Low-light image enhancement method based on multi-scale perception enhancement structure block decomposition

Abstract

The invention is applied to the technical field of low-light image enhancement, and particularly discloses a low-light image enhancement method based on multi-scale perception enhancement structure block decomposition, which comprises the following steps: firstly, carrying out SPD decomposition on a low-illumination image to obtain four components including perception gain, signal intensity, signal structure and average intensity; a perception gain component, a signal intensity component, a signal structure component and an average intensity component are respectively enhanced through an artificially set enhancement strategy, and the four enhanced components are fused into one image again through a multi-scale framework. According to the low-light image enhancement method based on multi-scale perception enhancement structure block decomposition, the image is subjected to adjustment enhancement after being decomposed so as to generate a result with rich information and vivid perception, a synthesized multi-exposure image is obtained by adopting simple logarithm stretching, the enhancement performance can be further improved, the enhanced image is more natural and real, and the image quality is improved. The method has advantages and effectiveness in the aspects of keeping detail information and keeping image consistency.

Description

Technical Field

[0001] The present invention relates to the technical field of low-light-level image enhancement, and in particular to a low-light-level image enhancement method based on multi-scale perception enhancement structural block decomposition. Background Art

[0002] In complex environments, imaging equipment is difficult to achieve ideal imaging effects due to exposure time limitations. Therefore, the captured images exhibit problems such as low brightness, high noise, low contrast, and color distortion. In order to restore image quality and obtain sufficient information from the image to meet the requirements of advanced computer vision processing tasks such as target detection and tracking under low-light conditions, enhancing such poor-quality low-light images has received widespread attention in the field of computer vision.

[0003] To enhance low-light images, existing common methods, such as the zero-reference depth curve estimation algorithm, have fast processing speed but limited noise suppression and enhancement capabilities. Histogram equalization is used to redistribute pixels concentrated in low grayscale levels evenly across all grayscale levels, making the histogram distribution of the enhanced image wider. However, the model processes the three channels of the color image independently, resulting in serious color deviation. The dark channel dehazing theory is applied to low-light image enhancement. That is, the inverted image is dehazed and then inverted again to obtain a low-light enhanced image. Due to its reliance on dark channel theory, this type of algorithm does not fully enhance the image brightness. Deep learning has also been successfully applied to image enhancement tasks such as image denoising and image super-resolution. By formulating the decomposition problem as an implicit prior regularization model, three learning-based modules are designed, responsible for data-dependent initialization, efficient unrolling optimization, and user-specified illumination enhancement. The algorithm has weak generalization ability, monotonous colors, and blurring, which loses a lot of structural information in the image, resulting in poor quality. The enhancement results of various existing methods suffer from monotonous colors, insufficient enhancement, image inconsistency, and loss of details. Summary of the Invention

[0004] The purpose of the present invention is to provide a low-light image enhancement method based on multi-scale perceptual enhancement block decomposition to solve the problems of color monotony, insufficient enhancement, image inconsistency and detail loss in the enhancement results proposed in the above background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a low-light-level image enhancement method based on multi-scale perception enhancement block decomposition, comprising the following steps: S1, first perform SPD decomposition on the low-light image to obtain the four components of perceptual gain, signal strength, signal structure and average strength; S2, enhance the four components of perception gain, signal strength, signal structure and average strength respectively through manually set enhancement strategies; S3, the enhanced four components are re-fused into one image through a multi-scale framework.

[0006] Preferably, the perceptual gain enhancement is set with a low threshold value Cl, and the perceptual gain enhancement is based on the idea of pseudo-fog image, which is to invert the low-illumination image and then enhance it based on the atmospheric scattering model; The specific relationship of the atmospheric scattering model is: ⊙T+A⊙(1-T); in For the collected images, is the image to be restored, A is the atmospheric light intensity, T is the medium transmittance, and ⊙ represents the Hadamard product; The illumination intensity matrix A is taken as 1, and the medium transmittance T is T 1-ω Make estimates; Where ω is the fidelity coefficient, which is the global mean of the inverted image, and I is the inverted image of the input image X; The quantization noise low threshold Cl is used as a small bias of the medium transmittance T, using J= +A= +A , perform brightness enhancement on the input image; Get a preliminary estimate of the normal exposure image , the grayscale value is limited to [0, 1], which is used for the next step of structural decomposition; The mean value of the patch is used as the perceptual gain i of the local patch mean(x).

