Image detail enhancement and noise suppression method in low-light environment

Abstract

The application discloses a kind of low-light environment under image detail enhancement and noise suppression method, belong to image processing technical field.The application first extracts low-light image luminance value to generate original luminance graph, obtains disturbance sensitive feature graph by adding random disturbance;Then, based on the local gradient of pixel, the local luminance information quantity is calculated, and the information effectiveness weight graph is generated by combining luminance randomness and structure disorder degree;Subsequently, the luminance detail graph is extracted from the original luminance graph, and the anti-disturbance, reliable detail graph is obtained by fusing the above feature graph;Finally, through image reconstruction neural network processing, the detail enhancement denoising luminance graph is output.The application realizes the accurate balance of low-light image detail retention and noise suppression, and improves image quality.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to a method for enhancing image details and suppressing noise in low-light environments. Background Technology

[0002] Low-light environments are prevalent in numerous practical applications such as security monitoring, nighttime driving assistance, and outdoor remote sensing. Due to insufficient lighting, acquired images often exhibit problems such as low brightness, blurred and obscured details, and severe noise interference. This not only reduces the visual quality of the images but also adversely affects the accuracy and reliability of subsequent advanced visual tasks such as image recognition, object detection, and feature extraction. Therefore, detail enhancement and noise suppression techniques for low-light images have become a research hotspot in the field of computer vision. The core objective is to effectively recover hidden details and suppress noise caused by insufficient lighting while improving image brightness, thus providing high-quality image data support for subsequent visual tasks.

[0003] Among existing low-light image enhancement techniques, Retinex-based methods are widely used. The core idea of ​​these methods is to decompose the image into a luminance component reflecting changes in illumination and a reflection component characterizing the essential features of the image. Image enhancement is achieved by adjusting the luminance component, while simultaneously aiming to preserve the detail information in the reflection component. However, existing Retinex-based methods have significant drawbacks in practical applications. Specifically, when enhancing the luminance component to improve the overall image brightness, it is difficult to accurately distinguish between real details and noise in the image, easily leading to the simultaneous amplification of noise. This noise amplification is particularly pronounced in low-light areas, resulting in the enhanced image still exhibiting both blurred details and noise interference, demonstrating low precision in both image detail enhancement and noise suppression. Summary of the Invention

[0004] To address the aforementioned shortcomings in the prior art, this invention provides an image detail enhancement and noise suppression method for low-light environments, which solves the problem of low accuracy in image detail enhancement and noise suppression in the prior art.

[0005] To achieve the aforementioned objectives, the present invention employs the following technical solution: a method for enhancing image details and suppressing noise under low-light conditions, comprising the following steps:

[0006] S1. Extract brightness values ​​from the low-light image to obtain the original brightness map;

[0007] S2. Add random perturbation to the original brightness map, and obtain the perturbation-sensitive feature map based on the brightness difference before and after the perturbation.

[0008] S3. Based on the local gradient of each pixel, obtain the local brightness information. Based on the brightness randomness and structural disorder, enhance the local brightness information to obtain the information effectiveness weight map.

[0009] S4. Obtain the brightness detail map from the original brightness map, and enhance the details by combining the perturbation-sensitive feature map and the information validity weight map respectively, to obtain the perturbation-resistant detail map and the reliable detail map.

[0010] S5. An image reconstruction neural network is used to process the anti-disturbance detail map, the reliable detail map, and the original brightness map to obtain a detail-enhanced and denoised brightness map.

[0011] Furthermore, S2 includes the following sub-steps:

[0012] S21. Multiply the disturbance amplitude by a standard normal distribution random number to obtain the random disturbance;

[0013] S22. Add random perturbation to the brightness values ​​in the original brightness map to obtain a perturbed brightness map;

[0014] S23. Subtract the luminance variance of each pixel in the perturbation luminance map from the luminance variance of each pixel in the original luminance map, and take the absolute value of the difference to obtain the perturbation response map:

[0015] S24. Normalize the disturbance response map to obtain the disturbance sensitivity feature map.

