An image enhancement method with adaptive brightness gain control under low illumination

Abstract

This invention discloses an image enhancement method with adaptive brightness gain control under low illumination conditions, belonging to the field of image data processing technology. The method includes: S1: acquiring raw image data and auxiliary processing data under low illumination conditions, preprocessing them, and storing the data to construct an image processing database; S2: performing illumination distribution optimization and brightness correction operations through brightness deviation analysis and outputting an illumination enhancement component; S3: performing denoising and detail enhancement processing based on noise and color deviation results and outputting a corrected reflectance component; S4: using image quality enhancement analysis and a loss function to guide a dual U-Net architecture for detail reconstruction and color optimization; S5: adjusting processing parameters and driving algorithm closed-loop optimization based on quality feedback and difficult samples. This method solves the multiple degradation problems of insufficient image brightness, noise interference, color distortion, and detail loss under low illumination conditions.

Description

Technical Field

[0001] This invention relates to the field of image data processing technology, specifically to an image enhancement method with adaptive brightness gain control under low illumination. Background Technology

[0002] In the field of computer vision, low-light image enhancement is one of the core challenges in image processing. In low-light environments, images captured by cameras not only lack brightness but also often suffer from various degradation phenomena such as noise interference, color shift, loss of detail, and blurred edges. To address these issues, existing technologies typically employ independent image enhancement methods, such as using brightness enhancement algorithms to improve insufficient brightness or utilizing denoising algorithms to reduce image noise.

[0003] For example, the invention patent publication number CN109801240B discloses an image enhancement method and an image enhancement device. The image enhancement method includes: acquiring an original grayscale image; dividing the original grayscale image into bright and dark areas based on the total number of pixels and the number of pixels corresponding to each grayscale value; calculating the grayscale mean of the dark areas; determining the magnification factor of each pixel based on the grayscale mean and a preset grayscale threshold; and performing brightness enhancement processing on the original grayscale image based on the magnification factor of each pixel to obtain a target grayscale image. This invention determines the magnification factor of each pixel based on the grayscale distribution of the original grayscale image to enhance the image brightness. It employs a point processing algorithm, resulting in high processing efficiency, low resource consumption, and good real-time performance.

[0004] For example, the invention patent with publication number CN119477710B discloses a low-light video image enhancement method based on automatic brightness perception. It constructs a paired dataset of low-light images and corresponding reference images; then, the images are converted to the HSV color space and the V channel is enhanced; the CLAHE algorithm is used to enhance the low-light images in the dataset, and the enhancement results are obtained by iterating through cropping thresholds; the optimal cropping threshold is determined based on the evaluation results of the low-light enhanced images; the optimal cropping threshold is clustered using the K-means algorithm and the best value is selected; a convolutional neural network is used to perceive the brightness of the image, and the cluster centers of the K-means algorithm are used as the optimal cropping thresholds to be assigned to different brightness levels; finally, the CLAHE algorithm with adaptive cropping thresholds is used for processing. This method integrates deep learning methods and the traditional image enhancement method CLAHE algorithm, ensuring both efficiency and enhancement effect.

[0005] However, these methods generally suffer from drawbacks, such as handling only a single aspect and struggling to simultaneously address degradation issues related to brightness, color, and noise. Furthermore, existing methods often struggle with complex low-light environments and cannot provide personalized optimization for different image types. The inadequate quality assessment and feedback mechanisms for image enhancement lead to instability in enhancement effects during practical applications, making it difficult to meet real-world needs.

[0006] Therefore, in order to address the above problems, there is an urgent need for an image enhancement method that adaptively adjusts the brightness gain under low illumination. Summary of the Invention

[0007] Technical problems to be solved

[0008] To address the shortcomings of existing technologies, this invention provides an image enhancement method with adaptive brightness gain control under low illumination, which solves the problems of insufficient image brightness, noise interference, color distortion, and multiple degradations such as loss of detail in low-light environments.

[0009] Technical solution

[0010] To achieve the above objectives, the present invention provides the following technical solution: an image enhancement method with adaptive brightness gain control under low illumination, comprising: S1: acquiring raw image data and auxiliary processing data under low illumination conditions, preprocessing them, and storing the raw image data and auxiliary processing data to construct an image processing database; S2: performing brightness deviation analysis using illumination component data, performing differentiated illumination distribution optimization and brightness correction operations, and outputting the illumination enhancement component; S3: fusing brightness control, noise suppression, and color correction, quantifying image noise and color deviation, performing denoising and detail enhancement processing based on the noise and color deviation results, and outputting the corrected reflectivity component; S4: performing quality enhancement analysis based on the image's noise level, detail richness, and color fidelity, and using a loss function to guide a dual U-Net architecture for detail reconstruction and color optimization of the image; S5: adjusting processing parameters and driving algorithm closed-loop optimization based on quality feedback and difficult samples.

[0011] Further, the specific steps for preprocessing raw image data and auxiliary processing data acquired under low-light conditions are as follows: Image data is acquired under low-light conditions, including raw image data and auxiliary processing data; the raw image data includes raw pixel brightness values ​​and the total number of pixels; the auxiliary processing data includes illumination components, reflectivity components, average illumination components, illumination standard deviation, average signal intensity, noise standard deviation, reference reflectivity components, and color standard deviation; raw images under low-light conditions are acquired in real-time using a camera, directly obtaining the raw pixel brightness value of each pixel; the total number of pixels in the image is obtained by calculating the product of width and height in the raw image resolution; the global pixel brightness value is calculated by summing all raw pixel brightness values ​​and dividing by the total number of pixels. The average signal strength is calculated as follows: The original image is processed using a fast Retinex decomposition algorithm based on guided filtering to separate and output the illumination component and reflectance component; the arithmetic mean of all pixel values ​​of the illumination component is calculated to obtain the average illumination component; the standard deviation of all pixel values ​​of the illumination component is calculated as the illumination standard deviation; a standard reference image under normal lighting conditions is acquired, and the corresponding reference reflectance component is obtained by applying the same Retinex decomposition algorithm; a gray-level co-occurrence matrix texture analysis method is used to select uniformly textured sky and flat wall areas in the original image, and the standard deviation of pixel brightness in these areas is calculated to estimate the global noise standard deviation; the standard deviation of the current reflectance component in each color channel is calculated, and their average is taken as the color standard deviation.

