Homomorphic Filtering Detail-Enhanced Multi-Exposure Image Fusion Method and Storable Medium

Abstract

The present invention discloses a method for detail-enhanced multi-exposure image fusion based on homomorphic filtering and a storage medium, which relates to the field of image processing. The present invention includes the following steps: acquiring an image sequence; calculating the weight information of each pixel in the image sequence according to a local contrast weight algorithm and an exposure amount weight algorithm; combining the weight information of each pixel to obtain an initial weight map; filtering and normalizing the initial weight map by using a guided filter to obtain a weight map; reconstructing a pyramid by using the Gaussian pyramid of the weight map and the Laplacian pyramid of the homomorphic filtering detail enhancement of the image sequence to obtain a fused image. The present invention can generate a high-quality fused image, which retains most of the details of the original input image while also maintaining the natural transition of light and shadow.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and more specifically, to a method for enhancing details of homomorphic filtering for multi-exposure image fusion and a storage medium. Background Art

[0002] Currently, the methods for wide dynamic range fusion mainly include: high dynamic range (HDR) imaging and multi-exposure image fusion (MEF). HDR imaging needs to reconstruct the HDR image by inversely solving the camera response function (CRF). However, the CRF estimation requires multiple exposure parameters and constraints of the camera and the algorithm is complex with low practicability. The methods of multi-exposure image fusion often have a slow operation speed, and problems such as lack of spatial domain information, unnatural light and dark transitions, local color distortion, unclear texture details, and failure to adapt the overall brightness of the input image sequence.

[0003] High Dynamic Range (HDR) Imaging Method

[0004] HDR imaging generally includes two main steps: HDR reconstruction and tone mapping. First, multiple low dynamic range (LDR) images with different exposure levels are taken in the same scene, and then the HDR image is reconstructed by inversely solving the camera response function (CRF). The HDR image generated by the HDR imaging method cannot be directly displayed on a conventional LDR device. Therefore, tone mapping is used to map the HDR image into a low dynamic range image. However, generating the HDR image requires parameters of the source images, such as exposure time, exposure value, and camera response function (CRF), etc., and constraints, for example, assuming certain specific parameter forms of the CRF, but these values are usually unknown to ordinary users. Moreover, the calculation of the CRF is very complex and the tone mapping is very time-consuming, so the practicability of this method is low.

[0005] Multi-Exposure Image Fusion (MEF) Method

[0006] The multi-exposure image fusion (MEF) method is a series of low dynamic range images with different exposure levels obtained by an ordinary image acquisition device taking pictures of the same scene with different exposures. Using these images, a high-quality image with rich details and conforming to the human eye perception characteristics can be directly generated. The fusion process does not require steps such as CRF estimation, HDR image reconstruction, and tone mapping. The algorithm is simple and efficient and easy to implement. There are already some classic multi-exposure image fusion algorithms, but there are still many problems and challenges that have not been solved.

[0007] Most existing MEF methods generate fused images by weighted averaging of input images at different exposure levels, such as multi-exposure image fusion algorithms based on multi-resolution. This method uses three quality metrics: contrast, saturation, and exposure rate to estimate the weight map. Then, the source images and the obtained weight map are decomposed into Laplacian pyramids and Gaussian pyramids respectively. Then, the images with multiple exposures are blended to form a fused image. The images generated by this algorithm have good saturation. However, the details in the bright and dark regions of the fused image of this method will be lost. In addition, moving object artifacts cannot be eliminated. Since Laplacian pyramid fusion will lose some high-frequency details, some recent detail enhancement methods have become the focus of multi-exposure image fusion research. After that, some researchers proposed a technique using median and recursive filters to calculate color dissimilarity to remove artifacts, but the color appearance of the images obtained by this method is very dull and color information is lost. These multi-exposure image fusion algorithms often have problems such as detail loss, unnatural brightness transition, local color distortion, and failure to adapt well to the overall brightness of the input image sequence.

