Low-illumination gray image enhancement and denoising method

Abstract

The application relates to a low-illumination gray-scale image enhancement and denoising method, which comprises the following steps: constructing an optimal exposure ratio image according to an input low-illumination gray-scale image; decomposing the low-illumination gray-scale image and the optimal exposure ratio image by adopting a LatLRR method to obtain a low-rank part and a significant part; constructing a low-rank part weight and a significant part weight according to the obtained low-rank part and the significant part; performing normalization processing on the obtained low-rank part weight and the significant part weight to obtain a normalized weight; and performing image reconstruction and synthesis according to the low-rank part, the significant part and the normalized weight to obtain a multi-exposure fusion enhanced image. The application can enhance and denoise a single low-illumination image, restore structure information, maintain details, avoid block effects and reduce the calculation complexity.

Description

Technical Field

[0001] This invention relates to a method for enhancing and denoising low-light grayscale images. Background Technology

[0002] Existing technologies suffer from poor practicality and generalizability in low-light grayscale images: Most existing technologies are designed for color images. For example, the Retinex method is based on the premise that each color region can be composed of the three primary colors of red, green, and blue with fixed wavelengths; when solving the illumination map, the brightness values ​​of the three color channels are used. When applying this type of method to grayscale images, the grayscale image must first be converted into a pseudo-color image (red channel = green channel = blue channel = original grayscale image). This makes the Retinex method unsuitable for solving the true illumination map. For enhancement methods based on multi-exposure fusion, compared to color images, grayscale images contain less useful information for constructing deweighting factors, resulting in poor detail recovery or block artifacts. At the same time, existing technologies lack methods that can simultaneously address random noise in low-light images. In practical engineering applications, it is difficult to obtain multiple images of the same scene with different exposures simultaneously, which also limits deep learning methods. Summary of the Invention

[0003] Therefore, it is necessary to provide a method for enhancing and denoising low-light grayscale images.

[0004] This invention provides a method for enhancing and denoising low-light grayscale images, the method comprising the following steps: a. Based on the input low-light grayscale image X src Construct an image X with optimal exposure ratio vir b. The LatLRR method is used to process the low-light grayscale image X. src and the optimal exposure ratio image X vir Decomposition yields the low-rank part and significant parts c. Construct the low-rank component weights based on the obtained low-rank and salient components. and significant part weights d. Obtain the weights of the lower-rank components. and significant part weights Normalization is performed to obtain normalized weights. e. Based on the low-rank component, the salient component, and the normalized weights, the image is reconstructed to obtain... and Then and The image is enhanced by combining multiple exposures and fusion.

[0005] Specifically, step a includes:

[0006] For any pixel p within the defined region g The corresponding level l is calculated using Formula 1. g :

[0007] The optimal exposure ratio image X is constructed using Formula 1. vir :

[0008]

[0009] Where γ=X white X d =X src X w,max This represents the maximum brightness of the mapped image.

[0010] Specifically, step a further includes:

[0011] Based on Formula 2, the information entropy criterion is used to find the image with the optimal exposure ratio in Formula 1:

[0012]

[0013] Specifically, step b includes:

[0014] X src X vir Substituting these values ​​into the formula below, we obtain the low-rank component and the significant component after decomposition:

[0015]

[0016] st,X=XZ+LX+E Formula 5

[0017] λ represents the weighting factor r, ||.|| * Let ||.||1 represent the nuclear norm, where ||.||1 is the l1-norm, and X represents the input image to be decomposed. XZ(X L ) represents the lower-rank part, LX(X) S () represents the salient part, and E represents the noisy part. Here, image noise is separated.

[0018] The LatLRR method was used to analyze X. src and X vir Decomposition yields the low-rank part and significant parts

[0019] Specifically, step c includes:

[0020] The method for constructing the weights of the lower-rank component is as follows:

[0021]

[0022] D0 = |conv(X) L *H laplacian )|*ψ2 Formula 7

[0023]

[0024] In the formula, H laplacian This is the Laplacian operator, conv(.) is the convolution symbol, D0 is the initial contrast weight factor, ψ2 is the two-dimensional Gaussian filtering method, and x represents the pixels in the image. λ1 and ||.||1 represent the l2 and l1 norms respectively, and λ1 and λ2 are weighting factors. and These are the first-order gradients in the vertical and horizontal directions, respectively. Ω(x) is the w×w neighborhood centered at x, and y is the pixel within Ω(x).

