present application is not limited to the following specific embodiments, and any equivalent transformation made on the basis of the technical solutions of the present application falls within the scope of protection of the present application.

[0064] Embodiment:

[0065] The present embodiment gives a constraint correction method for non-uniform illumination multispectral imaging data, which comprises the following steps.

[0066] Step one, introduction of imaging data:

[0067] Collect multispectral images of cultural relics, register and crop the multispectral images to obtain non-uniform illumination mural images, and introduce dark frame images under no light conditions as reference data.

[0068] In step one, the differences in optical transmittance of each part of the optical element in the imaging device, the differences in the response of the sensor to incident light, and the differences in the spectral distribution of the environmental light source will directly affect the radiation intensity distribution of the incident light reaching the photosensitive surface, causing spectral errors of dark current noise on the resulting image.

[0069] Step two, dark noise removal:

[0070] According to the dark frame image introduced in step one, estimate the environmental dark noise by laboratory calibration method, and remove the environmental dark noise from the non-uniform illumination mural image obtained in step one according to the environmental dark noise to obtain the original multi-channel spectral image I(x, y).

[0071] In step two, in order to correct the dark noise problem caused by the imaging system device characteristics, the dark frame image under no light condition is introduced, the dark noise correction is carried out using the laboratory calibration method, the random noise existing in a short time is eliminated, and finally the average dark current value of the imaging device under the action of no external light source is obtained.

[0072] In step two, according to the average processing of the dark frame output image of the imaging device collected continuously for a long time, the dark current value measured is deducted in the data processing process. Thus, the influence of the dark current on the image quality is weakened, and the image noise caused by the non-uniformity of the dark current is reduced. The calculation method of the environmental dark noise removal is:

[0073]

[0074] In the formula:

[0075] DN (i,j,λ) DN is the response value corresponding to the camera after removing the dark noise in the spectral channel with the center wavelength λ, the i-th row and the j-th column;

[0076] R is the response value of the camera in the spectral channel with the center wavelength λ, the i-th row and the j-th column;

[0077] DN andianliu(i,j,λ,k) DN is the dark current value of the pixel in the spectral channel with the center wavelength λ, the i-th row and the j-th column at the k-th time point;

[0078] N is the total number of observation time points.

[0079] Step three, region adaptive non-uniform illumination image correction:

[0080] In step three, the following steps are specifically included:

[0081] Step 301, according to the original multi-channel spectral image I(x, y) obtained in step two, the guided filter multi-scale Retinex model based on the local pixel linear relationship is used to calculate the original light component L(x, y) estimated by the guided filter.

[0082] In step 301, the multi-scale Retinex model is:

[0083]

[0084] In the formula:

[0085] L(x, y) is the original light component estimated by the guided filter;

[0086] i is the filter scale;

[0087] N is the number of filter scales,

[0088] w i The weight of different scales is 1 / 3.

[0089] I(x,y) is the guide image;

[0090] q i (x,y) is the guide filter function.

[0091] Step 302, based on the original illumination component L(x,y) estimated by the guide filter in step 301, the image is regionally segmented, as shown in Figure 1 For the brightness characteristics of each region in the image, a dynamically adjusted gamma correction method is used to achieve balanced processing of over-bright and over-dark regions, and to obtain the adaptive gamma-corrected illumination component L(x,y). g

[0092] In step 302, the dynamically adjusted gamma correction method is:

[0093]

[0094] In the formula:

[0095] L g (x,y) is the adaptive gamma-corrected illumination component;

[0096] L(x,y) is the original illumination component estimated by the guide filter;

[0097] γ(x,y) is the gamma correction parameter.

[0098] In step 302, the calculation method of the gamma correction parameter is:

[0099]

[0100] In the formula:

[0101] μ i is the mean of the illumination component in the i-th region;

[0102] σ i is the variance of the illumination component in the i-th region.

[0103] In this embodiment, the mean and variance characteristics of the region are introduced to calculate the gamma correction parameter γ(x,y); the gamma correction parameter γ(x,y) controls the intensity of brightness correction, and its value is crucial to the brightness adjustment effect of the image region.

[0104] Step 303, using the CLAHE algorithm to enhance the adaptive gamma-corrected illumination component L g ​(x, y) and the detail definition of the contrast of (x, y), further enhance the brightness of the illumination component, maintain the texture details, and obtain the enhanced adaptive gamma corrected illumination component L c (x, y).

[0105] In step 303, the calculation formula of the CLAHE algorithm enhancement is as follows:

[0106] L c (x, y) = f c (L g (x, y))

[0107] In the formula:

[0108] L c (x, y) is the enhanced adaptive gamma corrected illumination component;

[0109] L g (x, y) is the adaptive gamma corrected illumination component;

[0110] f c (·) is the CLAHE algorithm operation.

