Weak-light image enhancement method and system based on structural texture preservation of high and low frequency information

Abstract

The present invention belongs to the field of image enhancement technology, and specifically discloses a method and system for enhancing weak-light images based on structural texture preservation of high- and low-frequency information. The system includes an image acquisition module and a processing module. The image acquisition module is used to acquire weak-light images, input the weak-light images to be enhanced into the processing module, divide the images into the product of reflection and illumination, optimize the illumination, obtain the optimal solution for illumination, decompose the weak-light images into high-frequency layers and low-frequency layers, extract structural priors and texture priors, propose and solve the optimization problem of reflection, and combine the optimal solution for illumination with the solution for reflection optimization to obtain an enhanced image. By adopting this technical solution, for images acquired by transmission line inspection drones under insufficient lighting, the illumination and reflection are solved separately in sequence to avoid noise transfer, and then the optimized illumination and reflection are combined to obtain an enhanced image, thereby improving the visual quality of weak-light images.

Description

Technical Field

[0001] The present invention belongs to the technical field of image enhancement, and relates to a method and system for low-light image enhancement based on structural texture preservation of high- and low-frequency information. Background Art

[0002] When a transmission line inspection drone is used at night or in dim lighting conditions, the images and videos captured by the acquisition equipment often have low visibility and inevitable noise, and some useful details are directly lost. Such images are called low-light images. Weak-light images are characterized by low brightness, dark contrast, severe color distortion, and a large amount of detail information lost.

[0003] In recent years, image enhancement technology for foggy, rainy and weak light conditions has received a lot of attention and made breakthrough progress, which will have broad application prospects in target detection, tracking and image processing. In a variety of harsh climate environments, enhancing and preprocessing low-quality images caused by various reasons will be an important part of computer vision algorithms, used to reduce the impact of environmental factors such as fog, rain, underwater and dimness on the performance of computer vision algorithms mainly used for high-quality image input.

[0004] Many researchers have focused on the research of low-light image enhancement algorithms and have proposed a variety of effective methods. The simplest and most convenient one is the histogram-based image enhancement method, which stretches the dynamic range of its grayscale to achieve the effect of improving the overall brightness and details of the image. However, this method is prone to over-enhancement and is not easy to use. With the application of the Retinex theory (computational theory of color constancy perception) based on human visual perception in the field of image enhancement, the image is regarded as the product of reflection and illumination, from the single-scale Retinex algorithm to the multi-scale Retinex algorithm, and then to the multi-scale Retinex algorithm with color restoration, and then various image enhancement effects based on the Retinex algorithm are getting better and better. However, these Retinex-based methods lack the protection of the naturalness of illumination, ignore the strong noise in low-light images, and do not suppress it. Summary of the invention

[0005] The purpose of the present invention is to provide a method and system for low-light image enhancement based on structural texture preservation of high- and low-frequency information, so as to suppress noise and improve the visual quality of low-light images.

[0006] In order to achieve the above object, the basic scheme of the present invention is: a method for enhancing a low-light image by preserving structural texture based on high- and low-frequency information, comprising the following steps:

[0007] Input the low-light image to be enhanced and divide the image into the product of reflection and illumination;

[0008] The maximum value of the three channels R, G, and B is used as the initial value of the lighting. While maintaining the smoothness of the lighting segmentation, the lighting is optimized and the optimized lighting is made close to its initial value to obtain the optimal lighting solution.

[0009] Decompose the low-light image into high-frequency layer and low-frequency layer;

[0010] Filter the high-frequency layer to suppress the noise in the flat area, and extract the structure prior and texture prior from the high-frequency layer and the low-frequency layer;

[0011] Based on high and low frequency information, structure prior and texture prior, the optimization problem of reflection is proposed and solved;

[0012] The optimal lighting solution is corrected, and the corrected optimal lighting solution is combined with the solution obtained by reflection optimization to obtain an enhanced image.