[0007] The above technical solution can moderately enhance the contrast and brightness of the image while retaining image details, avoid distortion caused by excessive enhancement, improve the visibility of the image in low-light environments, and have good adaptability to different types of images.

[0008] Preferably, the perceptual gain is denoted as I after being aggregated into the entire image, and the maximum perceptual gain of all patches is used to estimate the fused perceptual gain, combined with the characteristics of the input image X; The expression of perception gain Î is: Î ; in To obtain the global maximum value of the image, To obtain the global mean of the image, the value of p is 2 for all layers except the 0th layer, which is 1.

[0009] By using the above technical solution, the low-light image is inverted and enhanced in an atmospheric scattering mode, which can effectively improve the image brightness and contrast and highlight the details. At the same time, simulating atmospheric scattering can make the enhancement process more natural, avoid over-processing, and improve the visual effect of low-light images.

[0010] Preferably, the perceptual gain is a global variable, and the range of the perceptual gain i is limited to [0.55, 0.85].

[0011] The above technical solution can enhance the contrast between areas of different brightness in an image.

[0012] Preferably, the signal intensity is calculated as the signal intensity of the local plaque, and the signal intensity of the local plaque is obtained by decomposition calculation. ║ - ║, and aggregated into C, based on C, a logarithmic linear stretching transformation is performed, the formula is k), where k is the maximum value in C rounded up, and Take the larger value in the image.

[0013] The above technical solution can ensure that the structural information and intensity information of the image are coordinated with each other, avoiding the situation where the image structure is deformed or blurred due to improper intensity adjustment.

[0014] Preferably, the structural strength , signal structure Combine The formula is ⊙ ⊙( , where p is an exponential parameter greater than or equal to 0, and its value is 4.

[0015] The above technical solution can enhance dark details, adjust bright areas, improve image contrast and visibility, and make the overall processing more adaptable and effective.

[0016] Preferably, the average intensity for the image at non-minimum scale uses the mean of each local patch as the brightness value of the patch. For the image at minimum scale, the average intensity is calculated for each local patch and then subjected to a nonlinear stretching, using an exponential function. , perform benchmark brightness compensation, and the average brightness after aggregation is recorded as , .

[0017] By adopting the above technical solution, low-frequency information can retain the overall outline and general structure of the image, while high-frequency information can highlight the details and edges.

[0018] Preferably, the plaque reconstruction image formula of the image at the minimum scale is: ), the image patch aggregation formula at a certain scale is ⊙ ) ⊙B)⊙ ⊙B⊙ ), where B , ) indicates that local mean filtering is performed.

[0019] The above technical solution is adopted to enhance the image clarity and stabilize the image.

[0020] Preferably, the formula for reconstructing the image from the patches of the image at the minimum scale is It consists of low-frequency information and high-frequency information.

[0021] By adopting the above technical solution, the combination of low-frequency information and high-frequency information makes the reconstructed image better. Preferably, the formula of the low-frequency information is expressed as L ; The formula of the low-frequency information is expressed as H ; Downsample by 2 times to obtain the second scale information ,based on , we can calculate 、 、 and ; According to the formula ） ） Get high-frequency information of the second scale ; Decomposition continues until the smallest scale level , Z , where h and w are the height and width of the original scale image respectively; The low-frequency information at different scales can be calculated as The low-frequency information and high-frequency information are finally fused through the bilinear upsampling operator. , Upsampling; The formula is .

[0022] The above technical solution enables more targeted processing to achieve color balance in the image.