[0111] In step 304, the enhanced adaptive gamma corrected illumination component L c (x, y) obtained in step 303 is weighted and fused with the original illumination component L g (x, y) estimated by the guided filtering, to obtain the final illumination correction component L enhancement .

[0112] In step 304, the calculation method of the weighted fusion is as follows:

[0113] L enhancement = mL(x, y) + (1-m) L c (x, y).

[0114] In the formula:

[0115] m is a weight value, and the value range is 0 to 1.

[0116] In step 304, the optimal weight value of m is determined by the experimental trial and error method. The experimental trial and error method is as follows: taking 0.1 as an interval to take values one by one in the range, comparing the spectral distance between the corrected spectrum and the standard spectrum reflectivity each time, and taking the weight value when the spectral distance is the smallest as the best weight value that meets the correction model under the scene.

[0117] In step 305, based on the illumination correction component L enhancement , a corrected multi-channel spectral image is obtained.

[0118] Step four, global correction based on spectral prior constraint:

[0119] In step four, a linear regression model based on prior spectral constraints is introduced to deal with the problem of loss of gray value caused by single-channel processing. The correction coefficient and bias item are obtained by minimizing the objective function to correct the multi-channel spectral image. The standard spectral data is used as prior knowledge to provide the correlation information between spectral channels for the regression model. The multi-channel spectral information is input into the regression model as a complete whole, so that the model can represent the spectral channel features and spatial structure features at the same time, and the regional adaptive correction can maintain the overall information of the multi-channel spectrum.

[0120] In step 401, based on the original multi-channel spectral image I(x, y) obtained in step two, the standard spectral reflectance of the sampling point is introduced to construct a linear regression model.

[0121] In step 401, the linear regression model is represented as Y = XW + B.

[0122] In the formula:

[0123] Y is an output matrix of n x N, and Y represents all output values of all samples.

[0124] X is an input matrix of n x N, which is the multi-channel spectral data after the black and white board correction;

[0125] W is a regression coefficient matrix of N x N size; each column w j in W represents the regression coefficient of the jth output variable;

[0126] B is an n x N column vector, and B represents the bias item error, each row is the same bias vector β j .

[0127] n is the total number of samples;

[0128] N is the total number of channels;

[0129] j is the sequence of the channel, j = 1, 2, …, N.

[0130] In step 401, in the linear regression model, the jth output value y ij of the ith sample is obtained by linear combination of the N input features of the sample, plus the bias vector β j , that is:

[0131] y ij = x i1 w 1j +x i2 w 2j +…+x iN w Nj +β j; i is the sequence of samples, i = 1, 2, …, n; for easy calculation, the linear regression model is represented as: Y = XW + B.

[0132] In step 402, the structural features and the spatial domain spectral features are fused to construct a loss function L.

[0133] The expression of the loss function L is:

[0134]

[0135] In the formula:

[0136] α1 and α2 are regularization parameters for balancing the spatial structure item and the spectral gradient item of the loss function L;

[0137] is the spatial structure item of the loss function L;

[0138] is the spectral gradient item of the loss function L;

[0139] is the 16-dimensional standard spectral data in the i-th row;

[0140] y i is the 16-dimensional spectral data output by the linear regression model in the i-th row;

[0141] ‖·‖ F is the Frobenius norm;

[0142] SSIM(·) is the structural similarity calculation;

[0143] is the spectral gradient of the 16-dimensional standard spectral data in the i-th row;

[0144] is the 16-dimensional spectral gradient output by the linear regression model in the i-th row.

[0145] In step 402, the construction process of the loss function L is:

[0146] In step 40201, the spectral structure loss function l structure is used to measure the difference between the multispectral image output by the linear regression model and the standard uniform multispectral image.

[0147] The spectral structure loss function l structure is:

[0148]

[0149] Step 40202, add gradient information constraint reflecting the trend of change between adjacent wavelengths, increase the structure perception smoothing constraint of the multispectral image, and obtain a spectral gradient loss function l gradient For:

[0150]

[0151] Step 40203, in order to simultaneously maintain spatial structure information and suppress spectral distortion, a loss function L is defined which fuses structure features and spatial domain spectral features.

[0152] Step 403, based on the minimization of the loss function L, the linear regression model is optimized to obtain the optimal linear regression model, that is, the illumination non-uniform global correction model with spectral prior constraint.