[0016] Furthermore, S3 includes the following sub-steps:

[0017] S31. Obtain the gradient vector for each pixel in the original brightness map, and multiply the magnitude of the gradient vector by the corresponding brightness value in the original brightness map to obtain the local brightness information.

[0018] S32. Taking each pixel in the original brightness map as the center, calculate the brightness variance of the neighborhood at the center and normalize it to obtain the brightness randomness.

[0019] S33. Taking each pixel in the original brightness map as the center, the structural disorder is obtained based on the gradient vector of all pixels in the central neighborhood.

[0020] S34. Take the average of the luminance randomness and structural disorder to obtain the luminance uncertainty;

[0021] S35. Subtract the brightness uncertainty from 1, multiply the result of the subtraction with the local brightness information, and then normalize the result of the multiplication to obtain the information effectiveness weight map.

[0022] Furthermore, S33 includes the following sub-steps:

[0023] S331. Taking each pixel in the original brightness map as the center, add the gradient vectors of all pixels in the neighborhood of the center, and take the modulus of the addition result to obtain the neighborhood gradient vector and the magnitude.

[0024] S332. Take the modulus of the gradient vector of each pixel within the central neighborhood, and add all the modulus results within the central neighborhood to obtain the sum of the neighborhood gradient magnitudes;

[0025] S333. The structural disorder is obtained by subtracting the ratio of the neighborhood gradient vector and magnitude to the sum of the neighborhood gradient magnitudes from 1.

[0026] Furthermore, S4 includes the following sub-steps:

[0027] S41. Perform Gaussian smoothing on the original brightness map to obtain a Gaussian smoothed brightness map;

[0028] S42. Subtract the brightness values ​​of the original brightness map from the Gaussian smoothed brightness map at the pixel level, and take the absolute value to obtain the brightness detail map.

[0029] S43. The luminance detail map is enhanced by using the perturbation-sensitive feature map and the information validity weight map respectively, so as to obtain the perturbation-resistant detail map and the reliable detail map.

[0030] Furthermore, the expression for the disturbance-resistant detail diagram is obtained as follows:

[0031] ,

[0032] in, To prevent disturbances to pixels in the detail image The brightness value at that location, For pixels in the perturbation-sensitive feature map Sensitive values ​​at that location For pixels in the brightness detail image The brightness value at that location;

[0033] The expression for obtaining the reliable detail map is:

[0034] ,

[0035] in, For pixels in a reliable detail image The brightness value at that location, For pixels in the information validity weighting graph Information validity weighting.

[0036] Furthermore, the image reconstruction neural network in S5 includes: a perturbation-resistant detail coding branch, a reliable detail coding branch, a luminance structure weight modulation branch, a luminance structure guided feature fusion module, a reconstruction decoding network, and a luminance residual stacking layer.

[0037] Furthermore, S5 includes the following sub-steps:

[0038] S51. The disturbance-resistant detail coding branch is used to extract the disturbance-resistant detail code from the disturbance-resistant detail map;

[0039] S52. Trusted detail coding branch is used to extract trusted detail codes from the trusted detail graph;

[0040] S53. Use the brightness structure weight modulation branch to extract the first brightness structure weight and the second brightness structure weight from the original brightness map;

[0041] S54. Based on the first luminance structure weight and the second luminance structure weight, the luminance structure guided feature fusion module is used to weight the anti-disturbance detail coding and the reliable detail coding to obtain the luminance structure guided fusion feature.

[0042] S55. A reconstruction decoding network is used to reconstruct the brightness structure-guided fusion features to obtain a brightness detail residual map.

[0043] S56. The original brightness map and the brightness detail residual map are added together using a brightness residual overlay layer to obtain a detail-enhanced and noise-reduced brightness map.

[0044] Furthermore, the perturbation-resistant detail coding branch and the trusted detail coding branch have the same structure, both including the following sequentially connected components: a first convolutional block, a second convolutional block, and a first convolutional layer; the first convolutional block has a kernel size of 3×3 and 32 channels; the second convolutional block has a kernel size of 3×3 and 64 channels; and the first convolutional layer has a kernel size of 1×1 and 32 channels.