[0008] How to solve the problems existing in the existing solutions, such as the lack of spatial domain information, unnatural brightness transition, local color distortion, blurred details, and failure to adapt well to the overall darkness or brightness of the input picture sequence, is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0009] In view of this, the present invention provides a detail enhancement multi-exposure image fusion method and a storage medium based on homomorphic filtering. First, for the input LDR image sequence, the weight information of each pixel in each picture is calculated by using the locally contrast weight algorithm, exposure amount weight algorithm, and color dissimilarity weight algorithm innovatively proposed in this article. Then, two methods are respectively used to construct a comprehensive weight map for the static image sequence and the dynamic image sequence, and the fast guided filter is used to denoise the weight map. Finally, the Gaussian pyramid of the denoised weight map and the Laplacian pyramid of the input image sequence innovatively improved by the present invention are reconstructed to fuse and obtain the final fused image.

[0010] To achieve the above object, the present invention adopts the following technical solutions:

[0011] A detail enhancement multi-exposure image fusion method based on homomorphic filtering, comprising the following steps:

[0012] Obtain an image sequence;

[0013] Calculate the weight information of each pixel in the image sequence according to the locally contrast weight algorithm and the exposure amount weight algorithm;

[0014] Combine the weight information of each pixel to obtain an initial weight map;

[0015] Filter the initial weight map using a guided filter and normalize it to obtain the weight map;

[0016] Reconstruct the pyramid using the Gaussian pyramid of the weight map and the Laplacian pyramid of the homomorphic filtered details of the image sequence to obtain the fused image.

[0017] Optionally, when the image sequence is a dynamic image sequence, calculate the weight information of each pixel in the image sequence using the local contrast weight algorithm, exposure weight algorithm, and color dissimilarity weight algorithm; combine the weight information of each pixel to obtain the initial weight map.

[0018] Optionally, the local contrast weight algorithm is as follows:

[0019] Apply the Laplacian filter to the grayscale image of each image. The specific algorithm is as follows:

[0020]

[0021] where the value range of n is 1, 2, …, N, N is the number of input image sequences, A n (i,j) is the local contrast value at the (i,j) position of the nth image, |·| represents taking the absolute value calculation, represents the value at the (i,j) position of the grayscale image of the nth image, * represents the convolution operation, h is the Laplacian filter kernel, and the value of h is as follows:

[0022]

[0023] Use the maximum value at the same pixel position in all images as the local contrast weight W 1 , and the specific algorithm is as follows:

[0024]

[0025] where represents the local contrast weight value at the (i,j) position of the nth image.

[0026] Optionally, the exposure weight algorithm is as follows:

[0027] First, normalize the grayscale image of the image sequence to the [0,1] interval, and use a threshold to divide each image into two parts to calculate the exposure weight separately. The specific algorithm of the threshold is as follows:

[0028]

[0029] where It represents the value at the position (i, j) of the grayscale image after normalizing the nth image. mean{·} represents the mean calculation. The initial value of T1 is 0.5, and T0 is a very small number. If |T1 - T2| < T0 holds, then T2 is the optimal threshold. Otherwise, assign the value of T2 to T1, and repeat the above steps for iteration until the optimal threshold T2 is obtained;

[0030] Divide the image into two parts, G1 and G2, according to the optimal threshold. The G1 part is composed of pixels with grayscale values greater than T2, and the G2 part is composed of pixels with grayscale values less than or equal to T2. Calculate the means of the G1 and G2 parts respectively, and then calculate the total average, the adaptive exposure weight control factors α n and β n , and the calculation formula is as follows:

[0031]

[0032]

[0033] where T n is the optimal threshold of the nth image, α n and β n respectively represent the adaptive exposure weight control factors of the G 1 and G 2 parts; finally, use the Gaussian curve to distribute the exposure weight, and the formula is as follows:

[0034]

[0035]

[0036] where σ controls the amplitude of the curve.

