A multi-exposure dynamic range enhancement method for low-light images

Abstract

A multi-exposure dynamic range enhancement method for low-light images, for a group of low-light image sequences which are registered, continuous multi-exposure imaging of the same scene, for different exposure time original images, first, calculate the contrast, texture and exposure weight factors for each pixel, the three weighted multiplication and normalization to get the weight map corresponding to different exposure time; second, the original image of different exposure time is decomposed by Laplacian pyramid, and the weight map corresponding to different exposure time is decomposed by Gaussian pyramid; third, the Gaussian pyramid weight coefficient of different exposure time is weighted and averaged with the Laplacian pyramid layer by layer, and the new Laplacian pyramid is fused; finally, the fused Laplacian pyramid is reconstructed, and the final fused target scene high dynamic range low-light image is obtained.

Description

TECHNICAL FIELD

[0001] The present application relates to a multi-exposure dynamic range enhancement method for low-light images, and belongs to the field of image processing. BACKGROUND

[0002] Image sensors are widely used in industrial detection, medical diagnosis, military detection, aerospace, etc. Since the 1960s, image sensors have been installed on aircraft, airships and satellites, etc. The ground images obtained by them can be used for military reconnaissance, map surveying, etc. The current commonly used detectors in aerospace remote sensing, such as Time Delay Integration Charge Coupled Device (TDICCD) detectors, CMOS (Complementary Metal Oxide Semiconductor) detectors, etc., have a dynamic range of about 60dB, which is much smaller than the actual target scene dynamic range of more than 180dB. Therefore, the direct use of such detectors will lose a lot of information in the target scene. Night detection has extremely important strategic significance for military reconnaissance, but the dynamic range problem is more prominent when monitoring night scenes. The brightness and contrast of low-light images obtained at night are significantly lower, and local detail information is easily hidden, so that the aerospace remote sensing detection system cannot conduct more in-depth research on the obtained night color images. At present, researchers have tried to improve the dynamic range of the detector itself, but there are still many problems such as large dark noise and strong signal interference, and the design and flow sheet have not yet reached the mature and applicable level. In addition, the wide dynamic range device itself also has the problem of high cost. The traditional image processing field often uses a multi-exposure image fusion method, that is, a plurality of pictures with different exposure times are fused to obtain information of high-brightness areas in low-exposure images and low-dark areas in high-exposure images. SUMMARY

[0003] The technical problem solved by the present application is that the multi-exposure image fusion method used in the traditional image processing field in the prior art has various defects, and a multi-exposure dynamic range enhancement method for low-light images is proposed.

[0004] The present application solves the above technical problems by the following technical solutions:

[0005] A multi-exposure dynamic range enhancement method for low-light images, comprising:

[0006] Performing pixel-by-pixel calculation of contrast factor C, texture T, and exposure appropriateness E on a group of low-light image sequences;

[0007] The contrast factor C, the texture T and the exposure appropriateness E are weighted and multiplied and normalized to determine the weight of the pixel coordinate (i, j) of the k-th image

[0008] The gray value L of each input image in the low-light image sequence k The weights of each input image are obtained by performing Laplacian pyramid decomposition Gaussian pyramid decomposition is performed

[0009] The Gaussian pyramid decomposition result and the Laplacian pyramid decomposition result are weighted and averaged

[0010] The weighted and averaged result is subjected to inverse Laplacian pyramid transformation to obtain the inverse transformation final result as a group of target scene high dynamic range low-light images.

[0011] The low-light image sequence is registered, and the same scene is continuously imaged by multiple exposures, the number of input images of each group of low-light image sequences is N, and the calculation method of the contrast factor C is

[0012] Each input image is subjected to Laplacian filtering, and the absolute value of the filtering result is taken as the contrast factor of the pixel coordinate (i, j):

[0013] C k (i,j)=|▽ 2 L k (i,j)|

[0014] In the formula, C k (i,j) represents the contrast factor of the pixel coordinate (i, j) of the k-th image (k = 1, 2,..., N), ▽ 2 is the Laplacian operator, and L k (i,j) is the gray value of the pixel coordinate (i, j) of the k-th image (k = 1, 2,..., N).