[0012] Further, the specific steps for constructing an image processing database after storing the original image data and auxiliary processing data are as follows: The collected monitoring link data are aligned according to timestamps to ensure data synchronization across data sources at the same point in time; all illumination and reflectance data are adjusted to zero mean and unit variance to eliminate the influence of different dimensions and value ranges on the analysis results; for time series data, a sliding window method is used to generate training samples to capture time-dependent features and nonlinear dynamic features; the standardized and normalized original image data and auxiliary processing data are stored, and the image processing database is constructed.

[0013] Further, the specific steps for brightness deviation analysis using illumination component data are as follows: Obtain the illumination component value of each pixel in the image, the average illumination component of the entire image, and the illumination standard deviation; calculate the absolute value of the difference between the illumination component of each pixel and the average illumination component, and then divide it by the illumination standard deviation to obtain the relative deviation of the brightness of each pixel from the average level; subtract the relative deviation from one, and then multiply it by the illumination component of the corresponding pixel to obtain the brightness adjustment contribution value; sum the brightness adjustment contribution values ​​of all pixels and divide them by the total number of pixels to finally obtain the brightness deviation value.

[0014] Furthermore, the specific steps for performing differentiated illumination distribution optimization and brightness correction operations and outputting illumination enhancement components are as follows: By comparing the brightness deviation value and the deviation threshold in real time, when the brightness deviation value is less than the deviation threshold, each illumination component of the image is obtained and sorted in ascending order to obtain a brightness sorting queue. The top-ranked components in the brightness queue are then selected. The region containing the illumination component is defined as the low-illuminance region. Brightness adjustment is applied to local regions: low-frequency brightness regions are extracted by convolving the image with a Gaussian kernel; the image is decomposed using Retinex to enhance the low-frequency brightness components, while a local spatial adaptive linear transformation is used to increase the brightness of the low-illuminance region, and bilateral filtering is applied to smooth the low-frequency brightness region; the enhanced illumination component is output and archived into the image processing database; when the brightness deviation value is greater than or equal to the deviation threshold, global brightness enhancement is performed using the Retinex method to improve the overall image brightness and ensure a natural brightness transition; brightness adjustment is still applied to local regions; finally, the enhanced illumination component is output.

[0015] Furthermore, the specific steps for quantifying image noise and color deviation by integrating brightness adjustment, noise suppression, and color correction are as follows: Obtain the reflectance component of each pixel in the image, the reference reflectance component in the reference image, the average signal strength, the noise standard deviation, the color standard deviation, and the total number of pixels; calculate the average signal strength of each pixel, divide it by the noise standard deviation, and multiply it by the brightness adjustment coefficient to obtain the brightness adjustment term; calculate the absolute value of the difference between the pixel reflectance component and the reference reflectance component, divide it by the color standard deviation to obtain the color deviation term; calculate the brightness adjustment term plus the color deviation term to obtain the noise and color correction term for each pixel; sum the noise and color correction terms for all pixels and divide by the total number of pixels to obtain the noise correction value.

[0016] Further, the specific steps for performing denoising and detail enhancement processing based on the noise deviation results and outputting the corrected reflectance component are as follows: By comparing the noise correction value and the correction threshold in real time, when the noise correction value is less than the correction threshold, Gaussian filtering is used to denoise the image while preserving detail information. For low-light areas in the image, a local weighted average filtering method is used. For edges and textures, bilateral filtering is used to smooth flat areas of the image and preserve details, outputting the corrected reflectance component. When the noise correction value is greater than or equal to the correction threshold, a convolutional neural network is used; based on color space conversion; the Sobel operator is used to increase the high-frequency information of the image; local area details are balanced using contrast-limited adaptive histogram; texture restoration technology is used to restore the high-frequency information of the image, enhance the image's clarity and detail, and output the corrected reflectance component.

[0017] Furthermore, the specific steps for quality enhancement analysis based on image noise level, detail richness, and color fidelity are as follows: The enhanced image is obtained by calculating the pixel-level product of the illumination enhancement component and the corrected reflectivity component; the noise level is obtained by calculating the mean absolute value of the Laplacian transform of the image; the detail richness is obtained by calculating the mean gradient magnitude of the image, and the detail standard deviation is calculated through statistical analysis of the detail richness values ​​of multiple images; the color fidelity is obtained by calculating the pixel-level L1 distance with the reference image; and the noise level, noise standard deviation, and detail richness are then obtained. The noise level is calculated by dividing the noise level by the noise standard deviation and taking a negative exponent, then multiplying by a noise suppression weighting factor to obtain the noise suppression term. The detail richness is calculated by dividing the detail richness by the detail standard deviation and taking a negative exponent, then multiplying by a detail enhancement weighting factor to obtain the detail enhancement term. The color fidelity is calculated by dividing the color standard deviation by the color standard deviation and taking a negative exponent, then multiplying by a color fidelity weighting factor to obtain the color fidelity term. The noise suppression term, detail enhancement term, and color fidelity term are added together to obtain the quality enhancement value, which reflects the overall quality level of the enhanced image.

[0018] Furthermore, the specific steps for using the loss function to guide the dual-U-Net architecture to perform detail reconstruction and color optimization on images are as follows: Real-time comparison of the quality enhancement value and the quality threshold; establishing a loss function through reconstruction loss, consistency loss, and smoothness loss; when the quality enhancement value is less than the quality threshold, constraining the network by adding reconstruction loss and smoothness loss in the dual-U-Net architecture, while expanding the receptive field of feature extraction to capture the global illumination structure, and performing enhanced color correction in reflectivity optimization; after enhancement, the quality enhancement value is recalculated. If the quality enhancement value is still less than the quality threshold after n consecutive enhancements, the historical best quality enhancement value is output, marked as a difficult sample, and the image's luminance component, detail information, and parameter settings are archived in the image processing database, triggering incremental training to optimize the weights of the dual-U-Net network; when the quality enhancement value is greater than or equal to the quality threshold, it indicates that the image has reached the basic quality requirements, and detail optimization is performed: in the dual-U-Net, the network is guided to enhance high-frequency information by adding the weights of the consistency loss, multi-scale feature fusion technology is used to finely reconstruct textures and edges, and the final enhanced image is output.