[0023] Compared with the prior art, the present invention has the following advantages: the low-light-level image enhancement method based on multi-scale perception enhancement block decomposition has the following advantages: 1. This low-light image enhancement method, based on multi-scale perceptual enhancement block decomposition, operates within a multi-scale framework, better capturing features at different scales within the image. This allows for targeted enhancement, resulting in a more natural and realistic enhanced image. It demonstrates advantages and effectiveness in preserving detail and maintaining image consistency. 2. Furthermore, an improved dehazing model is proposed to extract the initial brightness component estimated by the perceptual gain module, thereby obtaining richer color information and more accurate brightness components. This model can adapt to varying degrees of low-light conditions and achieves good enhancement effects even for images in extremely low light, achieving excellent performance in various scenarios. 3. Furthermore, the structural block decomposition method helps to improve the efficiency and stability of the algorithm. By decomposing the image into different structural blocks, it can be processed more targeted, reducing the amount of calculation while increasing the processing speed. It can maintain the color balance and tonal consistency of the image and avoid problems such as color distortion. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a schematic diagram of the structure of the steps of the present invention; Figure 2 Schematic diagram of the algorithm structure of the present invention; Figure 3 Schematic diagram of the structure of the perception gain algorithm of the present invention; Figure 4 Schematic diagram of the signal strength enhancement algorithm structure of the present invention; Figure 5 Schematic diagram of the average brightness enhancement structure of the present invention; Figure 6 This is a schematic diagram of the fusion and reconstruction structure of the present invention; Figure 7 This is a schematic diagram of the low-frequency information and high-frequency information structures of the present invention; Figure 8 Schematic diagram of the effects of different stretching functions of the present invention; Figure 9 This is a schematic diagram of the high and low frequency visualization information structure of the present invention. DETAILED DESCRIPTION