[0015] The calculation method of the texture T is that for any pixel (i, j) of an input image, a sliding window with a size of 10x10 is taken, and the normalized gray co-occurrence matrix of the window image with the pixel (i, j) as the center is calculated According to the angular second moment of the normalized gray co-occurrence matrix, the texture T of the pixel (i, j) is obtained.

[0016] The gray co-occurrence matrix H k,i,j is specifically represented as:

[0017] H k,i,j (r,c)=P(G(m,n)=r|G(m+d m ,n+d n )=c)

[0018] In the formula, G represents the image gray value matrix corresponding to the sliding window.

[0019] Normalized gray co-occurrence matrix Specifically represented as:

[0020]

[0021] Angular second moment T k (i,j) is specifically represented as:

[0022]

[0023] In the formula, T k (i,j) represents the texture factor at the pixel coordinate (i,j) of the k-th image, and the angular second moment T k (i,j) is the texture T of the pixel (i,j).

[0024] The calculation method of the exposure adequacy E is that the reflection degree image R k of the k-th image is obtained by processing each input image through a multi-scale Retinex algorithm, and the reflection degree image R k (i,j) at which the pixel (i,j) is located is specifically represented as:

[0025]

[0026]

[0027] In the formula, N r represents the number of scales, represents the scale n r corresponding to the weighting coefficient, δ1, δ2, and δ3 are fixed values, represents the value of the filter function at the pixel (i,j);

[0028] Let The exposure adequacy factor evaluation formula of each input image at the pixel (i,j) is:

[0029]

[0030] In the formula, E k (i,j) represents the exposure adequacy factor at the pixel coordinate (i,j) of the k-th image, and σ is a fixed value.

[0031] The determination method of the weight at the pixel coordinate (i,j) of the k-th image is:

[0032]

[0033]

[0034] ω C , ω T , ω E respectively correspond to the weights of contrast factor C, texture T, and exposure adequacy E.

[0035] The Gaussian pyramid decomposition method is as follows:

[0036] First, the input image to be decomposed is taken as the 0th layer of the Gaussian pyramid GP0, convolution is performed on the image after convolution by using a low-pass filter ψ, and down-sampling is performed on the image to remove even rows and columns, so as to obtain the image of the previous layer, and the image of the previous layer is taken as the input image to repeat the convolution and down-sampling operation until the top layer of the pyramid is reached, wherein:

[0037]

[0038] 1≤l≤P_Height, 0 G < Row l , 0 G < Col l

[0039] In the formula, GP l (r G , c G ) represents the value at the pixel coordinate (r G , c G ) of the lth layer of the Gaussian pyramid image, P_Height is the total number of pyramid decomposition layers, Row l , Col l are respectively the number of columns and the number of rows of the lth layer of the Gaussian pyramid image;

[0040] The determination method of the total number of pyramid decomposition layers P_Height is as follows:

[0041] P_Height=log2[min(H,W)]

[0042] In the formula, H and W are respectively the height and the width of the image to be decomposed.

[0043] The low-pass filter ψ is represented as:

[0044]

[0045] The Laplacian pyramid decomposition method is as follows:

[0046] First, the input image to be decomposed is subjected to Gaussian pyramid decomposition to obtain the image sequence GP0, GP1, …, GP P_Height , and then the decomposition result is subjected to extended prediction to obtain the image sequence GP0 * , GP1 *Gp, Gp,..., Gp * P_Height The decomposed images of the Laplacian pyramid are determined by the difference between the corresponding image sequences of the two groups;

[0047] The method of extended prediction for the decomposed result is:

[0048] The (l-1)th layer Gaussian decomposed image GP l-1 is predicted before down-sampling. l After extended interpolation, a low-pass filter ψ is used for convolution to obtain the predicted image GP l * wherein:

[0049]

[0050] 0≤l≤P_Height,0<r G <Row l ,0<c G <Col l

[0051]

[0052] In the formula, GP l * (r L ,c L ) represents the value at the pixel coordinate (r L ,c L ) of the lth layer Laplacian pyramid image, and the difference between GP l and GP l * is the lth layer Laplacian pyramid image LP l .