[0045] The luminance structure weight modulation branch consists of the following sequentially connected components: a third convolutional block, a fourth convolutional block, a second convolutional layer, and a sigmoid layer. The third convolutional block has a kernel size of 5×5 and 32 channels. The fourth convolutional block has a kernel size of 3×3 and 32 channels. The second convolutional layer has a kernel size of 1×1 and 2 channels. The sigmoid layer is used to map the features output by the first channel of the second convolutional layer to the first luminance structure weight, and the sigmoid layer is used to map the features output by the second channel of the second convolutional layer to the second luminance structure weight.

[0046] Furthermore, the reconstruction decoding network comprises the following sequentially connected components: a fifth convolutional block, a first residual block, a second residual block, a sixth convolutional block, and a third convolutional layer; the kernel size of the fifth convolutional block is 3×3, and the number of channels is 64; the kernel size of the convolutional layers in the first and second residual blocks is 3×3, and the number of channels is 64; the kernel size of the third convolutional layer is 1×1, and the number of channels is 1.

[0047] The beneficial effects of this invention are as follows:

[0048] 1. This invention introduces a random perturbation mechanism in the luminance domain (step S2). By analyzing the sensitivity of luminance changes before and after the perturbation, a perturbation-sensitive feature map is constructed. This feature map reflects the response characteristics of luminance changes to external perturbations, allowing regions that are highly sensitive to perturbations and have poor stability to be identified as noise-dominated regions, while regions that are insensitive to perturbations correspond to true structural details. Therefore, in the subsequent enhancement process, effective distinction is achieved between noise and true details, overcoming the problem of noise being amplified synchronously during luminance enhancement in existing Retinex-type methods.

[0049] 2. In step S3 of this invention, based on the local gradient information of pixels, brightness randomness and structural disorder are introduced to weight and enhance the local brightness information, constructing an information validity weight map. This weight map can measure the credibility of brightness details in different regions from the perspective of information theory and structural stability, so that regions with clear structure and high information content receive higher weights during the enhancement process, while regions with strong randomness and structural disorder are suppressed, effectively reducing noise interference and improving the accuracy and stability of the enhancement results.

[0050] 3. This invention constructs both an anti-disturbance detail map and a reliable detail map (step S4), constraining brightness details from two complementary dimensions: perturbation stability and information validity. This dual detail constraint mechanism can enhance realistic details while suppressing unreliable details, effectively avoiding the conflict between detail enhancement and noise suppression in traditional methods, and achieving synergistic optimization of the two.

[0051] 4. This invention inputs the anti-disturbance detail map, the reliable detail map and the original brightness map into the image reconstruction neural network for reconstruction. The neural network adaptively fuses the multi-source brightness and detail information, thereby improving the accuracy of image detail enhancement and noise suppression. Attached Figure Description

[0052] Figure 1 A flowchart of an image detail enhancement and noise suppression method under low-light conditions;

[0053] Figure 2 A schematic diagram of the structure of a neural network for image reconstruction;

[0054] Figure 3A schematic diagram of the structure of the perturbation-resistant detail coding branch and the trusted detail coding branch;

[0055] Figure 4 This is a schematic diagram of the brightness structure weighted modulation branch;

[0056] Figure 5 A schematic diagram of the reconstructed decoding network;

[0057] Figure 6 This is a schematic diagram of the structure of the first residual block and the second residual block. Detailed Implementation

[0058] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0059] like Figure 1 As shown, a method for enhancing image details and suppressing noise in low-light environments includes the following steps:

[0060] S1. Extract brightness values ​​from the low-light image to obtain the original brightness map;

[0061] S2. Add random perturbation to the original brightness map, and obtain the perturbation-sensitive feature map based on the brightness difference before and after the perturbation.

[0062] S3. Based on the local gradient of each pixel, obtain the local brightness information. Based on the brightness randomness and structural disorder, enhance the local brightness information to obtain the information effectiveness weight map.

[0063] S4. Obtain the brightness detail map from the original brightness map, and enhance the details by combining the perturbation-sensitive feature map and the information validity weight map respectively, to obtain the perturbation-resistant detail map and the reliable detail map.