[0025] Specifically, step c further includes:

[0026] The method for constructing the weights of the salient components is as follows:

[0027] D s (x)=||X S (x)-μ S (x)|| a Formula 9

[0028] Where, μ S It is the average value of the low-rank portion, and 'a' depends on the image.

[0029] Using the above formula, after low-rank decomposition, X src X vir The low-rank parts of each decomposition and significant parts Weighted graphs of the low-rank and salient parts were constructed for both, where the weights of the low-rank parts are: and The weight of the significant part is and

[0030] Specifically, step d includes:

[0031] Weight the lower-rank part and significant part weights Normalization process is performed to obtain and

[0032]

[0033]

[0034]

[0035]

[0036] Specifically, step e includes:

[0037] The results obtained from each decomposition Image synthesis and Image synthesis The formula is as follows:

[0038]

[0039]

[0040] Finally, and The images are synthesized to obtain the enhanced image:

[0041]

[0042] Compared with existing technologies, this application can enhance and denoise single low-light images. Based on a multi-exposure fusion framework, this application first proposes to construct an optimal virtual exposure image using an information entropy constraint criterion, building upon the low-light mapping tone mapping operator. Then, a low-rank decomposition method is used to decompose the intermediate virtual image and the original image to obtain low-rank and salient parts. Subsequently, an initial weight map is constructed based on the different information contained in each of the decomposed images. An adaptive weight correction algorithm is proposed to maximize the restoration of structural information and preserve details while avoiding block artifacts. In the weight map solution process, this application transforms the decomposition and optimization of nonlinear problems into a total variational model, employing an iterative method to reduce computational complexity. Attached Figure Description

[0043] Figure 1 This is a flowchart of the low-light grayscale image enhancement and denoising method of the present invention;

[0044] Figure 2 This is a schematic diagram illustrating the processing steps of the low-light grayscale image enhancement and denoising method provided in an embodiment of the present invention.

[0045] Figure 3 A schematic diagram of an input low-light grayscale image provided in an embodiment of the present invention;

[0046] Figure 4 A schematic diagram of the constructed optimal exposure ratio image provided in an embodiment of the present invention;

[0047] Figure 5 This is a schematic diagram of the enhanced image provided in an embodiment of the present invention. Detailed Implementation

[0048] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0049] Please see Figure 1 This is a flowchart of a preferred embodiment of the low-light grayscale image enhancement and denoising method of the present invention (please refer to it as well). Figure 2 ).

[0050] Step S1, based on the input low-light grayscale image X src (Please refer to) Figure 3 Construct an image X with optimal exposure ratio vir (Please refer to) Figure 4 Specifically:

[0051] The optimal exposure ratio image X is constructed using Formula 1. vir :

[0052]

[0053] In Formula 1, γ = X white X d =X src X w,max The maximum brightness is represented by Equation 2. Using the information entropy criterion, the optimal exposure ratio image in Equation 1 is found, as shown in Equation 2 below:

[0054]

[0055] The formula for calculating information entropy is shown in Formula 3:

[0056]

[0057] In Formula 1, p i is the probability of grayscale value i appearing in the image, N is the maximum grayscale value of the image, which is 255 in this embodiment, and H is the image information entropy.

[0058] Step S2, use the LatLRR method to process the low-light grayscale image X. src and the optimal exposure ratio image X vir Decomposition yields the low-rank part and significant parts Specifically:

[0059] X src and the constructed optimal exposure ratio image X vir Substituting these values ​​into the formula below, we obtain the low-rank component and the significant component after decomposition:

[0060]

[0061] st,X=XZ+LX+E Formula 5

[0062] λ represents the weighting factor r, ||.|| * Let |||1| represent the nuclear norm, where |||1 is the l1-norm, and X represents the input image to be decomposed. XZ(X L ) represents the lower-rank part, LX(X) S () represents the significant part, and E represents the noise part. Here, the image noise is separated.

[0063] The LatLRR method was used to analyze X. src and X vir Decomposition yields the low-rank part and significant parts

[0064] Step S3: Construct the low-rank component weights based on the obtained low-rank and salient components. and significant part weights Specifically:

[0065] Step S31, the method for constructing the low-rank partial weights includes:

[0066]

[0067] D0 = |conv(X) L *H laplacian )|*ψ2 Formula 7

[0068]

[0069] In the formula, H laplacian This is the Laplacian operator, conv(.) is the convolution symbol, D0 is the initial contrast weight factor, and ψ2 is a two-dimensional Gaussian filtering method. Here, x represents the pixels in the image. λ1 and ||.||1 represent the l2 and l1 norms respectively, and λ1 and λ2 are weighting factors. and These are the first-order gradients in the vertical and horizontal directions, respectively. Ω(x) is a w×w neighborhood centered at x, and y is a pixel within Ω(x).