[0053] The weighted average formula is:

[0054]

[0055] In the formula, GP k,l represents the lth layer Gaussian pyramid decomposed result of the kth weight image, LP k,l represents the lth layer Laplacian pyramid decomposed result of the kth original image, and LP l out represents the weighted average result of the lth layer.

[0056] The inverse transformation of the Laplacian pyramid starts from the top layer LP out P_Height of the pyramid, performs up-sampling, and then performs convolution with the next layer image LP out P_Height-1Add, continue up-sampling and superposition processing to the last layer to obtain the final result of inverse transform, and the result is a set of target scene high dynamic range low-light images.

[0057] Compared with the prior art, the present application has the following advantages:

[0058] (1) The multi-exposure dynamic range enhancement method for low-light images provided by the present application calculates contrast, texture and exposure adequacy three weight factors pixel by pixel for a multi-exposure image sequence, and for the exposure adequacy factor, first, the image sequence is processed by a multi-scale Retinex algorithm to restore the true color and details of the image, then the mean value of the retinex processing result of the image sequence is taken as the standard of ideal exposure, and finally the exposure degree is determined according to the difference between the pixel value and the ideal exposure standard, avoiding the error introduced by manual selection of a fixed value.

[0059] (2) The present application weights and multiplies the contrast, texture and exposure adequacy three weight factors, normalizes to obtain a weight map corresponding to different exposure times, and respectively performs weighted averaging through Gaussian pyramid decomposition and Laplacian pyramid to obtain a target scene high dynamic range low-light image after inverse transform of the Laplacian pyramid, thereby improving the dynamic range of the low-light scene and increasing the image information, providing support for subsequent image interpretation. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 A multi-exposure dynamic range enhancement method flowchart is provided for the present application. DETAILED DESCRIPTION

[0061] A multi-exposure dynamic range enhancement method for low-light images, for a set of low-light image sequences that are registered, imaged continuously and multi-exposure for the same scene, for original images of different exposure times, first, contrast, texture and exposure adequacy three weight factors are calculated pixel by pixel, and the three are weighted and multiplied and normalized to obtain a weight map corresponding to different exposure times; second, the original images of different exposure times are decomposed by a Laplacian pyramid, and the weight maps corresponding to different exposure times are decomposed by a Gaussian pyramid; third, the Gaussian pyramid weight coefficients of different exposure times are weighted and averaged with the Laplacian pyramid layer by layer to obtain a new Laplacian pyramid; finally, the fused Laplacian pyramid is reconstructed to obtain a fused target scene high dynamic range low-light image.

[0062] The specific steps are as follows:

[0063] The contrast factor C, the texture T and the exposure adequacy E are calculated pixel by pixel for a set of low-light image sequences.

[0064] The contrast factor C, the texture T and the exposure appropriateness E are weighted and multiplied and normalized to determine the weight of the pixel coordinate (i, j) of the k-th image

[0065] The gray value L of each input image in the low-light image sequence k The Laplacian pyramid decomposition is performed on the weight of each input image The Gaussian pyramid decomposition is performed

[0066] The weighted average of the Gaussian pyramid decomposition result and the Laplacian pyramid decomposition result is performed

[0067] The inverse Laplacian pyramid transformation is performed on the weighted average result to obtain the inverse transformation final result as a group of target scene high dynamic range low-light images.

[0068] Wherein:

[0069] The low-light image sequence is registered, and the same scene is continuously exposed and imaged, the number of input images of each group of low-light image sequences is N, and the calculation method of the contrast factor C is:

[0070] The Laplace filter is performed on each input image, and the absolute value of the filter result is taken as the contrast factor of the pixel coordinate (i, j):

[0071] C k (i,j)=|▽ 2 L k (i,j)|

[0072] In the formula, C k (i,j) represents the contrast factor of the pixel coordinate (i, j) of the k-th image (k = 1, 2,..., N), and △ 2 is the Laplace operator, and L k (i,j) represents the gray value of the pixel coordinate (i, j) of the k-th image (k = 1, 2,..., N).

[0073] The calculation method of the texture T is that for any pixel (i, j) of an input image, the sliding window size is taken as 10*10, the normalized gray co-occurrence matrix H According to the angular second moment of the normalized gray co-occurrence matrix as the texture T of the pixel (i, j).