[0025] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0026] See also Figure 1-9The present invention provides a technical solution: a low-light image enhancement method based on multi-scale perception enhancement structural block decomposition, comprising the following steps: S1, first perform SPD decomposition on the low-light image to obtain the four components of perceptual gain, signal strength, signal structure and average strength; S2, enhance the four components of perception gain, signal strength, signal structure and average strength respectively through manually set enhancement strategies; S3, the enhanced four components are re-fused into one image through a multi-scale framework; The perceptual gain enhancement is set with a low threshold Cl. The idea of perceptual gain enhancement through pseudo-fog image is to invert the low-light image and then enhance it based on the atmospheric scattering model; The specific relationship of the atmospheric scattering model is: ⊙T+A⊙(1-T); in For the collected images, is the image to be restored, A is the atmospheric light intensity, T is the medium transmittance, and ⊙ represents the Hadamard product; The illumination intensity matrix A is taken as 1, and the medium transmittance T is T 1-ω Make estimates; Where ω is the fidelity coefficient, which is the global mean of the inverted image, and I is the inverted image of the input image X; The quantization noise low threshold Cl is used as a small bias of the medium transmittance T, using J= +A= +A , perform brightness enhancement on the input image; Get a preliminary estimate of the normal exposure image , the grayscale value is limited to [0, 1], which is used for the next step of structural decomposition; The mean value of the patch is used as the perceptual gain i of the local patch mean(x); The perceptual gain is denoted as I after being aggregated into the entire image. The maximum perceptual gain among all patches is used to estimate the fused perceptual gain, combined with the features of the input image X. The expression of perception gain Î is: Î ; in To obtain the global maximum value of the image, To obtain the global mean of the image, the value of p is 2 for all layers except the 0th layer, which is 1; Perceptual gain is a global variable, and the range of perceptual gain i is limited to [0.55, 0.85]; Signal intensity calculation of the local plaque signal intensity, the local plaque signal intensity is calculated by decomposition c ║ - ║, and aggregated into C, based on C, a logarithmic linear stretching transformation is performed, the formula is k), where k is the maximum value in C rounded up, and Take the larger value in the image; Structural strength , signal structure Combine The formula is ⊙ ⊙( , where p is an exponential parameter greater than or equal to 0, and its value is 4; For images at non-minimum scale, the average intensity uses the mean of each local patch as the brightness value of the patch. For images at minimum scale, the average intensity calculates the mean of each local patch and then performs a nonlinear stretch on it, using an exponential function. , perform benchmark brightness compensation, and the average brightness after aggregation is recorded as , ; The image reconstruction formula of the patch at the minimum scale is: ), the image patch aggregation formula at a certain scale is ⊙ ) ⊙B)⊙ ⊙B⊙ ), where B , ) indicates that local mean filtering is performed; Formula for reconstructing the image from patches at the minimum scale It is composed of low-frequency information and high-frequency information; The formula for low-frequency information is L ; The formula for low-frequency information is H ; Downsample by 2 times to obtain the second scale information ,based on , we can calculate 、 、 and ; According to the formula ） ） Get high-frequency information of the second scale ; Decomposition continues until the smallest scale level , Z , where h and w are the height and width of the original scale image respectively; The low-frequency information at different scales can be calculated as , the low-frequency information and high-frequency information are finally fused through the bilinear upsampling operator , Upsampling; The formula is ; like Figure 2 As shown in the figure, first the low-light image is input, and then the multi-scale structure decomposition and fusion strategy will be used to preserve its details and enhance its brightness. The defogging model is used to obtain the perceptual gain of the image. Considering the quantization noise of the imaging device, a low threshold Cl is set in the algorithm. The area below this value is not enhanced. Then the perceptual gain is estimated based on the idea of pseudo-fog image. The idea of pseudo-fog image is to invert the low-light image and then enhance it based on the atmospheric scattering model. The atmospheric scattering model is ⊙T+A⊙(1-T), then in order to simplify the calculation, the illumination intensity matrix A is taken as 1, and the medium projection rate T is T 1-ω , the fidelity coefficient is taken as the global mean of the inverted image, where I is the inverted image of the input image x. At the same time, in order to avoid the denominator from having a zero value, the quantization noise low threshold Cl is used as a small offset of the medium transmittance T, and finally T is used. 1-ω , the input image is enhanced to obtain a preliminary estimated normal exposure image with grayscale values limited to [0, 1] for the next structural decomposition. The mean of the patches is taken as the perceptual gain of the local patches, and then the perceptual gain of the whole image is aggregated and recorded as I. Because the patches decomposed by an image are captured under the same imaging conditions, it is assumed that the perceptual gain of the whole image is uniform, so the maximum perceptual gain of all patches is used to estimate the fused perceptual gain, and taking into account the characteristics of the input image, an adaptive correction is performed to obtain the perceptual gain. The expression of the perceptual gain is: Î , where max is the global maximum value of the image, mean is the global mean value of the image, and the value of p is 2 for all layers except the 0th layer, which is 1. Since a single image is captured under the same imaging conditions, the expected perceptual gain of the entire image should be uniform, so the perceptual gain set here is a global variable, and the range of perceptual gain i is limited to [0.55, 0.85]. Compared with using the inverted image of the low-light image as the overexposed image, the dual illumination estimation strategy is adopted to perform forward illumination estimation on the input image to correct the underexposed area, and at the same time perform reverse illumination estimation on the input image to correct the overexposed area. Although it can enhance the low-brightness area and Maintaining high-brightness areas, but this algorithm requires a huge amount of calculation. By enhancing the perceptual gain, the perceptual characteristics of the image are specifically adjusted to make the enhanced image more consistent with the visual perception of the human eye. The human eye is sensitive to perceptual characteristics such as brightness and contrast of images. By adjusting the perceptual gain separately, the visual needs of the human eye can be better met, the visual quality of the image can be improved, and the important information in the image can be highlighted, making the key areas in the image more eye-catching. In low-light environments, some important details in the image may be difficult to detect due to insufficient brightness. By enhancing the perceptual gain, the visibility of these details can be effectively improved, thereby increasing the information content and value of the image; like Figure 4 As shown, the signal intensity is related to the local contrast. The local plaque signal intensity is calculated and aggregated into C. A logarithmic linear stretch transformation is performed on it to enhance the signal intensity while preventing image inconsistency. k), where k is the maximum value in C rounded up, Take the larger value in the image, and the range should not be too large, otherwise the image will appear structurally uneven. The range of the independent variable is [0, 1], and k is 1, such as Figure 8As shown in the figure, although the exponential function has the largest value, its contrast is the lowest and there are artifacts at the edge of the eaves. Simply multiplying by k or using stretching functions with large ranges such as threesigma and gamma will cause local unevenness and severe edge effects in the image. Choosing the log function has richer layering and contrast than using the global maximum. The algorithm based on the Retinex model estimates the brightness component of the image, enhances the brightness component, and then multiplies it with the reflection component to obtain the enhanced image. Compared with finding R , the maximum grayscale value of each pixel in the G and B channels is used to construct the brightness component, and an optimization model is constructed to accurately estimate the brightness component. Although LIME is very good at enhancing the brightness component of the image, poor color constancy and over-enhancement sometimes occur in normal lighting areas, making the enhancement effect unnatural. In low-light images, insufficient brightness is a common problem. By enhancing the signal strength, the overall brightness of the image can be effectively improved, making the image clearer and more visible. It has an important protective effect on the edges, textures and other details of the image. In low-light images, due to the influence of noise and low brightness, the structural information of the image is often easily destroyed. By enhancing the signal structure alone, the structure of the image can be effectively restored and strengthened, the clarity and three-dimensional sense of the image can be improved, and it helps to improve the recognition and comprehensibility of the image.