[0037] Optionally, the color dissimilarity weight algorithm is specifically as follows:

[0038]

[0039]

[0040] where I n (i, j) is the pixel value at the position (i, j) of the nth input image, Ieq{·} represents the histogram equalization calculation, is the image sequence after histogram equalization, median{·} represents the median filtering calculation, and I med (i, j) is the reference image of the static background;

[0041] Detect moving targets by calculating the color dissimilarity between the histogram equalized image and the static background image, and the calculation formula is as follows:

[0042]

[0043] where is the initial weight map of color dissimilarity, and δ takes a value of 0.1;

[0044] Use morphological operators to refine the initial weight map of color dissimilarity to remove noise estimation:

[0045]

[0046] where s 1 and s 2 respectively represent flat disk structuring elements with radii of n 1 and n 2 respectively, is the dilation operation, is the erosion operation; is the color dissimilarity weight value at the position (i, j) of the nth image.

[0047] Optionally, use the Gaussian pyramid of the weight map and the Laplacian pyramid reconstructed by homomorphic filtering details enhancement of the image sequence to obtain the fused image. The specific steps are as follows:

[0048] Use a guided filter to refine the weight map and normalize it. The calculation method is as follows:

[0049]

[0050]

[0051] where GF r,ε (S, G) represents fast guided filtering, r represents the filtering radius, ε manages the blurring degree of filtering, S represents the input image, and G represents the guidance image;

[0052] Pyramid fusion based on homomorphic filtering details enhancement: Decompose the weight map and the input image using the Gaussian pyramid and the Laplacian pyramid respectively. Among them, the Laplacian pyramid decomposition decomposes the input image into the base layer and the detail layers. The highest layer is the base layer, and the other layers are the detail layers. Enhance each detail layer using homomorphic filtering. The specific formula is as follows:

[0053]

[0054] L{I n (i, j)} (l) = homomorphic(L{I n (i, j)} (l) ) l = 1, 2,..., L - 1;

[0055] where n = 1, 2, …, N, floor(·) represents rounding down towards negative infinity, r and c are the height and width of the input image respectively, min(·) represents the minimum value function, represents the pixel value at the position (i, j) of the nth image in the lth layer, upsample(·) is the upsampling operation, L{·} (l) represents the Laplacian pyramid image of the lth layer, homomorphic(·) represents the homomorphic filtering operation;

[0056] After obtaining the Laplacian pyramid of the enhanced image sequence, the Gaussian pyramid of the weight map and the Laplacian pyramid of the enhanced image sequence are fused and reconstructed to obtain the final fused image. The formula is as follows:

[0057]

[0058] L{F(i, j)} (L-l) = L{F(i, j)} (L-l) + upsample(L{F(i, j)} (L-l+1) ) l = 1, 2, …, L - 1

[0059] F(i, j) = L{F(i, j)} (1) ;

[0060] where G{·} (l) represents the Gaussian pyramid image of the lth layer, F(i, j) is the pixel value at the position (i, j) of the fused image, and the final output image is obtained.

[0061] A computer storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of any one of the methods for detail-enhanced multi-exposure image fusion based on homomorphic filtering are implemented.

[0062] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses a method for detail-enhanced multi-exposure image fusion based on homomorphic filtering and a storage medium. It considers the spatial domain information, has a more natural light and dark transition, does not distort the color, and adapts to the overall darker or brighter situation of the input picture sequence. It innovatively uses iterative partition calculation of the exposure amount weight, innovatively proposes an adaptive exposure amount weight function, cooperates with functions of local contrast weight and color dissimilarity weight, comprehensively calculates the fusion weight of each pixel point, and finally innovatively uses homomorphic filtering to enhance the pyramid detail layer to generate a high-quality fused image, while maintaining most of the details of the original input image and also keeping the natural transition of light and shadow. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on the provided drawings.