[0019] Furthermore, based on quality feedback and hard samples, the specific steps for adjusting processing parameters and driving algorithm closed-loop optimization are as follows: Receive the final enhanced image and its quality enhancement value from the image enhancement output, and perform hierarchical verification: high-quality results are archived into the high-quality sample set, while substandard results are marked as hard samples and archived into the hard sample set. Based on the accumulation and analysis of hard samples, the threshold parameters of the preceding sequence are adjusted. When the number of accumulated hard samples reaches m, the incremental training process of deep learning is triggered. The weights of the dual U-Net network are optimized using the newly accumulated hard sample data, and the weight ratio of its multi-task loss function is updated to improve the image enhancement effect in complex low-light scenes.

[0020] Beneficial effects

[0021] The present invention has the following beneficial effects:

[0022] (1) This invention, through the analysis of illumination component data and the optimization of differentiated illumination distribution, can dynamically adjust the brightness to address the problem of insufficient brightness in low-light images, thereby achieving an adaptive control effect for image brightness enhancement, and effectively solving the problems of uneven brightness and local information loss in low-light images in the prior art.

[0023] (2) By using a dual U-Net architecture and combining reconstruction loss, consistency loss and smoothness loss, the present invention enables more accurate processing of image detail reconstruction and color optimization, thereby achieving fine optimization of image details and colors, and effectively solving the problems of low-light image detail loss and serious color deviation in the prior art.

[0024] (3) This invention, by calculating the noise standard deviation and color standard deviation and quantifying the image noise and color deviation, achieves effective suppression of noise and correction of color deviation in the image, thereby realizing the synergistic optimization effect of image noise control and color correction, and effectively solves the problem that noise and color deviation are difficult to optimize at the same time in the prior art.

[0025] (4) By accumulating and analyzing difficult samples, this invention can trigger the incremental training process of deep learning, improve the optimization effect of the weights of the dual U-Net network, and thus achieve continuous improvement of image enhancement effect, effectively solving the problem of insufficient generalization ability of image enhancement technology in complex low-light environment in the prior art.

[0026] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0027] Figure 1 This is a flowchart of an image enhancement method for adaptive brightness gain control under low illumination according to the present invention.

[0028] Figure 2 This is an image pixel brightness deviation analysis diagram for the present invention;

[0029] Figure 3 This is a schematic diagram of the entire process of reflectivity component correction processing in this invention;

[0030] Figure 4 This is a network architecture diagram of the dual U-Net architecture framework of the present invention;

[0031] Figure 5 This is a low-light image of the present invention and its corresponding enhancement result using dual U-Net. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] Please see Figures 1-5 This invention provides a technical solution: an image enhancement method with adaptive brightness gain control under low illumination, comprising: S1: collecting original image data and auxiliary processing data under low illumination environment, preprocessing them, and storing the original image data and auxiliary processing data to construct an image processing database; S2: performing brightness deviation analysis through illumination component data, performing differentiated illumination distribution optimization and brightness correction operations, and outputting illumination enhancement components; S3: fusing brightness control, noise suppression, and color correction, quantifying image noise and color deviation, performing denoising and detail enhancement processing based on the noise and color deviation results, and outputting corrected reflectivity components; S4: performing quality enhancement analysis through image noise level, detail richness, and color fidelity, and using a loss function to guide a dual U-Net architecture to perform detail reconstruction and color optimization of the image; S5: adjusting processing parameters and driving algorithm closed-loop optimization based on quality feedback and difficult samples.

[0034] Specifically, the preprocessing steps for acquiring raw image data and auxiliary processing data under low-light conditions are as follows: Image data is acquired under low-light conditions, including raw image data and auxiliary processing data. The raw image data mainly includes raw pixel brightness values ​​and the total number of pixels. Images under low-light conditions are acquired in real-time using a camera, directly obtaining the raw pixel brightness value of each pixel. The total number of pixels in the image is obtained by calculating the product of the width and height in the original image's resolution. To comprehensively evaluate the overall brightness of the image, the global average signal strength is calculated by summing all raw pixel brightness values ​​and dividing by the total number of pixels, reflecting the image's brightness level. Auxiliary processing data includes illumination components, reflectivity components, average illumination components, illumination standard deviation, average signal strength, noise standard deviation, reference reflectivity components, and color standard deviation. This data further helps improve image brightness and detail. By employing a fast Retinex decomposition algorithm based on guided filtering to process the raw image, the illumination and reflectivity components can be effectively separated and output. The arithmetic mean of all pixel values ​​for the illumination component is used to calculate the average illumination component. The standard deviation of all pixel values ​​for the illumination component is then calculated and used as the illumination standard deviation to characterize the distribution of image brightness. A gray-level co-occurrence matrix texture analysis method is used to select uniformly textured, flat areas of the sky and walls in the original image. The pixel brightness standard deviation of these areas is used to estimate the global noise standard deviation, further aiding in the assessment of image noise levels. The standard deviation of the current reflectance component across each color channel is calculated and averaged to provide support for subsequent color correction. Through this comprehensive data processing and analysis, more accurate parameters can be provided for enhancing low-light images, ensuring that subsequent image optimization operations effectively improve image quality.

[0035] This implementation scheme comprehensively acquires and processes image data under low-light conditions, including calculating the original image brightness value, illumination component, reflectance component, and noise and color standard deviation. This allows for accurate assessment of the image's brightness level, noise level, and color deviation. By applying the fast Retinex decomposition algorithm to separate illumination and reflectance, and using gray-level co-occurrence matrix analysis to extract noise information from textured, uniform regions, the global illumination and noise standard deviations can be effectively obtained, providing more reliable data support. This comprehensive analysis and processing of data lays the foundation for subsequent image brightness enhancement and noise suppression operations, ensuring improved image quality. Particularly in low-light image enhancement, more precise brightness and color optimization can be achieved, improving the overall visual effect.