[0013] The working principle and beneficial effect of this basic solution are: solving illumination and reflection separately in sequence can effectively avoid the problem of noise transfer. Adaptively separate the high-frequency layer and the low-frequency layer from the original image, and propose to use filtering on the high-frequency layer, combined with the characteristics of the high-frequency layer composed of details and noise, to suppress most of the flat remaining noise. Incorporate the high-frequency layer and the low-frequency layer into the optimization problem, and introduce structure and texture information. In the process of solving reflection, the structural prior and texture prior drive the result to reach a balance between structure and texture. By combining the corrected and adjusted illumination and reflection, an enhanced image is obtained to improve the visual quality of low-light images.

[0014] Further, the method to separate the image into the product of reflection and illumination is:

[0015] Input the low-light image to be enhanced and record it as O, O is composed of reflection R and illumination L, recorded as

[0016]

[0017] Among them, (x, y) is the coordinate of the pixel point, Represents element-wise product.

[0018] Decomposing the input image into two components, illumination and reflection, is beneficial for subsequent operations. Solving illumination and reflection separately in sequence can effectively avoid the problem of noise transfer.

[0019] Further, the method for optimizing lighting is as follows:

[0020] The initial value of lighting is

[0021]

[0022] Among them, (x, y) is the coordinate of the pixel point, and c refers to the selected R, G, and B channels. is the initial value of the illumination of the pixel (x, y), O c (x, y) is the intensity of the low-light image on channel c at the pixel point (x, y);

[0023] Construct the optimization problem of lighting L:

[0024]

[0025] in, is the distance fidelity term between the optimized lighting and the initial lighting, F refers to the Frobenius norm, is the smoothing term introduced for lighting smoothing, α is the balance coefficient, is the gradient operator.

[0026] Optimizing lighting facilitates subsequent image enhancement processing.

[0027] Furthermore, the method for obtaining the optimal lighting solution is:

[0028] According to the distance fidelity The optimized lighting L and the initial lighting are obtained should be as similar as possible, and the gradient of L should also be Similar, choose To approximate the sparse terms represents the gradient operator, and d represents the direction. Since the gradient of the graph contains vertical and horizontal directions, d has vertical and horizontal options. The lighting optimization problem is rewritten as:

[0029]

[0030] where ∈ is a positive constant used to prevent division by zero; the abbreviation A is used. d (x, y) instead d is used to indicate the direction of the gradient, and the optimization problem is rewritten as:

[0031]

[0032] Differentiate L and set the result to 0, and the solution to the lighting optimization problem is:

[0033]

[0034] Where D is the toplitz matrix obtained by the discrete gradient operator with forward difference, and D contains the vertical direction D v With horizontal direction D h ; I is the identity matrix of appropriate size, l, a is L, A is a vector form, T represents the matrix transpose, It is a simplified gradient operation. The operation Diag(x) constructs a diagonal matrix using the vector x.

[0035] The operation structure is simple and easy to use. It can calculate the optimal solution for lighting, facilitate subsequent combination with reflection, obtain enhanced images, and improve the visual quality of low-light images.

[0036] Furthermore, the method of decomposing the low-light image into a high-frequency layer and a low-frequency layer is as follows:

[0037] O=O low +O high

[0038]

[0039] λ=2σ

[0040] O high =OO low

[0041] Among them, O low represents the low-frequency layer, O high represents the high-frequency layer, σ is the global noise estimate, and l is the balance coefficient, which is defined as twice the global noise estimate σ. This definition helps to divide the noise and detail information into the high-frequency layer.

[0042] According to the characteristics of the high-frequency layer and the low-frequency layer of the image, the high-frequency layer and the low-frequency layer are separated using the total variation and global noise estimation method. The operation is simple and convenient for subsequent use.

[0043] Furthermore, the calculation formula of the global noise estimate σ is as follows:

[0044]

[0045]

[0046] Where * is the convolution operation, (x, y) is the coordinate of the corresponding pixel, W and H are the width and height of the low-light image O, respectively, and N is the proposed noise estimator.