[0064] S5. An image reconstruction neural network is used to process the anti-disturbance detail map, the reliable detail map, and the original brightness map to obtain a detail-enhanced and denoised brightness map.

[0065] S2 includes the following steps:

[0066] S21. Multiply the disturbance amplitude by a standard normal distribution random number to obtain the random disturbance;

[0067] S22. Add random perturbation to the brightness values ​​in the original brightness map to obtain a perturbed brightness map;

[0068] S23. Subtract the luminance variance of each pixel in the perturbation luminance map from the luminance variance of each pixel in the original luminance map, and take the absolute value of the difference to obtain the perturbation response map:

[0069] ,

[0070] ,

[0071] in, For pixels in the disturbance response graph The response value at that location, For pixels in the original brightness image The luminance variance at that location To perturb the pixels in the brightness map The luminance variance at that location For pixels in the original brightness image The brightness value at that location, To perturb the pixels in the brightness map The brightness value at that location, For the disturbance amplitude, To take the absolute value, These are standard normally distributed random numbers. The random perturbation is generated for each pixel, with one random perturbation corresponding to one pixel.

[0072] S24. Normalize the disturbance response map to obtain the disturbance sensitivity feature map:

[0073] ,

[0074] in, For pixels in the perturbation-sensitive feature map Sensitive values ​​at that location To obtain the minimum value, To obtain the maximum value.

[0075] In this embodiment, the disturbance amplitude The value range is 5 to 10.

[0076] In this embodiment, S2 specifically includes:

[0077] S21. Set the perturbation round K. During each perturbation, multiply the perturbation amplitude by a standard normal distribution random number to obtain the kth random perturbation for each pixel.

[0078] S22. Add the kth random perturbation to the brightness value of the pixel in the original brightness map to obtain the kth perturbation brightness map.

[0079] S23. Subtract the brightness variance of each pixel in the k-th perturbation brightness map from the brightness variance of each pixel in the original brightness map, and take the absolute value of the difference to obtain the k-th perturbation response map.

[0080] S24. Add the response values ​​of the K perturbation response maps by pixel, divide by K, and then normalize to obtain the perturbation sensitive feature map. After multiple perturbations, take the average value to improve accuracy.

[0081] In this embodiment, K is set to 10, and k is the number of the disturbance round.

[0082] Because directly using the variance of the original brightness map cannot distinguish between the variance of "true details" and the variance of "noise," adding perturbations further increases the variance in anomalous areas.

[0083] The luminance variance of each pixel is calculated as follows: taking each pixel as the center, the variance of each luminance value within the central neighborhood is used as the basis.

[0084] This invention can effectively characterize the stability differences of different pixel regions to random disturbances through the disturbance response. Regions with larger disturbance responses correspond to regions with strong brightness randomness and higher noise ratios, while regions with smaller disturbance responses reflect real and stable image structure and detail information.

[0085] In this embodiment, S3 includes the following sub-steps:

[0086] S31. Obtain the gradient vector for each pixel in the original brightness image, and multiply the magnitude of the gradient vector by the corresponding brightness value in the original brightness image to obtain the local brightness information.

[0087] ,

[0088] ,

[0089] in, For pixels The amount of local brightness information at that location For pixels The gradient vector at that point, For pixels in the original brightness image The brightness value at that location, The sign of the partial derivative. For modulo operation, and All coordinates are pixel coordinates;

[0090] S32. Taking each pixel in the original brightness map as the center, calculate the brightness variance of the neighborhood around the center, and normalize it to obtain the brightness randomness:

[0091] ,

[0092] ,

[0093] in, For pixels The randomness of the brightness at that location For pixels The luminance variance at that location The average brightness value within the neighborhood of the center. For the pixels in the neighborhood of the center The brightness value at that location, The number of pixels in the neighborhood of the center. The area surrounding the center. and The coordinates of the pixels within the neighborhood of the center are the variables. To obtain the minimum value, To obtain the maximum value;

[0094] In S32, a value is obtained centered on each pixel. Therefore, one pixel corresponds to one ;

[0095] S33. Taking each pixel in the original brightness map as the center, and based on the gradient vector of all pixels within the central neighborhood, obtain the structural disorder:

[0096] ,

[0097] in, For pixels The degree of structural disorder at the location For pixels within the central neighborhood The gradient vector at that point, For modulo operation, To avoid a constant in the denominator of 0, The area surrounding the center. and The coordinates of the pixels within the neighborhood of the center are the variables. The neighborhood of the center Ω is a 3×3 pixel neighborhood or a 5×5 pixel neighborhood.