[0036] Specifically, the steps for constructing an image processing database after storing the raw image data and auxiliary processing data are as follows: The collected monitoring link data are aligned according to timestamps to ensure data synchronization across data sources at the same point in time; all illumination and reflectance data are adjusted to zero mean and unit variance to eliminate the influence of different dimensions and value ranges on the analysis results; during this process, illumination and reflectance components are standardized by channel to ensure that the illumination and reflectance components of different channels are consistent within their numerical ranges, thereby improving data consistency and accuracy; for time-series data, a sliding window method is used to generate training samples to capture time-dependent features and nonlinear dynamic features. The length of the sliding window is typically set between 10 and 100, and the step size ranges from 1 to 10 to ensure sufficient data coverage and processing efficiency; the standardized and normalized image data is stored together with the auxiliary processing data to construct an image processing database, providing structured support for subsequent image enhancement and analysis.

[0037] In this implementation plan, the collected monitoring link data is synchronized and standardized to ensure that data from different sources can be processed uniformly, thereby eliminating the impact of inconsistent data units. Channel standardization is also performed on the illumination and reflectivity components of the images to ensure data consistency across different channels. For time-series data processing, a sliding window method is used to generate training samples, capturing temporal features and dynamic changes, further improving the efficiency and accuracy of data processing. The processed image data and auxiliary data are stored in an image processing database, providing high-quality, structured foundational data for subsequent image enhancement and analysis.

[0038] Specifically, the steps for brightness deviation analysis using illumination component data are as follows: First, obtain the illumination component value of each pixel in the image and calculate the average illumination component and illumination standard deviation for the entire image. This serves as an important basis for evaluating the global brightness level and distribution of the image. Second, calculate the absolute value of the difference between the illumination component of each pixel and the average illumination component, and divide it by the illumination standard deviation to obtain the relative deviation of each pixel's brightness from the average level. This relative deviation accurately reflects the position of each pixel in the global brightness distribution, thus quantifying the difference between its brightness and the average brightness of the entire image. Third, subtract the relative deviation from the average value and multiply it by the corresponding pixel's illumination component to obtain the brightness adjustment contribution value of that pixel. This brightness adjustment contribution value reflects the magnitude of each pixel's contribution to the overall brightness adjustment and can effectively adjust the gain of different brightness areas. Finally, sum the brightness adjustment contribution values ​​of all pixels and divide by the total number of pixels to obtain the image's brightness deviation value. This value measures the degree of overall image brightness adjustment and provides an important reference for subsequent brightness enhancement and optimization operations.

[0039] The brightness deviation value is calculated as follows: ;

[0040] In the formula, This represents the brightness deviation value, used to adjust the overall brightness of the image, reflecting the magnitude and direction of the brightness enhancement. Indicates the first in the image Each light component; It represents the average illumination component of an image, calculated as the mean of all illumination components in the image, and reflects the global brightness level of the image. This represents the standard deviation of illumination, typically the standard deviation of image brightness distribution or other indicators that measure the dispersion of image brightness. It represents the total number of pixels in an image, reflecting the pixel scale of the image.

[0041] The illumination component of pixel 1 in the image sample is 82, the average illumination component is 109.0, and the standard deviation of illumination is 53.2, resulting in a calculated brightness deviation of 76.81. The illumination component of pixel 2 in the image sample is 91, the average illumination component is 109.0, and the standard deviation of illumination is 53.2, resulting in a calculated brightness deviation of 84.90. The illumination component of pixel 3 in the image sample is 178, the average illumination component is 109.0, and the standard deviation of illumination is 53.2, resulting in a calculated brightness deviation of 140.68. The illumination component of pixel 4 in the image sample is 121, the average illumination component is 109.0, and the standard deviation of illumination is 53.2, resulting in a calculated brightness deviation of 112.67. The illumination component of pixel 5 in the image sample is 167, the average illumination component is 109.0, and the standard deviation of illumination is 53.2, resulting in a calculated brightness deviation of 135.10.

[0042] Table 1. Results of Image Brightness Adjustment and Deviation Analysis under Low Illumination Environment

[0043] Sample pixels Pixel illumination components Average light component Standard deviation of illumination Brightness deviation value 1 82 109.0 53.2 76.81 2 91 109.0 53.2 84.90 3 178 109.0 53.2 140.68 4 121 109.0 53.2 112.67 5 167 109.0 53.2 135.10

[0044] like Figure 2 As shown in Table 1, this is an image pixel brightness deviation analysis diagram provided in an embodiment of this application. Figure 2It can be seen that the brightness deviation value of the sample points is significantly positively correlated with their pixel illumination components; that is, the higher the original illumination intensity of a pixel, the greater its deviation from the average brightness of the image. For example, sample point 3 has the highest illumination component (178) and the largest brightness deviation value (140.68), indicating significant overexposure at this point. Conversely, sample point 1 has the lowest illumination component (82) and the smallest brightness deviation value (76.81), belonging to a dark area in the image. This pattern shows that under the same average illumination and standard deviation, the brightness deviation of each point is mainly determined by its own illumination intensity. This result can provide a basis for subsequent image enhancement algorithms, namely, suppressing brightness gain in high-deviation areas and improving detail visibility in low-deviation areas, thereby achieving local adaptive optimization and effectively improving the visual quality of low-light images.

[0045] In this implementation scheme, by calculating the difference between the illumination component of each pixel and the average illumination component of the entire image, and combining this with the illumination standard deviation, the deviation of each pixel from the overall image brightness level can be quantified. By adjusting the brightness deviation of each pixel and calculating its contribution to the overall brightness adjustment, a brightness deviation value is finally obtained. This process can accurately assess the distribution characteristics of image brightness and provide a basis for subsequent brightness optimization and adjustment, ensuring that image brightness enhancement is both targeted and globally consistent.