[0074] The gray co-occurrence matrix H k,i,j is specifically represented as:

[0075] H k,i,j (r,c)=P(G(m,n)=r|G(m+d m ,n+dn )=c)

[0076] In the formula, G represents the image gray value matrix corresponding to the sliding window.

[0077] Normalized gray co-occurrence matrix Specifically represented as:

[0078]

[0079] Angular second moment T k (i,j) is specifically represented as:

[0080]

[0081] In the formula, T k (i,j) represents the texture factor at the pixel coordinate (i,j) of the k-th image, and the angular second moment T k (i,j) is the texture T of the pixel (i,j).

[0082] The calculation method of the exposure adequacy E is that: the reflectance image R k of the k-th image is obtained by processing each input image through a multi-scale Retinex algorithm, and the reflectance image R k (i,j) in which the pixel (i,j) is located is specifically:

[0083]

[0084]

[0085] In the formula, N r represents the number of scales, represents the weighting coefficient corresponding to the scale n r , and δ1, δ2, and δ3 are fixed values, represents the value of the filter function at the pixel (i,j).

[0086] Let The exposure adequacy factor evaluation formula of each input image at the pixel (i,j) is:

[0087]

[0088] In the formula, E k (i,j) represents the exposure adequacy factor at the pixel coordinate (i,j) of the k-th image, and σ is a fixed value.

[0089] The determination method of the weight at the pixel coordinate (i,j) of the k-th image is:

[0090]

[0091]

[0092] ω C ,ω T ,ω E respectively correspond to the weights of contrast factor C, texture T, exposure appropriateness E;

[0093] The Gaussian pyramid decomposition method is as follows:

[0094] The input image to be decomposed is taken as the 0th layer of the Gaussian pyramid GP0, convolution is performed on the image after convolution by using a low-pass filter ψ, and the image is down-sampled to remove even rows and columns to obtain the image of the upper layer, and the image of the upper layer is taken as the input image to repeat the convolution and down-sampling operation until the top layer of the pyramid is reached, wherein:

[0095]

[0096] 1≤l≤P_Height,0<r G <Row l ,0<c G <Col l

[0097] In the formula, GP l (r G ,c G ) represents the value at the pixel coordinate (r G ,c G ) of the lth layer of the Gaussian pyramid image, P_Height represents the total number of pyramid decomposition layers, Row l ,Col l respectively represent the number of columns and rows of the lth layer of the Gaussian pyramid image;

[0098] The method for determining the total number of pyramid decomposition layers P_Height is as follows:

[0099] P_Height=log2[min(H,W)]

[0100] In the formula, H and W respectively represent the height and width of the image to be decomposed;

[0101] The low-pass filter ψ is represented as:

[0102]

[0103] The Laplacian pyramid decomposition method is as follows:

[0104] The input image to be decomposed is first subjected to Gaussian pyramid decomposition to obtain the image sequence GP0, GP1, …, GP P_Height , and then the decomposition result is subjected to extension prediction to obtain the image sequence GP0* , GP1 * , …, GP * P_Height , the difference of the two groups of image sequences is determined as the decomposition image of the Laplacian pyramid;

[0105] The method for performing the extended prediction on the decomposition result is:

[0106] The prediction is performed on the (l-1)th layer Gaussian decomposition image GP l-1 , after the extended interpolation on the lth layer image GP l , and then the low-pass filter ψ is used for convolution to obtain the predicted image GP l * , wherein:

[0107]

[0108] 0≤l≤P_Height, 0 G < Row l , 0 G < Col l

[0109]

[0110] In the formula, GP l * (r L , c L ) represents the value at the pixel coordinate (r L , c L ) of the lth layer Laplacian pyramid image, and the difference between GP l and GP l * represents the lth layer Laplacian pyramid image LP l ;

[0111] The weighted average formula is:

[0112]

[0113] In the formula, GP k,l represents the lth layer Gaussian pyramid decomposition result of the kth weight image, LP k,l represents the lth layer Laplacian pyramid decomposition result of the kth original image, and LP l out represents the weighted average result of the lth layer;

[0114] The inverse transformation of the Laplacian pyramid starts from the top layer LP out P_Height , performs the up-sampling, and then performs the convolution with the next layer image LP outP_Height-1 Add, continue up-sampling and superposition processing to the last layer to obtain the inverse transform final result as a set of target scene high dynamic range low-light images.