[0070] The solution process for the above formula is the same as that of the variational RTV method, and will not be repeated here.

[0071] Step S32, the method for constructing the weights of the significant components includes:

[0072] D s (x)=||X S (x)-μS (x)|| a Formula 9

[0073] Where, μ S It is the average value of the low-rank portion, and 'a' depends on the image.

[0074] Using the above formula, after low-rank decomposition, the original image X src and intermediate process image X vir The low-rank parts of each decomposition and significant parts Weighted graphs of the low-rank and salient components were constructed for both. The weights of the low-rank components are as follows: and The weight of the significant part is and

[0075] Step S4, obtain the low-rank partial weights and significant part weights Normalization is performed to obtain normalized weights. Specifically:

[0076] Weight the lower-rank part and significant part weights Normalization process is performed to obtain and

[0077]

[0078]

[0079]

[0080]

[0081] Step S5: Based on the low-rank component, the salient component, and the normalized weights, perform image reconstruction to obtain... and Then and The image is enhanced by combining multiple exposures and fusion. (Please refer to) Figure 5 ):

[0082] The results obtained from each decomposition Image synthesis and Image synthesis The formula is as follows:

[0083]

[0084]

[0085] Finally, and The images are synthesized to obtain the enhanced image:

[0086]

[0087] Although the present invention has been described with reference to the present preferred embodiments, those skilled in the art should understand that the above preferred embodiments are only used to illustrate the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for enhancing and denoising low-light grayscale images, characterized in that, The method includes the following steps: a. Based on the input low-light grayscale image Construct an image with optimal exposure ratio ; b. The low-light grayscale image is processed using the LatLRR method. and the image with the optimal exposure ratio Decomposition yields the low-rank part ={ , } and significant parts ={ , }; c. Construct the low-rank component weights based on the obtained low-rank and salient components. , and significant part weights , ; d. Obtain the weights of the lower-rank components. , and significant part weights , Normalization is performed to obtain normalized weights. , , , ; e. Based on the low-rank component, the salient component, and the normalized weights, the image is reconstructed to obtain... and Then and The image is enhanced by combining multiple exposures and fusion. ;in: Step a includes: For any pixel p within the defined region g The corresponding level is calculated using Formula 1. : Construct the optimal exposure ratio image using Formula 1 : Official 1 in, , = , This represents the maximum brightness of the mapped image.

2. The method as described in claim 1, characterized in that, Step a further includes: Based on Formula 2, the information entropy criterion is used to find the image with the optimal exposure ratio in Formula 1: Official 2.

3. The method as described in claim 2, characterized in that, Step b specifically includes: Will , Substituting these values ​​into the formula below, we obtain the low-rank component and the significant component after decomposition: Official 4 Official 5 Represents the weighting factor r, Represents the nuclear norm. yes , Represents the input image to be decomposed, XZ( ) represents the lower-rank part, LX( () represents the salient part, and E represents the noisy part. Here, image noise is separated. Using the LatLRR method for and Decomposition yields the low-rank part ={ , } and significant parts ={ , } 4. The method as described in claim 3, characterized in that, Step c includes: The method for constructing the weights of the lower-rank component is as follows: Official 6 Official 7 Official 8 In the formula, It is the Laplace operator. It is a convolution symbol. This is the initial contrast weighting factor. It is a two-dimensional Gaussian filtering method. Representing pixels in an image, and Represent and Norm, and As a weighting factor, and These are the first-order gradients in the vertical and horizontal directions, respectively. Therefore Centered The neighborhood of y is The pixels within.

5. The method as described in claim 4, characterized in that, Step c further includes: The method for constructing the weights of the salient components is as follows: Official 9 in, It is the average value of the lower-rank portion. It depends on the image; Using the above formula, after low-rank decomposition... , The low-rank parts of each decomposition ={ , } and significant parts ={ , Weighted graphs of the low-rank and salient parts were constructed for each part, where the weights of the low-rank parts are: and The weight of the significant part is and .

6. The method as described in claim 5, characterized in that, Step d specifically includes: Weight the lower-rank part , and significant part weights , Normalization process is performed to obtain , , and : Official 10 Official 11 Official 12 Official 13.

7. The method as described in claim 6, characterized in that, Step e specifically includes: The results obtained from each decomposition ={ , Image synthesis ,and ={ , Image synthesis The formula is as follows: Official 14 Official 15 Finally, and The images are synthesized to obtain the enhanced image: Official 16.

Images