[0064] Figure 1 It is a schematic flowchart of the present invention;

[0065] Figure 2 It is a Gaussian curve graph of the present invention. Specific embodiments

[0066] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0067] The embodiments of the present invention disclose a detail-enhanced multi-exposure image fusion method based on homomorphic filtering, as Figure 1 shown, including the following steps:

[0068] Obtain an image sequence;

[0069] Calculate the weight information of each pixel in the image sequence according to the local contrast weight algorithm and the exposure amount weight algorithm;

[0070] Combine the weight information of each pixel to obtain an initial weight map;

[0071] Filter and normalize the initial weight map using a guided filter to obtain a weight map;

[0072] Reconstruct the pyramid using the Gaussian pyramid of the weight map and the Laplacian pyramid of the homomorphic filtering detail enhancement of the image sequence to obtain a fused image.

[0073] Among them, the local contrast weight is specifically:

[0074] The local contrast measurement of the input image sequence can be used to retain important details such as edges and textures. These edge and texture information are contained in the gradient changes. Therefore, a Laplacian filter with good edge detection is used to calculate the local contrast weight. Apply the Laplacian filter to the grayscale map of each image. The specific algorithm is as follows:

[0075]

[0076] where the value range of n is 1, 2, …, N, N is the number of input image sequences, A n (i, j) is the local contrast value at the position (i, j) of the n-th picture, |·| represents taking the absolute value calculation, represents the value at the position (i, j) of the grayscale image of the n-th image, * represents the convolution operation, h is the Laplacian filter kernel, and the value of h is as follows:

[0077]

[0078] The maximum value at the same pixel position in all images is used as the local contrast weight W 1 , and the specific algorithm is as follows:

[0079]

[0080] where represents the local contrast weight value at the position (i, j) of the n-th picture.

[0081] Among them, the exposure weight is specifically:

[0082] The purpose of the exposure weight is to select the area with better exposure and avoid the areas of under-exposure (value 0) and over-exposure (value 1) in the fused image. The maximum value of the Gaussian curve is 1, and the values at the too large and too small places are infinitely close to 0. Therefore, the Gaussian curve is used to assign higher weights to the areas with good exposure and lower weights to the areas with bad exposure. The algorithm for assigning weights using the Gaussian curve is as follows:

[0083]

[0084] Generally, the value of μ is 0.5 and the value of σ is 0.2, as Figure 2 shown.

[0085] Since the value of μ being 0.5 is applicable to images with medium overall brightness, for images with too high or too low overall brightness values, this way of assigning weights is not the best. In order to better adapt to the situation where the input image sequence is overall dark or overall bright, an adaptive exposure weight algorithm is innovatively proposed, and the specific method of this algorithm is as follows:

[0086] First, normalize the grayscale images of the image sequence to the interval [0, 1]. Since in a grayscale image with good exposure, the grayscale values of the darker parts are lower and the grayscale values of the lighter parts are higher, but it also contains color and detail information. To avoid misinterpreting the darker and lighter parts as having bad exposure, an iterative threshold algorithm is innovatively proposed. The threshold is used to divide each image into two parts to calculate the exposure weights separately, and the specific algorithm of the threshold is as follows:

[0087]

[0088] where represents the value at the (i, j) position of the grayscale image after normalization of the nth image, mean{·} represents the mean calculation, the initial value of T1 is 0.5, T0 is a very small number. If |T1 - T2| < T0 holds, then T2 is the optimal threshold; otherwise, assign the value of T2 to T1, and repeat the above steps for iteration until the optimal threshold T2 is obtained.