[0115] Further illustration is made below according to specific embodiments:

[0116] In the current embodiment, as shown in Figure 1 Step 1: For a set of low-light image sequences (the number of input images is N) that are registered, continuous multi-exposure imaging of the same scene, calculate three weight factors of contrast C, texture T and exposure appropriateness E pixel by pixel. The calculation method of the contrast factor C is: perform Laplace filtering on each image, and take the absolute value of the filtering result as the contrast factor, that is:

[0117] C k (i,j)=|▽ 2 L k (i,j)|

[0118] In the formula, C k (i,j) represents the contrast factor at the pixel coordinate (i,j) of the k-th image (k=1, 2,..., N), ▽ 2 is the Laplace operator, and L k (i,j) represents the gray value at the pixel coordinate (i,j) of the k-th image (k=1, 2,..., N).

[0119] The calculation method of the texture T is: for a certain pixel (i,j) of an image, take the sliding window size as 10x10, calculate the normalized gray co-occurrence matrix of the window image with the pixel (i,j) as the center The element H k,i,j (r,c) in the gray co-occurrence matrix represents the occurrence probability of the event that “when the gray value of a certain pixel point (m+d m , n+d n ) of the window image is r, the gray value of the pixel point (m,n) of the window image is c”, that is:

[0120] H k,i,j (r,c)=P(G(m,n)=r|G(m+d m ,n+d n )=c)

[0121] In the formula, G represents the image gray value matrix corresponding to the sliding window. The normalized gray co-occurrence matrix is specifically represented as:

[0122]

[0123] The angular second moment of the normalized gray level co-occurrence matrix is taken as the texture feature of pixel (i, j). The angular second moment is the square sum of each element of the gray level co-occurrence matrix, i.e.

[0124]

[0125] In the formula, T k (i, j) represents the texture factor at pixel coordinate (i, j) of the kth image, represents the value of the rth row and the cth column of the normalized gray level co-occurrence matrix.

[0126] The calculation method of the exposure fitness E is as follows: the reflectance image R k i.e.

[0127]

[0128]

[0129] In the formula, N r represents the number of scales, generally three scales of large, medium and small are taken, represents the scale n r corresponding weighting coefficient, generally δ1, δ2 and δ3 are respectively taken as 15, 80 and 250, represents the value of the filter function at pixel (i, j).

[0130] Let then the exposure fitness factor of each image is evaluated by the following formula:

[0131]

[0132] In the formula, E k (i, j) represents the exposure fitness factor at pixel coordinate (i, j) of the kth image, and σ is taken as 0.2.

[0133] The three weight factors are weighted and normalized to obtain the weight i.e.

[0134]

[0135]

[0136] In the formula, ω C , ω T and ω E are respectively used to control the influence degree of the three weight factors of contrast C, texture T and exposure fitness E, and generally all take 1, which can be adjusted according to actual conditions.

[0137] Step two: Laplacian pyramid decomposition is performed on each original image L k (i,j) and each weight image W is decomposed by Gaussian pyramid.

[0138] The process of Gaussian pyramid decomposition is that, taking the weight image to be decomposed as the bottom layer (i.e. the 0th layer of Gaussian pyramid, GP0), convolving it with a low-pass filter ψ, then downsampling (removing even rows and columns) the convolved image to obtain the image of the upper layer, and repeating the convolution and downsampling operation with the image as the input until the top layer of the pyramid is reached. The formula is expressed as

[0139]

[0140] 1≤l≤P_Height,0<r G <Row l ,0<c G <Col l

[0141] In the formula, GP l (r G ,c G ) represents the value at the pixel coordinate (r G ,c G ) of the lth layer of the Gaussian pyramid image, P_Height represents the total number of pyramid decomposition layers, Row l , Col l represent the number of columns and rows of the lth layer of the Gaussian pyramid image, respectively.