[0098] S34. Take the average of the luminance randomness and structural disorder to obtain the luminance uncertainty. ;

[0099] S35. Subtract the luminance uncertainty from 1, multiply the result by the local luminance information, and then normalize the result to obtain the information effectiveness weight map:

[0100] ,

[0101] ,

[0102] in, For pixels in the information validity weighting graph Information validity weighting at the location, For pixels The amount of information in the local brightness enhancement at that location For pixels The amount of local brightness information at that location For pixels The brightness uncertainty at that location.

[0103] This invention obtains local brightness information by multiplying the gradient vector magnitude by the brightness value, which preserves the detailed texture features reflected by the gradient and highlights the information of the effective brightness area in combination with the brightness value. Next, the brightness randomness obtained by the brightness variance can identify noise areas with irregular brightness fluctuations. The structural disorder obtained by the analysis of the gradient vector can distinguish between noise areas with messy textures and real detail areas with regular structures. The brightness uncertainty calculated by combining the two can accurately quantify the unreliability of the area. Finally, multiplying the uncertainty by 1 by the local brightness information is equivalent to enhancing the information of low uncertainty (real detail) areas and suppressing the information of high uncertainty (noise) areas. The resulting information effectiveness weight map is then normalized.

[0104] In this embodiment, S33 includes the following sub-steps:

[0105] S331. Taking each pixel in the original brightness map as the center, add the gradient vectors of all pixels in the neighborhood of the center, and take the modulus of the addition result to obtain the neighborhood gradient vector and the magnitude.

[0106] S332. Take the modulus of the gradient vector of each pixel within the central neighborhood, and add all the modulus results within the central neighborhood to obtain the sum of the neighborhood gradient magnitudes;

[0107] S333. The structural disorder is obtained by subtracting the ratio of the neighborhood gradient vector and magnitude to the sum of the neighborhood gradient magnitudes from 1.

[0108] When the gradient directions of pixels in the neighborhood are consistent, the vector and magnitude are close to the sum of the magnitudes, and the structural disorder is close to 0, indicating that the region is a well-structured and realistic detail. When the gradient directions of pixels in the neighborhood are disordered, the vector and magnitude are much smaller than the sum of the magnitudes, and the structural disorder is close to 1, indicating that the region is a disordered and noisy region.

[0109] In this embodiment, S4 includes the following sub-steps:

[0110] S41. Perform Gaussian smoothing on the original brightness map to obtain a Gaussian smoothed brightness map. The Gaussian kernel size used in the Gaussian smoothing process is 3×3, and the standard deviation σ ranges from 0.5 to 1.5.

[0111] S42. Subtract the brightness values ​​of the original brightness map from the Gaussian smoothed brightness map at the pixel level, and take the absolute value to obtain the brightness detail map.

[0112] S43. The luminance detail map is enhanced by using the perturbation-sensitive feature map and the information validity weight map respectively, so as to obtain the perturbation-resistant detail map and the reliable detail map.

[0113] In this embodiment, the expression for obtaining the disturbance-resistant detail map is:

[0114] ,

[0115] in, To prevent disturbances to pixels in the detail image The brightness value at that location, For pixels in the perturbation-sensitive feature map Sensitive values ​​at that location For pixels in the brightness detail image The brightness value at that location;

[0116] The expression for obtaining the reliable detail map is:

[0117] ,

[0118] in, For pixels in a reliable detail image The brightness value at that location, For pixels in the information validity weighting graph Information validity weighting.

[0119] ,

[0120] in, For pixels in the brightness detail image The brightness value at that location, For pixels in the original brightness image The brightness value at that location, For pixels in a Gaussian-smoothed brightness map The brightness value at that location, To take the absolute value.