[0046] Specifically, the steps for performing differentiated illumination distribution optimization and brightness correction operations and outputting illumination enhancement components are as follows: By comparing the brightness deviation value and the deviation threshold in real time, when the brightness deviation value is less than the deviation threshold, each illumination component of the image is obtained and sorted in ascending order to obtain a brightness sorting queue. The region containing the top k% of the illumination components in the brightness queue is defined as the low-light region, where the value of k% is usually set to the range of 5% to 30%. Brightness adjustment is performed on the low-light region: the low-frequency brightness region is extracted by convolving the image with a Gaussian kernel to ensure a smooth transition of image brightness; then, the image is decomposed using Retinex theory to enhance the low-frequency brightness components to improve the uniformity of overall brightness. At the same time, a local spatial adaptive linear transformation method is used to increase the brightness of the low-light region, optimize the visual effect of the region, and smooth the low-frequency brightness region through bilateral filtering to preserve details and edges. The enhanced illumination components are output and archived into the image processing database to ensure the integrity of the basic data for subsequent analysis. When the brightness deviation value is greater than or equal to the deviation threshold, the Retinex method is used to perform global brightness enhancement to improve the overall image brightness and ensure that the image brightness transition is natural and there are no obvious boundary changes. At the same time, the brightness of local areas is still adjusted to enhance details and improve the visual quality of the image. Finally, the illumination enhancement component is output to ensure that the enhancement result meets the quality requirements.

[0047] In this implementation, by performing brightness deviation analysis and identifying low-light regions in the image, Gaussian kernel convolution is used to extract low-frequency brightness regions, and Retinex theory is applied to enhance the low-frequency brightness components. Local spatial adaptive linear transformation is used to increase the brightness of low-light regions, combined with bilateral filtering for smoothing, ensuring that image details are preserved while brightness is enhanced. When the brightness deviation value is less than the deviation threshold, fine-tuning of the brightness in low-light regions is performed, and the enhanced illumination component is output. When the brightness deviation value is greater than or equal to the deviation threshold, global brightness enhancement and local brightness adjustment are used to ensure a natural transition in image brightness, optimizing the overall brightness effect. The enhanced illumination component is archived in the image processing database, providing high-quality image data for subsequent processing.

[0048] Specifically, the steps for quantifying image noise and color deviation by integrating brightness adjustment, noise suppression, and color correction are as follows: Obtain the reflectance component of each pixel in the image, the reference reflectance component in the reference image, the average signal intensity, the noise standard deviation, the color standard deviation, and the total number of pixels; use linear regression analysis to establish a quantitative mapping between the brightness assessment value and the correction intensity adjustment amount to obtain the brightness adjustment coefficient; divide the average signal intensity of each pixel by the noise standard deviation and multiply it by the brightness adjustment coefficient to obtain the brightness adjustment term for each pixel; calculate the color deviation term for each pixel, which measures the degree of color deviation in the image by calculating the difference between the pixel's reflectance component and the reference reflectance component and dividing it by the color standard deviation, to obtain the noise and color correction term for each pixel; sum the noise and color correction terms for all pixels and divide by the total number of pixels to obtain the image's noise and color correction value. This value reflects the effect of noise and color correction in the image, providing a reliable basis for further image enhancement.

[0049] The noise correction value is calculated as follows: ;

[0050] In the formula, This represents the noise correction value, reflecting the effectiveness of noise suppression and color correction in the image; This represents the total number of pixels in the image; It represents the average signal strength in an image and is used to measure the image's brightness information; It represents the standard deviation of noise in an image, measuring the intensity of image noise; This represents the brightness adjustment coefficient, reflecting the effect of image brightness adjustment, and is used to improve the adaptability and dynamic adjustment capability of the formula; Indicates the first in the image The reflectance component of each pixel reflects the color information of the object's surface in the image; This represents the reference reflectance component in the reference image, used to compare the color deviation between the current image and the ideal image; It represents the color standard deviation in an image, measuring the range of color variation in an image.

[0051] In this implementation scheme, image noise and color correction are achieved through comprehensive analysis of the image's reflectance components, signal strength, noise, and color deviation. The reflectance components of each pixel in the image and the reference image are extracted, and the average signal strength, noise standard deviation, and color standard deviation of the image are calculated. Brightness adjustment and color deviation terms are calculated, and combined with a brightness adjustment coefficient, to correct the brightness and color of each pixel, ultimately yielding the noise and color correction terms. The final noise and color correction value is obtained by summing and averaging the correction terms for all pixels, thereby optimizing image brightness and color, improving image quality, and especially enhancing image performance in low-light environments.

[0052] Specifically, the steps for performing denoising and detail enhancement processing based on the noise deviation results and outputting the corrected reflectivity component are as follows: By comparing the noise correction value with the correction threshold in real time, such as... Figure 3 This diagram illustrates the entire process of reflectivity component correction in this implementation. It determines whether the noise and color deviation of the current image meet the correction criteria. When the noise correction value is less than the correction threshold, Gaussian filtering is used to denoise the image, smoothing it and removing noise while preserving detail. In low-light areas, a local weighted average filtering method is used to better balance noise suppression and detail preservation. For edges and textures in the image, bilateral filtering is employed to smooth flat areas and maintain edge details. The corrected reflectivity component is then output to ensure effective noise removal and maintain image structure. When the noise correction value is greater than or equal to the correction threshold, a convolutional neural network is used for image enhancement, combining color space transformation and the Sobel operator to enhance high-frequency information. Contrast-limited adaptive histogram equalization is used to process details in local areas, combined with texture restoration techniques to restore image details and high-frequency information, thereby improving image clarity and precision. The optimized corrected reflectivity component is then output.