[0027] Further, such as Figure 4 As shown, in k), in order to ensure that the faint details in the image can also be enhanced, k) Add a global bias and use it as the estimated structural strength. The structural strength is , further, in order to maximize the preservation of structural information and match the structural strength, while considering the need for comprehensive and To estimate the signal structure, the signal structure is obtained by ⊙ ⊙( Calculation, where p is an exponential parameter greater than or equal to 0, and its value is 4; when performing average brightness enhancement, such as Figure 5 As shown in the figure, for non-minimum scale images, the mean of each local patch is used as the brightness value of the patch; for minimum scale images, the mean of each local patch is calculated and then a nonlinear stretch is performed on it, and the exponential function is used to perform the baseline brightness compensation. The average brightness after aggregation is recorded as , so that the average intensity processing can balance the overall brightness distribution of the image, avoid the situation of local over-brightness or over-darkness, and ensure that the image will not have problems such as excessive distortion or color deviation during the enhancement process, thereby ensuring the stability and reliability of the entire enhancement process; After enhancing the four components of perceptual gain, signal strength, signal structure and average strength respectively, we start to reconstruct the image patches. ) describes the smallest scale patch reconstruction image, ⊙ ) ⊙B)⊙ ⊙B⊙ ) describes the aggregation of image patches at a certain scale, Based on the above, a local mean filter is performed on a specific module to improve the uniformity of the image. The filter size and the window size of the local patch are both p×p, p=2r+1. Furthermore, in the multi-scale framework, the image can be composed of two parts. The low-frequency information in the image is the brightness basis of the image, and the high-frequency information in the image is the detail feature of the image. The high-frequency component at the original scale also contains image color information and noise. Figure 9 The visualization information of high-frequency and low-frequency components at the largest scale and the second largest scale is shown, such as Figure 9 As shown, according to ） ） The high-frequency information of the second scale is obtained. This decomposition is carried out until the smallest scale layer is reached. The low-frequency information at different scales can be obtained by Calculation, the low-frequency information and high-frequency information are added together by the bilinear upsampling operator to obtain the final fused image, such as ,Depend on Figure 9It can be seen that most of the noise in the image is contained in the high-frequency information of the first layer. The corresponding high-frequency information is denoised by weighted guided filtering and BM3D, and the relevant parameters are set. The weighted guided filtering is used as the default method. In order to improve the visual effect, a weight is given to the high-frequency and low-frequency information to alleviate the problem of high image saturation. The specific setting is high-frequency weight 0.75, low-frequency weight 1.2. After the image is decomposed by the multi-scale framework of high-frequency information and low-frequency information, targeted enhancement is performed. The defogging algorithm is improved and the dependent dark channel prior information is discarded to enhance the brightness. It has better effects in terms of color richness and perceived authenticity, and can effectively capture the feature information of different scales in the image, in low light conditions. In an image, different objects and details may have different scales. Multi-scale analysis can better adapt to the characteristics of these different scales, thereby achieving more comprehensive and accurate image enhancement, and has a better enhancement effect on small targets and subtle structures in the image. In traditional single-scale methods, small targets and subtle structures are often easily ignored or blurred, while multi-scale analysis can enhance these small targets and subtle structures at different scales, improving their visibility and clarity. This method can better maintain the color balance of the image during the enhancement process, avoiding color distortion or deviation problems. By reasonably adjusting the enhancement degree of the four components, natural color restoration can be achieved, making the enhanced image more realistic and beautiful.

[0028] Working principle: When using this low-light image enhancement method based on multi-scale perceptual enhancement block decomposition, the SPD model is applied to single low-light image enhancement, which improves the color richness and perceptual realism of the image and increases the overall practicality.