[0121] This invention effectively removes low-frequency illumination components from an image by subtracting the original brightness map from a Gaussian-smoothed brightness map and taking the absolute value. This allows for the precise extraction of brightness detail maps that reflect texture, edges, and other information, avoiding interference from uneven illumination in detail recognition. Furthermore, a perturbation-sensitive feature map is introduced to suppress and enhance brightness details, weakening detail components that are highly sensitive to random perturbations and have poor stability, thus forming a perturbation-resistant detail map with strong noise resistance. Simultaneously, an information validity weight map is used to enhance brightness details, adaptively amplifying structurally clear and reliable detail components to form a reliable detail map.

[0122] like Figure 2 As shown, the image reconstruction neural network in S5 includes: a perturbation-resistant detail coding branch, a reliable detail coding branch, a luminance structure weight modulation branch, a luminance structure guided feature fusion module, a reconstruction decoding network, and a luminance residual stacking layer.

[0123] In this embodiment, S5 includes the following sub-steps:

[0124] S51. The disturbance-resistant detail coding branch is used to extract the disturbance-resistant detail code from the disturbance-resistant detail map;

[0125] S52. Trusted detail coding branch is used to extract trusted detail codes from the trusted detail graph;

[0126] S53. Use the brightness structure weight modulation branch to extract the first brightness structure weight and the second brightness structure weight from the original brightness map;

[0127] S54. Based on the first luminance structure weight and the second luminance structure weight, the luminance structure guided feature fusion module is used to weight the anti-disturbance detail coding and the reliable detail coding to obtain the luminance structure guided fusion feature.

[0128] S55. A reconstruction decoding network is used to reconstruct the brightness structure-guided fusion features to obtain a brightness detail residual map.

[0129] S56. A brightness residual overlay layer is used to add the brightness values ​​of the original brightness map and the brightness detail residual map at the pixel level to obtain a detail-enhanced and noise-reduced brightness map.

[0130] In this invention, the anti-disturbance detail coding branch, the reliable detail coding branch, and the luminance structure weight modulation branch extract corresponding features for "anti-disturbance details," "reliable details," and "original luminance structure," respectively. The first and second luminance structure weights are generated by combining image information from the original luminance map. These weighted fusions of the anti-disturbance detail coding and reliable detail coding ensure that different types of detail features are guided and constrained by the original luminance structure during the fusion process, further improving the accuracy of detail reconstruction.

[0131] The reconstruction decoding network generates a brightness detail residual map based on fusion features, and then adds the brightness residual map to the original brightness map to obtain the final result. This not only enhances the real details through the residual map, but also preserves the basic structure and overall style of the image based on the original brightness map, avoiding problems such as image distortion and color shift, while further suppressing noise residue.

[0132] like Figure 3 As shown, the perturbation-resistant detail coding branch and the trusted detail coding branch have the same structure, both including the following sequentially connected components: a first convolutional block, a second convolutional block, and a first convolutional layer; the kernel size of the first convolutional block is 3×3, and the number of channels is 32; the kernel size of the second convolutional block is 3×3, and the number of channels is 64; the kernel size of the first convolutional layer is 1×1, and the number of channels is 32.

[0133] like Figure 4 As shown, the luminance structure weight modulation branch includes the following sequentially connected components: a third convolutional block, a fourth convolutional block, a second convolutional layer, and a sigmoid layer. The third convolutional block has a kernel size of 5×5 and 32 channels. The fourth convolutional block has a kernel size of 3×3 and 32 channels. The second convolutional layer has a kernel size of 1×1 and 2 channels. The sigmoid layer is used to map the features output by the first channel of the second convolutional layer to the first luminance structure weight, and the sigmoid layer is used to map the features output by the second channel of the second convolutional layer to the second luminance structure weight.

[0134] In the brightness structure weight modulation branch, the second convolutional layer outputs feature maps for two channels. By applying the Sigmoid activation function to the output of each channel, the feature values ​​are mapped to the [0,1] interval, thereby obtaining the brightness structure weights—the Sigmoid output of the first channel corresponds to the first brightness structure weight, and the Sigmoid output of the second channel corresponds to the second brightness structure weight.