[0142] The total number of pyramid decomposition layers P_Height is determined by the following formula, and H and W represent the height and width of the image to be decomposed, respectively:

[0143] P_Height=log2[min(H,W)]

[0144] The low-pass filter ψ can be expressed as:

[0145]

[0146] The process of Laplacian pyramid decomposition is that, first, the image is decomposed by Gaussian pyramid to obtain the image sequence GP0, GP1, …, GP P_Height , then the decomposition result is extended to obtain the image sequence GP0 * , GP1 * , …, GP * P_Height, two image sequences corresponding difference can be obtained Laplacian pyramid decomposition image. Gaussian pyramid extended prediction refers to the prediction in the first l-1 layer of Gaussian decomposition image GP l-1 , first l layer image GP l extended interpolation and then using low-pass filter ψ convolution, get prediction image GP l * , that is:

[0147]

[0148] 0≤l≤P_Height,0<r G <Row l ,0<c G <Col l

[0149]

[0150] In the formula, GP l * (r L ,c L ) represents the value of the lth layer Laplacian pyramid image pixel coordinates (r L ,c L ) at. GP l and GP l * The difference between GP l .

[0151] Step three: the weight image Gaussian pyramid decomposition results and the original image Laplacian pyramid decomposition results are weighted and averaged, that is:

[0152]

[0153] In the formula, GP k,l represents the lth layer Gaussian pyramid decomposition result of the kth weight image, LP k,l represents the lth layer Laplacian pyramid decomposition result of the kth original image, and LP l out represents the weighted average result of the lth layer.

[0154] Step four: the weighted average result obtained in step three is subjected to Laplacian pyramid inverse transform. The inverse transform process starts from the top layer of the pyramid LP out P_Height , which is up-sampled and added to the next layer image LP out P_Height-1 , and is stacked until the last layer to obtain the final inverse transform result. The result is the final target scene high dynamic range low-light image.

[0155] The application provides a multi-exposure dynamic range enhancement method for a low-light image.

[0156] Although the application has been disclosed as above with preferred embodiments, it is not intended to limit the application, and any person skilled in the art can make possible changes and modifications to the technical solutions of the application by using the disclosed methods and technical contents without departing from the spirit and scope of the application.

[0157] The contents not described in detail in the specification of the application belong to the known technology of the person skilled in the art.

Claims

1. A multi-exposure dynamic range enhancement method for low-light images, characterized in that Comprise: Performing pixel-by-pixel calculation of contrast factor C, texture T, and exposure adequacy E on a group of low-light image sequences; The contrast factor C, the texture T, and the exposure appropriateness E are weighted and multiplied and normalized to determine the weight of the kth image pixel coordinate at (i, j) The gray value L of each input image in the low-light image sequence k Perform Laplacian pyramid decomposition on the weight of each input image Perform Gaussian pyramid decomposition; Performing weighted average on the Gaussian pyramid decomposition result and the Laplacian pyramid decomposition result; Performing inverse Laplacian pyramid transformation on the weighted average result to obtain an inverse transformation final result as a group of target scene high dynamic range low-light images; The method for calculating the exposure mode E is as follows: Each input image is processed using the multi-scale Retinex algorithm to obtain the reflectance image R of the k-th image. k The reflectance image R where pixel (i,j) is located k (i,j) is specifically represented as: where N r denotes the number of scales, denotes the scale n r corresponding weighting coefficients, δ1, δ2, δ3, are fixed values, denotes the value of the filter function at pixel (i,j); Let N is the number of input images, and the exposure adequacy factor at pixel (i,j) for each input image is evaluated by the formula: where E k (i,j) denotes the exposure adequacy factor at pixel coordinate (i,j) of the kth image, and σ is a constant value.

2. The multi-exposure dynamic range enhancement method for low-light images according to claim 1, characterized in that: The low-light image sequences are registered to continuously image the same scene at multiple exposures, the number of input images of each group of low-light image sequences is N, and the calculation method of the contrast factor C is: Performing Laplacian filtering on each input image, and taking the absolute value of the filtering result as the contrast factor at pixel coordinates (i, j): where C k (i,j) represents the contrast factor at pixel coordinate (i,j) of the k-th image (k = 1, 2,..., N), is the Laplacian operator, L k (i,j) is the gray value at pixel coordinate (i,j) of the k-th image (k = 1, 2,..., N).