[0153] The illumination non-uniform global correction model with spectral prior constraint is:

[0154]

[0155] In the formula:

[0156] I(x,y) is the original multi-channel spectral image;

[0157] f c (·) is the CLAHE algorithm operation;

[0158] f gid (·) is the multi-scale guided filter illumination component estimation operation;

[0159] W is an N×N size regression coefficient matrix;

[0160] B is an n×N column vector;

[0161] n is the total number of samples;

[0162] N is the total number of channels;

[0163] γ is a gamma correction parameter γ(x,y).

[0164] In step 403, adjusting the spatial structure term and the spectral gradient term of the loss function guides the regression model to maintain specific standard spectral data structure characteristics, and the regression coefficient matrix and the bias term are calculated by minimizing the loss function.

[0165] In step 403, the multi-channel global correction with spectral prior constraint adds spatial structure and spectral gradient constraint terms on the basis of non-uniform illumination correction, ensuring that the correlation and consistency of the original spectral features are maintained during the illumination change correction process, thereby maximizing the preservation of the integrity of the multi-channel spectral data during the correction of uneven illumination and ensuring that the corrected spectral data accurately reflects the true reflection characteristics of the material.

[0166] Step 404, the corrected multi-channel spectral image obtained in step three is input into a global correction model with added spectral prior constraint, global correction is performed, and corrected multi-spectral imaging data is obtained.

[0167] Application example:

[0168] The application example is based on the constraint correction method for non-uniformly illuminated multi-spectral imaging data given in the above embodiment. Specifically, in order to verify the application effect of the constraint correction method for non-uniformly illuminated multi-spectral imaging data proposed in the application, application effect tests are respectively performed on the simulated mural and the mixed pigment part of the thirteenth statue of Dule Temple.

[0169] A region-adaptive two-dimensional gamma correction method is used to balance the non-uniformly illuminated region, and then a CLAHE algorithm is used to process and improve the overall contrast and detail performance of the image. Then, a regression correction model based on spectral prior constraint is used to globally constrain the original multi-channel spectral image, so as to ensure that the correlation and consistency of the original spectral features are maintained during the illumination change correction process.

[0170] As shown in the results obtained by the method of the application, Figure 2 for non-uniform correction of the simulated mural scene, the method of the application still has good illumination correction effect, and the global correction method according to the prior spectral feature constraint can better restore the spectral features of the original channels.

[0171] For the verification effect of the mixed pigment real mural cultural relic, as shown in Figure 3 Even in the mixed pigment area, the edges of the classification results after correction are clearer, the similarity between the corrected spectral data and the standard reflectivity is obviously improved, the distinction between categories is significantly enhanced, and the overall classification visual effect is natural. Overall, the constraint correction method for non-uniformly illuminated multi-spectral imaging data has good effect.

Claims

1. A method of constrained correction of non-uniformly illuminated multispectral imaging data, characterized by, The method comprises the following steps: Step one, imaging data introduction; Step two, dark noise removal; According to the environmental dark noise, environmental dark noise removal is performed on the image to obtain an original multi-channel spectral image I(x, y); Step three, region adaptive non-uniform illumination image correction: In step 301, the original multi-channel spectral image I(x, y) obtained in step two is used to calculate an original light component L(x, y) estimated by guided filtering based on a multi-scale Retinex model based on local pixel linear relationship; Step 302, based on the guided filter estimated original illumination component L(x, y) obtained in step 301, the image is regionally segmented, and for the brightness feature of each region in the image, a dynamically adjusted gamma correction method is used to achieve balanced processing of over-bright regions and over-dark regions, to obtain an adaptive gamma corrected illumination component L g (x, y); The gamma correction parameter γ(x, y) is used in the dynamic gamma correction method; The calculation method of the gamma correction parameter γ(x, y) is: In the formula: μ i is the mean of the illumination component in the i-th region; σ i σi is the variance of the illumination component in the i-th region; Step 303, using CLAHE algorithm to enhance the adaptive gamma corrected illumination component L obtained in step 302 g (x,y) of the contrast and the detail definition, to obtain the enhanced adaptive gamma corrected illumination component L c (x,y) of the contrast and the detail definition, to obtain the enhanced adaptive gamma corrected illumination component L Step 304, the enhanced self-adaptive gamma corrected illumination component L obtained in step 303 is weighted fused with the original illumination component L estimated by the guided filter to obtain the final illumination correction component L c (x, y) and the original illumination component L estimated by the guided filter to obtain the final illumination correction component L g (x, y) and the original illumination component L estimated by the guided filter to obtain the final illumination correction component L enhancement : Step 305, based on the illumination correction component L enhancement a corrected multi-channel spectral image is obtained; Step four, global correction based on spectral prior constraint: In step 401, based on the original multi-channel spectral image I(x, y) obtained in step two, a standard spectral reflectance of a sampling point is introduced to construct a linear regression model; In step 402, a loss function L is constructed by fusing structural features and spatial domain spectral features; The expression of the loss function L is: In the formula: α1 and α2 are regularization parameters for balancing the spatial structure term and the spectral gradient term of the loss function L; is a spatial structure term for the loss function L; is the spectral gradient term of the loss function L; is the 16-dimensional standard spectral data in the i-th row; y i The i-th row of 16-dimensional spectral data output by the linear regression model; ‖·‖ F is the Frobenius norm; SSIM(·) is a structural similarity calculation; Spectral gradient for the i-th row of 16-dimensional standard spectral data; The 16-dimensional spectral gradient for the i-th row output by the linear regression model; In step 403, based on minimization of the loss function L, the linear regression model is optimized to obtain an optimal linear regression model, that is, an illumination non-uniform global correction model with added spectral prior constraint; The illumination non-uniform global correction model with added spectral prior constraint is: In the formula: I(x, y) is an original multi-channel spectral image; f c (·) is a CLAHE algorithm operation; f gid (·) is a multi-scale guided filter illumination component estimation operation; W is a regression coefficient matrix of N×N size; B is a column vector of n×N; n is the total number of samples; N is the total number of channels; γ is a gamma correction parameter γ(x, y); In step 404, the corrected multi-channel spectral image obtained in step three is input into the illumination non-uniform global correction model with added spectral prior constraint as a whole to perform global correction, and corrected multi-spectral imaging data is obtained.