[0135] In this embodiment, the expression for the luminance structure weighted modulation branch is:

[0136] ,

[0137] in, Luminance structure guides fusion features. As the first brightness structure weight, As the weight of the second brightness structure, For disturbance-resistant detail coding, Encode trusted details.

[0138] like Figure 5As shown, the reconstruction decoding network includes the following components connected in sequence: a fifth convolutional block, a first residual block, a second residual block, a sixth convolutional block, and a third convolutional layer; the kernel size of the fifth convolutional block is 3×3, and the number of channels is 64; the kernel size of the convolutional layers in the first and second residual blocks is 3×3, and the number of channels is 64; the kernel size of the third convolutional layer is 1×1, and the number of channels is 1.

[0139] like Figure 6 As shown, the first residual block and the second residual block have the same structure, both including: a fourth convolutional layer, a fifth convolutional layer and an adder A1.

[0140] In this embodiment, all convolutional blocks include: a convolutional layer and a ReLU activation function.

[0141] This invention combines random perturbation analysis with information validity modeling in the luminance domain to finely determine the stability and reliability of luminance details before detail enhancement. On one hand, it utilizes luminance changes before and after perturbation to construct perturbation-sensitive features, effectively distinguishing noise-sensitive regions from structurally stable real details. On the other hand, it constructs information validity weights based on local gradients, luminance randomness, and structural disorder to suppress the enhancement intensity in regions dominated by random noise. Furthermore, it uses a neural network to jointly reconstruct the original luminance and multi-dimensional detail features, achieving targeted enhancement of real details and adaptive noise suppression, thus solving the problem of low accuracy in detail enhancement and noise suppression in existing low-light image enhancement methods.

[0142] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. 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 image details and suppressing noise in low-light environments, characterized in that, Includes the following steps: S1. Extract brightness values ​​from the low-light image to obtain the original brightness map; S2. Add random perturbation to the original brightness map, and obtain the perturbation-sensitive feature map based on the brightness difference before and after the perturbation. S3. Based on the local gradient of each pixel, obtain the local brightness information. Based on the brightness randomness and structural disorder, enhance the local brightness information to obtain the information effectiveness weight map. S4. Obtain the brightness detail map from the original brightness map, and enhance the details by combining the perturbation-sensitive feature map and the information validity weight map respectively, to obtain the perturbation-resistant detail map and the reliable detail map. S5. An image reconstruction neural network is used to process the anti-perturbation detail map, the reliable detail map, and the original brightness map to obtain a detail-enhanced and denoised brightness map. S2 includes the following steps: S21. Multiply the disturbance amplitude by a standard normal distribution random number to obtain the random disturbance; S22. Add random perturbation to the brightness values ​​in the original brightness map to obtain a perturbed brightness map; S23. Subtract the luminance variance of each pixel in the perturbation luminance map from the luminance variance of each pixel in the original luminance map, and take the absolute value of the difference to obtain the perturbation response map: S24. Normalize the disturbance response map to obtain the disturbance sensitivity feature map; S3 includes the following steps: S31. Obtain the gradient vector for each pixel in the original brightness map, and multiply the magnitude of the gradient vector by the corresponding brightness value in the original brightness map to obtain the local brightness information. S32. Taking each pixel in the original brightness map as the center, calculate the brightness variance of the neighborhood at the center and normalize it to obtain the brightness randomness. S33. Taking each pixel in the original brightness map as the center, the structural disorder is obtained based on the gradient vector of all pixels in the central neighborhood. S34. Take the average of the luminance randomness and structural disorder to obtain the luminance uncertainty; S35. Subtract the brightness uncertainty from 1, multiply the subtraction result by the local brightness information, and then normalize the multiplication result to obtain the information effectiveness weight map. S33 includes the following steps: S331. Taking each pixel in the original brightness map as the center, add the gradient vectors of all pixels in the neighborhood of the center, and take the modulus of the addition result to obtain the neighborhood gradient vector and the magnitude. S332. Take the modulus of the gradient vector of each pixel within the central neighborhood, and add all the modulus results within the central neighborhood to obtain the sum of the neighborhood gradient magnitudes; S333. The structural disorder is obtained by subtracting the ratio of the neighborhood gradient vector and magnitude to the sum of the neighborhood gradient magnitudes from 1.