[0089] Then, divide the image into two parts, G1 and G2, according to the optimal threshold. The G1 part consists of pixels with grayscale values greater than T2, and the G2 part consists of pixels with grayscale values less than or equal to T2. Calculate the means of the two parts of each image respectively, and then calculate the total average. If the mean of the G1 part (or G2 part) of an image is less than or equal to the total average of their corresponding parts, this part of the image has a lower brightness in the entire image sequence. Then, the part with higher brightness in the image is the part with good exposure, and the weight of the part with higher brightness needs to be increased appropriately, and the weight of the part with lower brightness needs to be decreased appropriately. Similarly, if the mean of the G1 part (or G2 part) of an image is greater than the total average of their corresponding parts, the weight of the part with lower brightness needs to be increased appropriately, and the weight of the part with higher brightness needs to be decreased appropriately. Therefore, an innovative adaptive exposure weight control factor α n and β n are proposed. This factor can reflect the offset between the exposure of the input image and 0.5, and the calculation formula is as follows:

[0090]

[0091]

[0092] where T n is the optimal threshold of the nth image, α n and β n respectively represent the adaptive exposure weight control factors of the G 1 and G 2 parts. Finally, use the Gaussian curve to distribute the exposure weight, and the formula is as follows:

[0093]

[0094]

[0095] where σ controls the amplitude of the curve.

[0096] Among them, the color dissimilarity weight is specifically:

[0097] Multiple-exposure images are taken at different times, during which not all objects may be stationary. If the weight map only considers local contrast and exposure, this may lead to ghost artifacts in the fused image. Therefore, for dynamic images, the influence of moving objects must also be considered when estimating the weight map. Color dissimilarity is used to measure the color difference between the source image pixels and the static background pixels. The specific algorithm is as follows:

[0098] First, a static background is calculated as a reference image. Histogram equalization is performed on each input image to convert the color distribution of the input image into a similar color distribution, and then the reference image is selected using median filtering. The specific algorithm formula is as follows:

[0099]

[0100]

[0101] where I n (i,j) is the pixel value at position (i,j) of the nth input image, Ieq{·} represents the histogram equalization calculation, is the image sequence after histogram equalization, median{·} represents the median filtering calculation, and I med (i,j) is the reference image of the static background. Then, the moving objects are detected by calculating the color dissimilarity between the histogram equalized image and the static background image. The calculation formula is as follows:

[0102]

[0103] where is the initial weight map of color dissimilarity, and δ takes a value of 0.1. Finally, morphological operators are used to refine the initial weight map of color dissimilarity to remove noise estimation:

[0104]

[0105] where s 1 and s 2 represent flat disk structuring elements with radii of n 1 and n 2 respectively, is the dilation operation, is the erosion operation. is the color dissimilarity weight value at position (i,j) of the nth image. Its advantage is that it does not require the user to specify a reference image.

[0106] Among them, the pyramid fusion based on homomorphic filtering detail enhancement is specifically as follows:

[0107] An initial weight map is constructed from three calculated metrics, and the obtained initial weight map is noisy and discontinuous. Therefore, it is crucial to refine the initial weight map before using these weight maps for fusion processing. Edge-preserving filters can be used to refine the weight mapping, taking the input image as the guidance image or the joint image to ensure that pixels from similar objects have comparable weights. Among them, the fast guidance filter is more effective, and the fused image it generates has better color. Therefore, the guidance filter is selected to refine the weight map and normalize it.

[0108]

[0109]

[0110] Among them, GF r,ε (S, G) represents the fast guidance filter, r represents the filtering radius, ε manages the blurring degree of the filter, S represents the input image, and G represents the guidance image. The output image obtained by traditional weighted fusion contains seams, which are caused by different weights and transitional pixels resulting in seams and blurring in the final result. The pyramid-based multi-resolution method can solve this problem, but pyramid fusion will blur some texture details and colors. To retain the detail and color information of the image, detail enhancement is required. Many detail enhancement algorithms lose some low-frequency signals when enhancing high-frequency details, while homomorphic filtering can retain low-frequency signals while enhancing details. Therefore, a pyramid fusion based on homomorphic filtering detail enhancement is innovatively proposed: the weight map and the input image are decomposed using Gaussian pyramids and Laplacian pyramids respectively. Among them, the Laplacian pyramid decomposition decomposes the input image into a base layer and detail layers. The highest layer is the base layer, and the other layers are detail layers. Each detail layer is enhanced using homomorphic filtering. The specific formula is as follows:

[0111]

[0112] L{I n (i, j)} (l) =homomorphic(L{I n (i, j)} (l) ) l = 1, 2, …, L - 1;

[0113] Among them, n = 1, 2, …, N, floor(·) represents rounding down towards negative infinity, r and c are the height and width of the input image respectively, min(·) represents the minimum value function, represents the pixel value of the nth image at the position (i, j) in the lth layer, upsample(·) is the upsampling operation, L{·} (l)Denote the Laplacian pyramid image of the l-th layer, and homomorphic(·) represents the homomorphic filtering operation. After obtaining the Laplacian pyramid of the enhanced image sequence, the Gaussian pyramid of the weight map and the Laplacian pyramid of the enhanced image sequence are fused and reconstructed to obtain the final fused image. The formula is as follows:

[0114]

[0115] L{F(i,j)} (L-l) =L{F(i,j)} (L-l) +upsample(L{F(i,j)} (L-l+1) )l = 1,2,…,L - 1

[0116] F(i,j)=L{F(i,j)} (1) ;

[0117] where G{·} (l) denotes the Gaussian pyramid image of the l-th layer, and F(i, j) is the pixel value at the position (i, j) of the fused image, that is, the final output image.

[0118] A computer storage medium is also disclosed. A computer program is stored on the computer storage medium. When the computer program is executed by a processor, the steps of a method for detail-enhanced multi-exposure image fusion based on homomorphic filtering according to any one of the above are implemented.

[0119] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A detail-enhanced multi-exposure image fusion method based on homomorphic filtering, characterized in that, it includes the following steps: Obtain an image sequence; Calculate the weight information of each pixel in the image sequence according to the local contrast weight algorithm and the exposure weight algorithm; Combine the weight information of each pixel to obtain an initial weight map; Use a guided filter to filter and normalize the initial weight map to obtain a weight map; Use the Gaussian pyramid of the weight map and the Laplacian pyramid of the homomorphic filtering detail enhancement of the image sequence to reconstruct the pyramid to obtain a fused image; The local contrast weight algorithm is specifically as follows: Apply a Laplacian filter to the grayscale image of each image, and the specific algorithm is as follows: where the value range of n is 1, 2, …, N, N is the number of input image sequences, A n (i, j) is the local contrast value at the position (i, j) of the n-th picture, |·| represents taking the absolute value calculation, represents the value at the position (i, j) of the grayscale image of the n-th picture, * represents the convolution operation, h is the Laplacian filter kernel, and the value of h is as follows: Use the maximum value at the same pixel position in all images as the local contrast weight W 1 , and the specific algorithm is as follows: Among them represents the local contrast weight value at the position (i, j) of the nth figure; The exposure weight algorithm is as follows: First, normalize the grayscale images of the image sequence to the interval [0, 1], and use a threshold to divide each image into two parts to calculate the exposure weights separately. The specific algorithm of the threshold is as follows: where represents the value at the (i, j) position of the grayscale image after normalization of the nth image, mean{·} represents the mean calculation, the initial value of T1 is 0.5, T0 is a user-defined value. If |T1 - T2| < T0 holds, then T2 is the optimal threshold; otherwise, the value of T2 is assigned to T1, and the above steps are repeated for iteration until the optimal threshold T2 is obtained; The image is divided into two parts, G1 and G2, according to the optimal threshold. The G1 part consists of pixels with gray values greater than T2, and the G2 part consists of pixels with gray values less than or equal to T2. Calculate the means of the G1 and G2 parts respectively, and then calculate the total average. The adaptive exposure weight control factors α n and β n , and the calculation formulas are as follows: Among them, T n is the optimal threshold of the nth image, and α n and β n respectively represent the adaptive exposure amount weight control factors of the G 1 and G 2 parts; finally, the Gaussian curve is used to allocate the exposure amount weights, and the formula is as follows: Among them, The amplitude of the σ control curve.