2. The method of constrained correction of non-uniformly illuminated multispectral imaging data of claim 1, wherein, The specific process of step one is: collecting multi-spectral images of cultural relics, performing registration and clipping processing on the multi-spectral images to obtain a non-uniformly illuminated mural image, and introducing a dark frame image under no-light conditions as reference data.

3. The method of constrained correction of non-uniformly illuminated multispectral imaging data of claim 2, wherein, According to the dark frame image introduced in step one, the environmental dark noise is estimated by using a laboratory calibration method, and the non-uniformly illuminated mural image obtained in step one is subjected to environmental dark noise removal according to the environmental dark noise to obtain an original multi-channel spectral image I(x, y).

4. The method of constrained correction of non-uniformly illuminated multispectral imaging data of claim 1, wherein, In step two, the calculation method of the environmental dark noise removal is: In the formula: DN (i,j,λ) DN is the removed dark noise value of the i-th row, j-th column camera in the spectral channel with central wavelength λ. DH raw(i,j,λ) Rij is the response value of the i-th row, j-th column camera in the spectral channel with central wavelength λ; DN andianliu(i,j,λ,k) is the dark current value of the pixel in the i-th row and j-th column in the spectral channel with central wavelength λ at the k-th time point. N is the total number of observation time points.

5. The method of constrained correction of non-uniformly illuminated multispectral imaging data of claim 1, wherein, In step 301, the multi-scale Retinex model is: In the formula: L(x, y) is an original light component estimated by guided filtering; i is a filter scale; N is the number of filter scales, w i The weight of different scales is 1 / 3. I(x, y) is a guide image; q i (x,y) is a steering filter function.

6. The method of constrained correction of non-uniformly illuminated multispectral imaging data of claim 1, wherein, In step 302, the dynamic gamma correction method is: In the formula: L g (x, y) is the adaptive gamma-corrected illumination component; L(x, y) is an original light component estimated by guided filtering; γ(x, y) is a gamma correction parameter.

7. The method of constrained correction of non-uniformly illuminated multispectral imaging data of claim 1, wherein, In step 303, the calculation formula of the CLAHE algorithm enhancement is: L c (x,y) = f c (L g (x,y)) In the formula: L c (x,y) is the enhanced adaptive gamma corrected illumination component; L g (x, y) is the adaptive gamma-corrected illumination component; f c (·) is the CLAHE algorithm operation.

8. The method of constrained correction of non-uniformly illuminated multispectral imaging data of claim 1, wherein, In step 304, the calculation method of the weighted fusion is: L enhancement = mL(x, y) + (1 - m)L c (x, y); In the formula: m is a weight value, and ranges from 0 to 1.

9. The method of constrained correction of non-uniformly illuminated multispectral imaging data of claim 1, wherein, In step 401, the linear regression model is represented as: Y=XW+B; In the formula: Y is an output matrix of n x N, Y represents all output values of all samples; X is an input matrix of n x N, which is multi-channel spectral data after blackboard correction; W is a regression coefficient matrix of size N x N; each column w j represents the regression coefficient for the jth output variable; B is an n x N column vector, B represents the bias term error each row is the same bias vector β j ; n is the total number of samples; N is the total number of channels; j is the sequence of channels, j=1, 2, …, N.

Images