[0047] Obtain a global noise estimate to ensure smooth separation of the high-frequency and low-frequency layers of the image.

[0048] Furthermore, the high-frequency layer is filtered to suppress the noise in the flat area as follows:

[0049]

[0050]

[0051] in, is the high-frequency layer after Gaussian filtering, and the noise in some flat areas is suppressed. c is a constant, G represents the Gaussian filter that performs convolution operation on the high-frequency layer;

[0052] The formula for extracting structure prior from low-frequency layer and texture prior from high-frequency layer is as follows:

[0053]

[0054]

[0055] Among them, S is the structure prior extracted from the low-frequency layer, T is the texture prior extracted from the high-frequency layer, and S in and T in It is the gradient corresponding to the low-frequency layer and the high-frequency layer.

[0056] The high-frequency layer is filtered, and considering the characteristics of the high-frequency layer consisting of details and noise, most of the remaining noise is suppressed.

[0057] Furthermore, the optimization problem of reflection is proposed and solved as follows:

[0058]

[0059] Among them, α, β1, β2, δ1, δ2 are the balance coefficients of each item, The constructed structure prior and texture prior are used as weights to constrain the structure and texture of the reflection R respectively; It is to calculate the difference between the gradient of the reflection and the gradient of the low-frequency layer, so that the reflection maintains the main structure of the original image; It is to calculate the difference between the gradient of the reflection and the gradient of the high-frequency layer, so that the reflection retains the details of the original image; Also used for constraints Convert this formula into vector form:

[0060]

[0061] Here, d is used to represent the diagonal matrix form of the vector, d t =Diag(t),d s =Diag(s),r,o,l,s in ,t in It is R, O, L, S in ,T in The corresponding vector form;

[0062] Differentiating the reflection r and setting the result to 0, we get the solution formula for r:

[0063]

[0064] Extracting the structural and texture prior information to construct the reflection optimization problem is a classic least squares problem. After solving it, the reflection obtained will find a balance between structure and texture, which is beneficial to improving image quality.

[0065] Furthermore, the formula for the enhanced image obtained by combining the corrected illumination optimal solution with the solution obtained by reflection optimization is as follows:

[0066]

[0067] Among them, EO is the enhanced image and γ is set to 2.2.

[0068] Combining the corrected lighting with reflections creates an enhanced image that improves the visual quality of low-light images.

[0069] The present invention also provides a low-light image enhancement system based on structural texture preservation of high- and low-frequency information, comprising an image acquisition module and a processing module. The image acquisition module is used to acquire low-light images. The output end of the image acquisition module is connected to the processing module. The processing module executes the method of the present invention to perform low-light image enhancement.

[0070] The system is beneficial to the low-light image enhancement based on the structural texture of high- and low-frequency information, and has a simple structure and is easy to use. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 An image of a power line captured by a drone under insufficient lighting conditions in a preferred embodiment;

[0072] Figure 2 In order to utilize the method of the present invention Figure 1 An enhanced transmission line image obtained by processing;

[0073] Figure 3 It is a flow chart of a method for enhancing a low-light image by preserving structural texture based on high- and low-frequency information according to the present invention. DETAILED DESCRIPTION

[0074] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0075] In the description of the present invention, it is necessary to understand that the terms "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0076] In the description of the present invention, unless otherwise specified and limited, it should be noted that the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a mechanical connection or an electrical connection, or it can be the internal connection between two components. It can be a direct connection or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to the specific circumstances.