2. A detail-enhanced multi-exposure image fusion method based on homomorphic filtering according to claim 1, characterized in that, it further includes that when the image sequence is a dynamic image sequence, calculate the weight information of each pixel in the image sequence by using the local contrast weight algorithm, the exposure weight algorithm, and the color dissimilarity weight algorithm; combine the weight information of each pixel to obtain an initial weight map.

3. A detail-enhanced multi-exposure image fusion method based on homomorphic filtering according to claim 2, characterized in that, the color dissimilarity weight algorithm is specifically as follows: where I n (i, j) is the pixel value at position (i, j) of the nth input image, and Ieq{·} represents histogram equalization calculation, is the image sequence after histogram equalization, median{·} represents median filtering calculation, and I med (i, j) is the reference image of the static background; Detect moving targets by calculating the color dissimilarity between the histogram equalized image and the static background image, and the calculation formula is as follows: Among them is the initial weight map of color dissimilarity, and the value of δ is 0.1; Use morphological operators to refine the color dissimilarity initial weight map to remove noise estimation: where s 1 and s 2 represent flat disk - shaped structuring elements with radii n 1 and n 2 respectively, is the dilation operation, is the erosion operation; is the color dissimilarity weight value at the position (i, j) of the n - th image.

4. A detail-enhanced multi-exposure image fusion method based on homomorphic filtering according to claim 2, characterized in that, Use the Gaussian pyramid of the weight map and the Laplacian pyramid of the homomorphic filtering detail enhancement of the image sequence to reconstruct the pyramid to obtain a fused image, and the specific steps are as follows: Use a guided filter to refine and normalize the weight map, and the calculation method is as follows: Among them, GF r,ε (S, G) represents fast guided filtering, r represents the filtering radius, ε manages the blurring degree of filtering, S represents the input image, and G represents the guidance image; Pyramid fusion based on homomorphic filtering detail enhancement: Decompose the weight map and the input image with the Gaussian pyramid and the Laplacian pyramid respectively, where the Laplacian pyramid decomposition decomposes the input image into a base layer and detail layers, the highest layer is the base layer, and the other layers are detail layers. Enhance each detail layer with homomorphic filtering, and the specific formula is as follows: L = floor(log 2 min(r, c)) - 2; L{I n (i,j)} (l) =homomorphic(L{I n (i,j)} (l) ) l=1,2,…,L - 1; where \(n = 1, 2, \ldots, N\), \(floor(\cdot)\) represents rounding down to negative infinity, \(r\) and \(c\) are the height and width of the input image respectively, \(min(\cdot)\) represents the minimum value function, represents the pixel value at the position \((i, j)\) of the \(n\)th image at the \(l\)th layer, \(upsample(\cdot)\) is the upsampling operation, \(L\{\cdot\}\) (l) represents the Laplacian pyramid image of the \(l\)th layer, \(homomorphic(\cdot)\) represents the homomorphic filtering operation; After obtaining the Laplacian pyramid of the enhanced image sequence, fuse and reconstruct the Gaussian pyramid of the weight map and the Laplacian pyramid of the enhanced image sequence to obtain the final fused image, and the formula is as follows: L{F(i,j)} (L-l) = L{F(i,j)} (L-l) + upsample(L{F(i,j)} (L-l+1) ) for l = 1, 2, …, L - 1 F(i,j) = L{F(i,j)} (1) ; where G{·} (l) represents the Gaussian pyramid image of the l-th layer, and F(i, j) is the pixel value at the position (i, j) of the fused image, to obtain the final output image.

5. A computer storage medium, characterized in that, a computer program is stored on the computer storage medium, and when the computer program is executed by a processor, it implements the steps of a detail-enhanced multi-exposure image fusion method based on homomorphic filtering according to any one of claims 1-4.

Images