[0029] While 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 these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A low-light image enhancement method based on multi-scale perceptual enhancement block decomposition, characterized in that: The following steps are involved: S1, first perform SPD decomposition on the low-light image to obtain the four components of perceptual gain, signal strength, signal structure and average strength; S2, enhance the four components of perception gain, signal strength, signal structure and average strength respectively through manually set enhancement strategies; S3, the enhanced four components are re-fused into one image through a multi-scale framework.

2. A low-light-level image enhancement method based on multi-scale perceptual enhancement block decomposition according to claim 1, characterized in that: The perception gain enhancement is set with a low threshold Cl, and the perception gain enhancement is based on the idea of pseudo-fog image, which is to invert the low-light image and then enhance it based on the atmospheric scattering model; The specific relationship of the atmospheric scattering model is: ⊙T+A⊙(1-T); in For the collected images, is the image to be restored, A is the atmospheric light intensity, T is the medium transmittance, and ⊙ represents the Hadamard product; The illumination intensity matrix A is taken as 1, and the medium transmittance T is T 1-ω Make estimates; Where ω is the fidelity coefficient, which is the global mean of the inverted image, and I is the inverted image of the input image X; The quantization noise low threshold Cl is used as a small bias of the medium transmittance T, using J= +A= +A , perform brightness enhancement on the input image; Get a preliminary estimate of the normal exposure image , the grayscale value is limited to [0, 1], which is used for the next step of structural decomposition; The mean value of the patch is used as the perceptual gain i of the local patch mean(x).

3. A low-light-level image enhancement method based on multi-scale perceptual enhancement block decomposition according to claim 2, characterized in that: The perceptual gain is recorded as I after being aggregated into the entire image. The maximum perceptual gain of all patches is used to estimate the fused perceptual gain, combined with the characteristics of the input image X; The expression of perception gain Î is: Î ; in To obtain the global maximum value of the image, To obtain the global mean of the image, the value of p is 2 for all layers except the 0th layer, which is 1.

4. A low-light-level image enhancement method based on multi-scale perceptual enhancement block decomposition according to claim 3, characterized in that: The perceptual gain is a global variable, and the range of the perceptual gain i is limited to [0.55, 0.85].

5. A low-light-level image enhancement method based on multi-scale perceptual enhancement block decomposition according to claim 1, characterized in that: The signal intensity is calculated for the local plaque, and the signal intensity of the local plaque is obtained by decomposition calculation. ║ - ║, and aggregated into C, based on C, a logarithmic linear stretching transformation is performed, the formula is k), where k is the maximum value in C rounded up, and Take the larger value in the image.

6. A low-light-level image enhancement method based on multi-scale perceptual enhancement block decomposition according to claim 5, characterized in that: The structural strength , signal structure Combine The formula is ⊙ ⊙( , where p is an exponential parameter greater than or equal to 0, and its value is 4.

7. A low-light-level image enhancement method based on multi-scale perceptual enhancement block decomposition according to claim 2, characterized in that: The average intensity is used for the image at non-minimum scale, and the average value of each local patch is used as the brightness value of the patch. The average intensity is used for the image at minimum scale, and the average value of each local patch is calculated and then subjected to nonlinear stretching, using an exponential function. , perform benchmark brightness compensation, and the average brightness after aggregation is recorded as , .

8. A low-light-level image enhancement method based on multi-scale perceptual enhancement block decomposition according to claim 7, characterized in that: The image formula for patch reconstruction of the image at the minimum scale is: ), the image patch aggregation formula at a certain scale is ⊙ ) ⊙B)⊙ ⊙B⊙ ), where B , ) indicates that local mean filtering is performed.

9. A low-light-level image enhancement method based on multi-scale perceptual enhancement block decomposition according to claim 7, characterized in that: The formula for reconstructing the image from the patch of the image at the minimum scale is It consists of low-frequency information and high-frequency information.

10. A low-light-level image enhancement method based on multi-scale perceptual enhancement block decomposition according to claim 9, characterized in that: The formula of the low-frequency information is expressed as L ; The formula of the low-frequency information is expressed as H ; Downsample by 2 times to obtain the second scale information ,based on , we can calculate 、 、 and ; According to the formula ） ） Get high-frequency information of the second scale ; Decomposition continues until the smallest scale level , Z , where h and w are the height and width of the original scale image respectively; The low-frequency information at different scales can be calculated as The low-frequency information and high-frequency information are finally fused through the bilinear upsampling operator. , Upsampling; The formula is .

Images