2. The image detail enhancement and noise suppression method under low-light conditions according to claim 1, characterized in that, S4 includes the following steps: S41. Perform Gaussian smoothing on the original brightness map to obtain a Gaussian smoothed brightness map; S42. Subtract the brightness values ​​of the original brightness map from the Gaussian smoothed brightness map at the pixel level, and take the absolute value to obtain the brightness detail map. S43. The luminance detail map is enhanced by using the perturbation-sensitive feature map and the information validity weight map respectively, so as to obtain the perturbation-resistant detail map and the reliable detail map.

3. The image detail enhancement and noise suppression method under low-light conditions according to claim 1 or 2, characterized in that, The expression for obtaining the disturbance-resistant detail map is: ， in, To prevent disturbances to pixels in the detail image The brightness value at that location, For pixels in the perturbation-sensitive feature map Sensitive values ​​at that location For pixels in the brightness detail image The brightness value at that location; The expression for obtaining the reliable detail map is: ， in, For pixels in a reliable detail image The brightness value at that location, For pixels in the information validity weighting graph Information validity weighting.

4. The image detail enhancement and noise suppression method under low-light conditions according to claim 1, characterized in that, The image reconstruction neural network in S5 includes: a perturbation-resistant detail coding branch, a reliable detail coding branch, a luminance structure weight modulation branch, a luminance structure guided feature fusion module, a reconstruction decoding network, and a luminance residual overlay layer.

5. The image detail enhancement and noise suppression method under low light conditions according to claim 4, characterized in that, S5 includes the following steps: S51. The disturbance-resistant detail coding branch is used to extract the disturbance-resistant detail code from the disturbance-resistant detail map; S52. Trusted detail coding branch is used to extract trusted detail codes from the trusted detail graph; S53. Use the brightness structure weight modulation branch to extract the first brightness structure weight and the second brightness structure weight from the original brightness map; S54. Based on the first luminance structure weight and the second luminance structure weight, the luminance structure guided feature fusion module is used to weight the anti-disturbance detail coding and the reliable detail coding to obtain the luminance structure guided fusion feature. S55. A reconstruction decoding network is used to reconstruct the brightness structure-guided fusion features to obtain a brightness detail residual map. S56. The original brightness map and the brightness detail residual map are added together using a brightness residual overlay layer to obtain a detail-enhanced and noise-reduced brightness map.

6. The image detail enhancement and noise suppression method under low light conditions according to claim 4, characterized in that, The perturbation-resistant detail coding branch and the trusted detail coding branch have the same structure, both including a first convolutional block, a second convolutional block, and a first convolutional layer connected in sequence; the first convolutional block has a kernel size of 3×3 and 32 channels; the second convolutional block has a kernel size of 3×3 and 64 channels; the first convolutional layer has a kernel size of 1×1 and 32 channels. The luminance structure weight modulation branch consists of the following sequentially connected components: a third convolutional block, a fourth convolutional block, a second convolutional layer, and a sigmoid layer. The third convolutional block has a kernel size of 5×5 and 32 channels. The fourth convolutional block has a kernel size of 3×3 and 32 channels. The second convolutional layer has a kernel size of 1×1 and 2 channels. The sigmoid layer is used to map the features output by the first channel of the second convolutional layer to the first luminance structure weight, and the sigmoid layer is used to map the features output by the second channel of the second convolutional layer to the second luminance structure weight.

7. The image detail enhancement and noise suppression method under low light conditions according to claim 4, characterized in that, The reconstruction decoding network comprises, in sequence: a fifth convolutional block, a first residual block, a second residual block, a sixth convolutional block, and a third convolutional layer; the kernel size of the fifth convolutional block is 3×3, and the number of channels is 64; the kernel size of the convolutional layers in the first and second residual blocks is 3×3, and the number of channels is 64; the kernel size of the third convolutional layer is 1×1, and the number of channels is 1.

Images