[0053] In this implementation, by comparing the noise correction value with the correction threshold in real time, different image processing strategies are employed to optimize image quality based on the degree of noise and color correction. When the noise correction value is below the correction threshold, Gaussian filtering is used to remove noise from the image while preserving details, especially in low-light areas, where local weighted average filtering is used to further improve image quality. For edges and textured areas, bilateral filtering is used to maintain details while smoothing other areas. When the noise correction value is greater than or equal to the correction threshold, a convolutional neural network is used to further optimize the image. High-frequency information is enhanced through color space transformation and the Sobel operator, combined with adaptive histogram equalization and texture restoration techniques, significantly improving image sharpness and detail accuracy. The final output corrected reflectance component exhibits higher sharpness and color consistency.

[0054] Specifically, the steps for quality enhancement analysis based on image noise level, detail richness, and color fidelity are as follows: The enhanced image is obtained by calculating the pixel-level product of the illumination enhancement component and the corrected reflectivity component; the noise level is obtained by calculating the mean absolute value of the Laplacian transform of the image; the detail richness is obtained by calculating the mean gradient magnitude of the image, and the detail standard deviation is calculated through statistical analysis of the detail richness values ​​of multiple images; the color fidelity is obtained by calculating the pixel-level L1 distance with the reference image; the noise level, noise standard deviation, detail richness, detail standard deviation, color fidelity, and color standard deviation are obtained; the noise level in the image is statistically analyzed, and a weighted regression method is used to obtain the noise suppression weight factor by combining the noise standard deviation and the distribution of noise intensity; the image is then analyzed by calculating the noise level in the image. The detail richness and texture variation are combined with the detail standard deviation, and an optimization algorithm is used to adjust the detail enhancement weight factor. The color fidelity of the image is evaluated by comparing the color difference between the enhanced image and the reference image, and a method to minimize the error is used to obtain the color fidelity weight factor. The noise suppression term is obtained by dividing the noise level value by the noise standard deviation and taking a negative exponent, then multiplying it by the noise suppression weight factor. The detail enhancement term is obtained by dividing the detail richness value by the detail standard deviation and taking a negative exponent, then multiplying it by the detail enhancement weight factor. Finally, the color fidelity term is obtained by dividing the color standard deviation and taking a negative exponent, then multiplying it by the color fidelity weight factor. These three terms are then added together to obtain the quality enhancement value, which reflects the overall quality level of the enhanced image and is used for subsequent quality assessment and enhancement optimization.

[0055] The mass enhancement value is calculated as follows:

[0056] ;

[0057] In the formula, The quality enhancement value reflects the overall quality level of the initially enhanced image and is used to evaluate the image enhancement effect and guide subsequent enhancement strategies. This represents the noise suppression weighting factor, with a value range of [0,1], used to balance the importance of noise suppression in the overall quality assessment; This represents the detail enhancement weighting factor, with a value range of [0,1], used to balance the importance of detail richness in the overall quality assessment; This represents the color fidelity weighting factor, with a value range of [0,1], used to balance the importance of color accuracy in the overall quality assessment; This indicates the noise level value, reflecting the severity of noise in the initially enhanced image; This represents the detail richness value, reflecting the degree of richness of texture and edge details in the initially enhanced image; It represents the color fidelity value, reflecting the degree of color difference between the initially enhanced image and the reference image; It represents the standard deviation of noise and measures the intensity of image noise; It represents the standard deviation of detail, which measures the variation in texture and edge detail in an image; It represents the color standard deviation, which measures the change in the color distribution of an image.

[0058] In this implementation, an enhanced image is generated by calculating the pixel-level product of the illumination enhancement component and the corrected reflectivity component. Noise level and detail richness values ​​are calculated using the Laplacian transform and image gradient magnitude, and then the detail standard deviation is derived through statistical analysis of the detail richness. These values ​​collectively reflect the image's noise intensity, detail variations, and color differences compared to the reference image. The detail standard deviation is derived from multi-image statistics, ensuring consistency with the training data during the online inference phase. By weighting noise, detail, and color fidelity, a comprehensive quality enhancement value is generated, thereby performing comprehensive image quality optimization and ensuring that the final image achieves ideal results in noise suppression, detail enhancement, and color fidelity.

[0059] Specifically, the steps for using a loss function to guide the dual-U-Net architecture in image detail reconstruction and color optimization are as follows: Real-time comparison of the quality enhancement value and the quality threshold; establishment of a loss function through reconstruction loss, consistency loss, and smoothness loss. Through this loss function, the network can effectively optimize image details and colors. When the quality enhancement value is less than the quality threshold, the network utilizes the following steps within the dual-U-Net architecture: Figure 4This is the network architecture diagram of the dual-U-Net architecture framework in this implementation case. By adding constraints to the network with reconstruction loss and smoothness loss, and expanding the receptive field of feature extraction, the global illumination structure is captured, and enhanced color correction is performed in reflectivity optimization. After enhancement, the quality enhancement value is recalculated. If the quality enhancement value is still less than the quality threshold after n consecutive enhancements, the historical best quality enhancement value is output and marked as a difficult sample. The image's luminance component, detail information, and parameter settings are archived into an image processing database. The value of n ranges from [3,10] and is adjusted according to the processing target and image complexity. If the quality enhancement value is greater than or equal to the quality threshold, it indicates that the image has reached the basic quality requirements. Next, detail optimization is performed: In the dual-U-Net, the network is guided to enhance high-frequency information by increasing the weight of the consistency loss, and multi-scale feature fusion technology is used to finely reconstruct texture and edges, outputting the final enhanced image. Figure 5 This example shows the low-light image and its corresponding enhancement result using dual U-Net. Specifically: Figure 5 As shown in (a), the comparison between the audience area and the aquarium area in an indoor aquarium scene before and after enhancement is presented. The left side is the original low-light image, and the right side is the enhancement result after processing using this method. Figure 5 As shown in (b), the contrast between the building facade and the foreground plaza before and after enhancement is illustrated. The enhancement improves architectural details and foreground brightness while maintaining the layered transition of the sky area; Figure 5 As shown in (c), the comparison between the station building, track, and sky areas in the urban street scene before and after enhancement is demonstrated. The enhancement highlights the station building signs and crowd details while maintaining a natural and consistent sky color.