3. The multi-exposure dynamic range enhancement method for low-light images according to claim 2, characterized in that: The method for calculating the texture T is as follows: for a certain pixel (i, j) of any input image, a sliding window with a size of 10 x 10 is taken, and a normalized gray-level co-occurrence matrix of the window image with the pixel (i, j) as the center is calculated The angle second moment of the normalized gray-level co-occurrence matrix is taken as the texture T of the pixel (i, j).

4. The multi-exposure dynamic range enhancement method for low-light images according to claim 3, characterized in that: gray level co-occurrence matrix H k,i,j is specifically expressed as: H k,i,j (r,c) = P(G(m, n) = r | G(m + d m ,n+d n ) = c) In the formula, G represents the image gray value matrix corresponding to the sliding window; Normalized gray level co-occurrence matrix Specifically represented as: Corner second-order distance T k (i,j) is specified as: where T k (i,j) represents the texture factor at pixel coordinate (i,j) of the k-th image, the angular second moment T k (i,j) is the texture T of pixel (i,j).

5. The multi-exposure dynamic range enhancement method for low-light images according to claim 4, characterized in that: weight of the kth image at pixel coordinate (i,j) The determination method is that: In the formula, ω C , ω T , and ω E correspond to the weights of the contrast factor C, the texture T, and the exposure adequacy E, respectively.

6. The multi-exposure dynamic range enhancement method for low-light images according to claim 5, characterized in that: The Gaussian pyramid decomposition method is: First, the input image to be decomposed is taken as the 0th layer of the Gaussian pyramid GP0, convolution is performed on the image using a low-pass filter ψ, and down-sampling is performed on the convolved image to remove even rows and columns, thereby obtaining the image of the previous layer, and then the image of the previous layer is taken as the input image to repeat the convolution and down-sampling operations until the top layer of the pyramid is reached, wherein: 1≤l≤P_Height,0<r G <Row l ,0<c G <Col l In the formula, GP l (r G ,c G ) represents the value at the pixel coordinate (r G ,c G ) of the lth layer Gaussian pyramid image, P_Height is the total number of pyramid decomposition layers, Row l , Col l are the column number and row number of the lth layer Gaussian pyramid image, respectively. The determination method of the total number of pyramid decomposition layers P_Height is: P_Height = log2[min(H, W)] In the formula, H and W are the height and width of the image to be decomposed, respectively; The low-pass filter ψ is expressed as:

7. The multi-exposure dynamic range enhancement method for low-light images according to claim 6, characterized in that: The Laplacian pyramid decomposition method is: First, the input image to be decomposed is decomposed by Gaussian pyramid to obtain image sequences GP0, GP1, …, GP P_Height , and then the decomposed results are extended and predicted to obtain image sequences GP0 * , GP1 * , …, GP * P_Height The decomposed images of Laplacian pyramid are determined by calculating the difference between the two groups of image sequences. The method for performing extension prediction on the decomposition result is: Predicted l-th layer Gaussian-decomposed image GP l-1 , the l-th layer image GP l is extended and interpolated and then convolved with a low-pass filter ψ to obtain a predicted image GP l * wherein: 0 < i < P_Height, 0 < r G < Row l , 0 < c G < Col l When GP l * (r L ,c L ) represents the value at pixel coordinates (r L ,c L ) of the l-th layer Laplacian pyramid image, GP l is the difference between GP l * and GP l .

8. The multi-exposure dynamic range enhancement method for low-light images according to claim 7, characterized in that: The weighted average formula is: In the formula, GP k,l denotes the lth layer Gaussian pyramid decomposition result of the kth weight image, LP k,l denotes the lth layer Laplacian pyramid decomposition result of the kth original image, LP l out denotes the weighted average result of the lth layer.

9. The multi-exposure dynamic range enhancement method for low-light images according to claim 8, characterized in that: The inverse Laplacian pyramid transform is performed from the top layer LP out P_Height First, the up-sampling is performed and added with the next layer image LP out P_Height-1 The up-sampling and superposition are continued until the last layer to obtain the final result of the inverse transform, which is a set of target scene high dynamic range low-light images.

Images