[0077] The present invention discloses a method for enhancing weak-light images based on high- and low-frequency information and maintaining structural texture. The method is widely used in target detection, tracking, image semantic segmentation and other fields as well as in daily life for weak-light images acquired by transmission line inspection drones under insufficient illumination. The original image is a weak-light image in RGB format. For example, Figure 1 As shown, a low-light image of pedestrians at night is collected. Through this embodiment, the brightness of the low-light image is improved while the details and texture are maintained, image enhancement is achieved, and the image with low brightness and low visual quality is enhanced to an image with good visual quality, such as Figure 2 The method comprises the following steps: Figure 3 As shown:

[0078] S1, input the low-light image to be enhanced, and divide the image into the product of reflection and illumination according to the Retinex theory that simulates human visual perception; the method of dividing the image into the product of reflection and illumination is:

[0079] Input the low-light image to be enhanced and record it as a three-dimensional array of double type, denoted as O. The observed image can be decomposed into the product of two components, illumination L and reflection R. O is composed of reflection R and illumination L, denoted as

[0080]

[0081] Among them, (x, y) is the coordinate of the pixel point, It stands for element-wise product. Element-wise product means the point-by-point product of two matrices, such as a11*b11, a22*b22.

[0082] S2, taking the maximum value of the three channels R, G, and B as the initial value of the lighting, while maintaining the smoothness of the lighting segmentation, optimizing the lighting and making the optimized lighting close to its initial value to obtain the optimal lighting solution;

[0083] S3, using global noise estimation and total variation regularization method, decomposes the low-light image into high-frequency layer and low-frequency layer; global noise estimation and total variation regularization method can perform denoising, taking into account the smooth denoising and edge preservation of the image. Applying global noise estimation to the high- and low-frequency decomposition of the image, it is possible to obtain its own global noise estimation according to different images, and adaptively separate the high-frequency layer and the low-frequency layer from the original image.

[0084] S4, Gaussian filtering is performed on the high-frequency layer (Gaussian filtering is a linear smoothing filter suitable for eliminating Gaussian noise), suppressing the noise in the flat area, and extracting the structure prior and texture prior from the high-frequency layer and the low-frequency layer; Gaussian filtering is used on the high-frequency layer, combined with the characteristics of the high-frequency layer consisting of details and noise, to suppress most of the remaining noise in the flat area. This scheme proposes a new reflection optimization problem, incorporating the high-frequency layer and the low-frequency layer into the optimization problem, introducing structure and texture information, and driving the structure prior and texture prior in the reflection solution process to achieve a balance between structure and texture.

[0085] S5, based on high and low frequency information, structure prior and texture prior, the optimization problem of reflection is proposed and solved;

[0086] S6, gamma correction is performed on the optimal lighting solution (also known as gamma nonlinearization or gamma encoding, which is used to perform nonlinear operations or inverse operations on the brightness or tri-color stimulus values ​​of light in a film or imaging system), and the corrected optimal lighting solution is combined with the solution obtained by reflection optimization to obtain an enhanced image.

[0087] In a preferred embodiment of the present invention, the method for optimizing lighting is as follows:

[0088] The initial value of lighting is

[0089]

[0090] Among them, (x, y) is the coordinate of the pixel point, and c refers to the selected R, G, and B channels. is the initial value of the illumination of the pixel (x, y), O c (x, y) is the intensity of the low-light image on channel c at the pixel point (x, y);

[0091] Construct the optimization problem of lighting L:

[0092]

[0093] in, is the optimized lighting L and the initial lighting The distance fidelity term between them is to ensure that the optimized lighting is as close as possible to the set initial lighting. F refers to the Frobenius norm, which is similar to the 2 norm. It is a smoothing term introduced for lighting smoothing. α is a balance coefficient used to control the influence of the lighting smoothing term on the entire optimization problem. In the embodiment, it is set to 0.0015. is a gradient operator used to find the gradient for optimizing lighting.