[0060] In this implementation, a loss function guides the dual-U-Net architecture to optimize image detail reconstruction and color adjustment. The network's loss function is dynamically adjusted based on the comparison between the quality enhancement value and a quality threshold, performing enhancement processing on low-quality images. By increasing reconstruction and smoothness losses, the receptive field of feature extraction is expanded to capture global illumination structure, and color correction is optimized. When the image reaches the quality standard, detail optimization enhances high-frequency information through consistency loss, and multi-scale feature fusion is used to reconstruct texture and edges, outputting the final optimized enhanced image.

[0061] Specifically, the steps for adjusting processing parameters and driving algorithm closed-loop optimization based on quality feedback and difficult samples are as follows: Receive the final enhanced image and its quality enhancement value from the image enhancement output, and perform tiered verification: high-quality results are archived into the high-quality sample set, while substandard results are marked as difficult samples and archived into the difficult sample set. Based on the accumulation and analysis of difficult samples, the threshold parameters of the preceding steps are adjusted. When the accumulated number of difficult samples reaches the set trigger condition m, the incremental training process of deep learning will be initiated. The value of m is greater than 1000 to ensure that the data is representative and can cover the diversity in low-light scenes. During incremental training, the early layers of the network are first frozen, and only the later layers are trained to adapt to the characteristics of difficult samples. As difficult samples accumulate, the early layers are gradually unfrozen, and the entire network is optimized together. The learning rate is typically set to 10⁻⁻⁶. 6 The algorithm is upgraded to 10⁻³, and dynamically adjusted based on training feedback to achieve higher quality image enhancement. This incremental training process optimizes the weight ratio of the dual U-Net network and the multi-task loss function, improving image enhancement performance in low-light environments.

[0062] In this implementation, this step achieves its effect by evaluating the quality of the image enhancement output, classifying and validating high-quality and difficult samples, and then adjusting the threshold parameters in the preceding steps. Once a certain number of difficult samples have accumulated, an incremental training process for deep learning is triggered. This process utilizes the newly accumulated difficult sample data to optimize the weights of the dual-U-Net network and updates the weight allocation of its multi-task loss function. This continuous optimization of the network improves image enhancement performance in low-light scenes, ensuring adaptability and robustness in complex environments, thereby achieving higher-quality image enhancement.

[0063] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0064] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. An image enhancement method with adaptive brightness gain control under low illumination, characterized in that, Includes the following steps: S1: Collect raw image data and auxiliary processing data under low light conditions, preprocess them, and then store the raw image data and auxiliary processing data to build an image processing database; the raw image data includes the raw pixel brightness value and the total number of pixels; the auxiliary processing data includes the illumination component, reflectivity component, average illumination component, illumination standard deviation, average signal intensity, noise standard deviation, reference reflectivity component, and color standard deviation. S2: Perform brightness deviation analysis using illumination component data, execute differentiated illumination distribution optimization and brightness correction operations, and output illumination enhancement components; the specific steps of performing brightness deviation analysis using illumination component data are as follows: obtain the illumination component value of each pixel in the image, and calculate the average illumination component and illumination standard deviation of the entire image; Calculate the absolute value of the difference between the illumination component of each pixel and the average illumination component, and divide it by the illumination standard deviation to obtain the relative deviation of the brightness of each pixel from the average level; subtract the relative deviation from one, and then multiply it by the illumination component of the corresponding pixel to obtain the brightness adjustment contribution value of that pixel. The brightness adjustment contribution values ​​of all pixels are summed and divided by the total number of pixels to obtain the brightness deviation value of the image. The specific steps for performing differentiated illumination distribution optimization and brightness correction operations and outputting illumination enhancement components are as follows: By comparing the brightness deviation value and the deviation threshold in real time, when the brightness deviation value is less than the deviation threshold, each illumination component of the image is obtained and sorted in ascending order to obtain a brightness sorting queue. The area where the illumination components of a preset proportion in the brightness queue are located is defined as the low-illuminance area; brightness adjustment is applied to the local area: the low-frequency brightness area is extracted by convolving the image with a Gaussian kernel; the image is decomposed using Retinex to enhance the low-frequency brightness components, and local spatial adaptive linear transformation is used to increase the brightness of the low-illuminance area, and bilateral filtering is applied to the low-frequency brightness area for smoothing. The enhanced illumination component is output and archived into the image processing database. When the brightness deviation value is greater than or equal to the deviation threshold, global brightness enhancement is performed using the Retinex method to improve the overall brightness of the image and ensure a natural brightness transition. Brightness adjustment is still applied to local areas. Finally, the enhanced illumination component is output. S3: Integrates brightness adjustment, noise suppression and color correction, quantifies image noise and color deviation, performs denoising and detail enhancement processing based on the noise and color deviation results and outputs the corrected reflectance component; S4: Quality enhancement analysis is performed based on the noise level, detail richness, and color fidelity of the image. The loss function guides the dual U-Net architecture to reconstruct details and optimize colors in the image. S5: Based on quality feedback and difficult samples, adjust processing parameters and drive closed-loop optimization of the algorithm.

2. The image enhancement method with adaptive brightness gain control under low illumination according to claim 1, characterized in that: The specific steps for preprocessing the raw image data and auxiliary processing data acquired under low-light conditions are as follows: The camera captures raw images in low-light environments in real time, directly obtaining the raw pixel brightness value of each pixel. The total number of pixels in the image is obtained by calculating the product of the width and height in the original image resolution. The global average signal strength is calculated by summing the brightness values ​​of all original pixels and dividing by the total number of pixels. The original image is processed by a fast Retinex decomposition algorithm based on guided filtering, and the illumination component and reflectance component are separated and output. The average illumination component is obtained by taking the arithmetic mean of all pixel values ​​of the illumination component. Calculate the standard deviation of all pixel values ​​of the illumination component, and use it as the illumination standard deviation; By acquiring a standard reference image under normal lighting conditions and applying the same Retinex decomposition algorithm, the corresponding reference reflectance components are obtained. The gray-level co-occurrence matrix texture analysis method is used to select flat areas of sky and walls with uniform texture in the original image, calculate the standard deviation of pixel brightness in the area, and estimate the global noise standard deviation in this way. Calculate the standard deviation of the current reflectance component in each color channel, and take the average as the color standard deviation.

3. The image enhancement method with adaptive brightness gain control under low illumination according to claim 1, characterized in that: The specific steps for constructing the image processing database after storing the original image data and auxiliary processing data are as follows: The collected monitoring link data are aligned according to timestamps to ensure data synchronization from all data sources at the same point in time; all illumination and reflectance data are adjusted to zero mean and unit variance to eliminate the influence of different dimensions and value ranges on the analysis results; for time series data, a sliding window method is used to generate training samples to capture time-dependent features and nonlinear dynamic features; the original image data and auxiliary processing data that have been standardized and normalized are stored in the image processing database, and an image processing database is constructed.

4. The image enhancement method with adaptive brightness gain control under low illumination according to claim 2, characterized in that: The specific steps for integrating brightness adjustment, noise suppression, and color correction to quantify image noise and color deviation are as follows: Obtain the reflectance component of each pixel in the image, the reference reflectance component in the reference image, the average signal strength, the noise standard deviation, the color standard deviation, and the total number of pixels; A linear regression analysis method was used to establish a quantitative mapping between the brightness assessment value and the correction intensity adjustment amount to obtain the brightness adjustment coefficient; The brightness adjustment term is obtained by calculating the average signal strength of each pixel, dividing it by the noise standard deviation, and then multiplying it by the brightness adjustment factor. The color deviation term is obtained by dividing the absolute value of the difference between the pixel reflectance component and the reference reflectance component by the color standard deviation. The noise and color correction term for each pixel is obtained by adding the brightness adjustment term to the color deviation term. The noise and color correction terms for all pixels are summed and divided by the total number of pixels to obtain the noise and color correction value.

5. The image enhancement method with adaptive brightness gain control under low illumination according to claim 4, characterized in that: The specific steps for performing denoising and detail enhancement processing based on the noise deviation results and outputting the corrected reflectance component are as follows: By comparing the noise correction value and the correction threshold in real time, when the noise correction value is less than the correction threshold, Gaussian filtering is used to denoise the image while preserving details. For low-light areas in the image, a local weighted average filtering method is used. For edges and textures, bilateral filtering is used to smooth flat areas of the image and preserve details, and the corrected reflectance component is output. When the noise correction value is greater than or equal to the correction threshold, a convolutional neural network is used; based on color space transformation; the Sobel operator is used to increase the high-frequency information of the image; local area details are balanced by contrast-limited adaptive histogram; texture restoration technology is used to restore the high-frequency information of the image, enhance the image's clarity and detail, and output the corrected reflectance component.

6. The image enhancement method with adaptive brightness gain control under low illumination according to claim 1, characterized in that: The specific steps for quality enhancement analysis based on image noise level, detail richness, and color fidelity are as follows: The enhanced image is obtained by calculating the pixel-level product of the illumination enhancement component and the corrected reflectivity component. The noise level is obtained by calculating the mean absolute value of the Laplacian transform of the image. The detail richness value is obtained by calculating the mean gradient magnitude of the image. The detail standard deviation is calculated by statistical analysis of the detail richness values ​​of multiple images. The color fidelity value is obtained by the pixel-level L1 distance with the reference image. Obtain noise level values, noise standard deviation, detail richness values, detail standard deviation, color fidelity values, and color standard deviation; The noise level value is calculated by dividing it by the noise standard deviation and taking the negative exponent, and then multiplied by the noise suppression weighting factor to obtain the noise suppression term; Calculate the detail richness value, divide it by the detail standard deviation and take the exponent, then add one, and multiply by the detail enhancement weighting factor to obtain the detail enhancement term; calculate the color fidelity value, divide it by the color standard deviation and take the negative exponent, and multiply by the color fidelity weighting factor to obtain the color fidelity term; The noise suppression, detail enhancement, and color fidelity terms are added together to obtain the quality enhancement value, which reflects the overall quality level of the enhanced image.

7. The image enhancement method with adaptive brightness gain control under low illumination according to claim 6, characterized in that: The specific steps for using a loss function to guide the dual U-Net architecture to perform detail reconstruction and color optimization of images are as follows: The quality enhancement value and the quality threshold are compared in real time, and a loss function is established through reconstruction loss, consistency loss, and smoothness loss. When the quality enhancement value is less than the quality threshold, the dual U-Net architecture is used to constrain the network by adding reconstruction loss and smoothness loss, while expanding the receptive field of feature extraction to capture the global illumination structure and performing enhanced color correction in reflectivity optimization. After enhancement, the quality enhancement value is calculated again. If the quality enhancement value is still less than the quality threshold after n consecutive enhancements, the historical best quality enhancement value is output, marked as a hard sample, and the brightness component, detail information, and parameter settings of the recorded image are archived to the image processing database. Incremental training is then triggered to optimize the weights of the dual U-Net network. When the quality enhancement value is greater than or equal to the quality threshold, it indicates that the image has reached the basic quality requirements. Then, detailed optimization is performed: in the dual U-Net, the network is guided to enhance high-frequency information by increasing the weights of the consistency loss, and multi-scale feature fusion technology is used to finely reconstruct texture and edges, and the final enhanced image is output.

8. The image enhancement method with adaptive brightness gain control under low illumination according to claim 7, characterized in that: The specific steps for adjusting processing parameters and driving algorithm closed-loop optimization based on quality feedback and difficult samples are as follows: The system receives the final enhanced image and its quality enhancement value from the image enhancement output and performs hierarchical verification: high-quality results are archived into the high-quality sample set, while substandard results are marked as difficult samples and archived into the difficult sample set. Based on the accumulation and analysis of difficult samples, the system adjusts the threshold parameters of the preceding sequence. When the number of accumulated difficult samples reaches the set trigger condition m, the system triggers the incremental training process of deep learning. The value of m is greater than 1000. The system optimizes the weights of the dual U-Net network using the newly accumulated difficult sample data and updates the weight ratio of its multi-task loss function to improve the image enhancement effect in complex low-light scenes.

Images