[0094] Method to find the optimal solution for lighting:

[0095] Fidelity based on distance To obtain the optimal solution to the optimization problem, optimize the lighting L and the initial lighting should be as similar as possible, and the gradient of L should also be Similar, choose To approximate the sparse terms d represents the direction. Since the gradient of the graph contains vertical and horizontal directions, d has vertical and horizontal options. The lighting optimization problem is rewritten as:

[0096]

[0097] Among them, ∈ is a very small positive constant used to prevent division by zero; for convenience, the abbreviation A is used. d (x, y) instead Since the gradient of the image has horizontal and vertical directions, d is used to represent the direction of the gradient, which has two values: vertical and horizontal. The optimization problem is rewritten as:

[0098]

[0099] The above equation contains only quadratic terms and can be regarded as a typical least squares problem with an optimal solution. For computational efficiency, L is differentiated and the result is set to 0. The solution to the lighting optimization problem is:

[0100]

[0101] Where D is the toplitz matrix obtained by the discrete gradient operator with forward difference, and D contains the vertical direction D v With horizontal direction D h ; I is the identity matrix of appropriate size, l, a is L, A is in vector form (a vector is a vector, converting an mxn matrix into a 1xmn vector), T represents matrix transpose, To simplify the gradient operation, the operation Diag(x) constructs a diagonal matrix using the vector x.

[0102] In a preferred embodiment of the present invention, a method for decomposing a low-light image into a high-frequency layer and a low-frequency layer is as follows:

[0103] O=O low +O high

[0104]

[0105] λ=2σ

[0106] O high =OO low

[0107] Among them, O low represents the low-frequency layer, O high represents the high-frequency layer, σ is the global noise estimate, and λ is the balance coefficient, which is used to determine the impact of the total variational sparse term of the low-frequency layer on the entire high- and low-frequency optimization problem. According to the image, it can be decomposed into two components: the high-frequency layer and the low-frequency layer. The high-frequency component mainly contains details and most of the noise, while the low-frequency component mainly contains a lot of structural information of the original image. In separating the high- and low-frequency layers of the input image, the optimization formula Ensure that the low-frequency layer is as close to the input image as possible, and The purpose is to ensure that the separated low-frequency layer is as smooth as possible without interference from high-frequency information such as noise. Define λ as twice the global noise estimate σ, which helps to divide the noise and detail information into high-frequency layers.

[0108] The global noise estimate σ is calculated as follows:

[0109]

[0110]

[0111] Where * is the convolution operation, (x, y) is the coordinate of the corresponding pixel point, W and H are the width and height of the low-light image O, respectively, and N is the proposed noise estimator, which regards thin lines as noise. Therefore, the noise map separated by it can be regarded as a high-frequency information map.

[0112] The calculation of Gaussian filtering on the high-frequency layer and suppressing the noise in the flat area is as follows:

[0113]

[0114]

[0115] in, is the high-frequency layer after Gaussian filtering, and the noise in some flat areas is suppressed. c is a constant, G represents a Gaussian filter that performs a convolution operation on the high-frequency layer; in the embodiment, δ in the Gaussian function c It is set to 1.2 and the sliding window of the filter is set to 5x5 so that the noise in some flat areas can be suppressed without destroying the detailed information too much.

[0116] The formula for extracting structure prior from low-frequency layer and texture prior from high-frequency layer is as follows:

[0117]

[0118]

[0119] Among them, S is the structure prior extracted from the low-frequency layer, T is the texture prior extracted from the high-frequency layer, and S in and T in It is the gradient corresponding to the low-frequency layer and the high-frequency layer.

[0120] The method of proposing and solving the reflection optimization problem is as follows:

[0121]

[0122] Among them, α, β1, β2, δ1, δ2 are the balance coefficients of each item. β1, β2, δ1, δ2 can be set to 0.1, 0.1, 0.05, 0.005 respectively. The constructed structure prior and texture prior are used as weights to constrain the structure and texture of the reflection R, ensuring that the gradient of the reflection R is as close as possible to the gradient of the low-frequency layer and the high-frequency layer, that is, the extracted structure and texture priors; It is to calculate the difference between the gradient of the reflection and the gradient of the low-frequency layer, so that the reflection maintains the main structure of the original image (i.e., low-frequency information); It is to calculate the difference between the gradient of the reflection and the gradient of the high-frequency layer, so that the reflection retains the details of the original image (i.e., high-frequency information); Also used for constraints Ensure that the gradient of the reflection R is closer to the gradient of the low-frequency layer and the high-frequency layer, rather than farther away, because when minimizing the value of the optimization problem, the distance between the gradients will also make the function get a smaller value. Convert this formula into vector form:

[0123]

[0124] Here, d is used to represent the diagonal matrix form of the vector, d t =Diag(t),d s =Diag(s),r,o,l,s in ,t inIt is R, O, L, S in ,T in The corresponding vector form;

[0125] Differentiating the reflection r and setting the result to 0, we get the solution formula for r:

[0126]

[0127] The formula for the enhanced image obtained by combining the corrected illumination optimal solution with the solution obtained by reflection optimization is as follows:

[0128]

[0129] Among them, EO is the enhanced image and γ is set to 2.2.

[0130] The present invention also provides a low-light image enhancement system based on structural texture preservation of high- and low-frequency information, comprising an image acquisition module and a processing module. The image acquisition module is used to acquire low-light images. The output end of the image acquisition module is electrically connected to the processing module. The processing module executes the method described in the present invention to perform low-light image enhancement.

[0131] This scheme decomposes the input image into two components, illumination and reflection, based on the weak light enhancement method of Retinex. Since a large amount of noise in the reflection will affect the illumination during the alternating iteration of illumination and reflection, solving illumination and reflection separately in sequence can effectively avoid the noise transfer problem. According to the characteristics of the high-frequency layer and low-frequency layer of the image, the image is decomposed into high-frequency layer and low-frequency layer using the total variation and global noise estimation method, and the structure and texture information are extracted from them respectively, so that the optimized reflection can maintain a balance between texture and structure. In addition, it can maintain the naturalness of non-uniform illumination and suppress noise. Improve the brightness of weak light images and maintain details and textures to achieve image enhancement, for example, enhance low brightness and low visual quality images to images with better visual quality. For images obtained by transmission line inspection drones under insufficient light, the present invention improves real-time performance while enhancing image visibility.

[0132] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0133] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A method for low-light image enhancement based on structural texture preservation of high- and low-frequency information, characterized in that: The steps include: Input the low-light image to be enhanced and divide the image into the product of reflection and illumination; The maximum value of the three channels R, G, and B is used as the initial value of the lighting. While maintaining the smoothness of the lighting segmentation, the lighting is optimized and the optimized lighting is made close to its initial value to obtain the optimal lighting solution. Decompose the low-light image into high-frequency layer and low-frequency layer; Filter the high-frequency layer to suppress the noise in the flat area, and extract the structure prior and texture prior from the high-frequency layer and the low-frequency layer; Based on high and low frequency information, structure prior and texture prior, the optimization problem of reflection is proposed and solved; Correct the optimal lighting solution, and combine the corrected optimal lighting solution with the solution obtained by reflection optimization to obtain an enhanced image; The calculation for filtering the high frequency layer and suppressing the noise in the flat area is as follows: in, is the high-frequency layer after Gaussian filtering, and the noise in some flat areas is suppressed. c is a constant, G represents the Gaussian filter that performs convolution operation on the high-frequency layer; The formula for extracting structure prior from low-frequency layer and texture prior from high-frequency layer is as follows: Among them, S is the structure prior extracted from the low-frequency layer, T is the texture prior extracted from the high-frequency layer, and S in and T in is the gradient corresponding to the low-frequency layer and the high-frequency layer; The method of proposing and solving the reflection optimization problem is as follows: Among them, α, β1, β2, δ1, δ2 are the balance coefficients of each item, The constructed structure prior and texture prior are used as weights to constrain the structure and texture of the reflection R respectively; It is to calculate the difference between the gradient of the reflection and the gradient of the low-frequency layer, so that the reflection maintains the main structure of the original image; It is to calculate the difference between the gradient of the reflection and the gradient of the high-frequency layer, so that the reflection retains the details of the original image; Also used for constraints Convert this formula into vector form: Here, d is used to represent the diagonal matrix form of the vector, d t =Diag(t),d s =Diag(s),r,o,l,s in ,t in It is R, O, L, S in ,T in The corresponding vector form; Differentiating the reflection r and setting the result to 0, we get the solution formula for r:

2. The method for low-light image enhancement based on structural texture preservation of high- and low-frequency information according to claim 1, characterized in that: The method to separate the image into the product of reflection and illumination is: Input the low-light image to be enhanced and record it as O, O is composed of reflection R and illumination L, recorded as Among them, (x, y) is the coordinate of the pixel point, Represents element-wise product.

3. The method for low-light image enhancement based on structural texture preservation of high- and low-frequency information as claimed in claim 1, characterized in that: The method for optimizing lighting is as follows: The initial value of lighting is Among them, (x, y) is the coordinate of the pixel point, and c refers to the selected R, G, and B channels. is the initial value of the illumination of the pixel (x, y), O c (x, y) is the intensity of the low-light image on channel c at the pixel point (x, y); Construct the optimization problem of lighting L: in, is the distance fidelity term between the optimized lighting and the initial lighting, F refers to the Frobenius norm, is the smoothing term introduced for lighting smoothing, α is the balance coefficient, is the gradient operator.

4. The method for enhancing low-light images based on structural texture preservation of high- and low-frequency information as claimed in claim 3, characterized in that: Method to find the optimal solution for lighting: According to the distance fidelity The optimized lighting L and the initial lighting are obtained should be as similar as possible, and the gradient of L should also be Similar, choose To approximate the sparse terms represents the gradient operator, and d represents the direction. Since the gradient of the graph contains vertical and horizontal directions, d has vertical and horizontal options. The lighting optimization problem is rewritten as: Here, ò is a positive constant used to prevent division by zero; the abbreviation A is used. d (x, y) instead d is used to indicate the direction of the gradient, and the optimization problem is rewritten as: Differentiate L and set the result to 0, and the solution to the lighting optimization problem is: Where D is the toplitz matrix obtained by the discrete gradient operator with forward difference, and D contains the vertical direction D v With horizontal direction D h ; I is the identity matrix of appropriate size, l, a is L, A is a vector form, T represents the matrix transpose, It is a simplified gradient operation. The operation Diag(x) constructs a diagonal matrix using the vector x.

5. The method for enhancing low-light images based on structural texture preservation of high- and low-frequency information according to claim 1, characterized in that: The method of decomposing a low-light image into a high-frequency layer and a low-frequency layer is as follows: O=O low +O high λ=2σ The high =OO low Among them, O low represents the low-frequency layer, O high represents the high-frequency layer, σ is the global noise estimate, and λ is the balance coefficient, which is defined as twice the global noise estimate σ. This definition helps to divide the noise and detail information into the high-frequency layer.

6. The method for enhancing low-light images based on structural texture preservation of high- and low-frequency information as claimed in claim 5, characterized in that: The calculation formula of the global noise estimate σ is as follows: Where * is the convolution operation, (x, y) is the coordinate of the corresponding pixel, W and H are the width and height of the low-light image O, respectively, and N is the proposed noise estimator.

7. The method for low-light image enhancement based on structural texture preservation of high- and low-frequency information as claimed in claim 1, characterized in that: The formula for the enhanced image obtained by combining the corrected illumination optimal solution with the solution obtained by reflection optimization is as follows: Among them, EO is the enhanced image and γ is set to 2.

2.

8. A low-light image enhancement system based on structural texture preservation of high- and low-frequency information, characterized in that: It comprises an image acquisition module and a processing module, wherein the image acquisition module is used to acquire weak-light images, the output end of the image acquisition module is connected to the processing module, and the processing module executes the method according to one of claims 1 to 7 to perform weak-light image enhancement.

Images