Low-illumination image enhancement method based on information fusion strategy

Abstract

The invention belongs to the technical field of image enhancement, and particularly discloses a low-illumination image enhancement method based on an information fusion strategy, and the method comprises the steps: carrying out the processing of an inputted low-illumination image through an image enhancement algorithm based on adaptive Gamma correction, and generating a first group of multi-exposure image sequences; processing the same input image through an improved quadratic function enhancement algorithm to generate a second group of multi-exposure image sequences; and performing multi-scale fusion on the first group of multi-exposure image sequences and the second group of multi-exposure image sequences, including distributing the weight of each image through a weight strategy, performing multi-scale decomposition on the images by using a Gaussian pyramid and a Laplacian pyramid, and reconstructing and generating a final enhanced image after layer-by-layer fusion. According to the method provided by the invention, the advantages of brightness enhancement and color optimization are combined, and an image with balanced brightness, rich details and natural colors can be generated in a low-illumination scene.

Description

Technical Field

[0001] The present invention relates to the technical field of image enhancement, and in particular to a low-illumination image enhancement method based on an information fusion strategy. Background Art

[0002] In dim scenes, such as those with point light sources, backlit scenes, or high dynamic range conditions with large reflectivity differences, underexposed or overexposed areas often occur, making it difficult to fully capture scene information and resulting in information loss. To address this issue, Multi-Exposure Fusion (MEF) technology, which integrates image sequences with varying exposure levels, has become a key approach to improving low-light image quality. However, existing methods still have limitations in extreme exposure conditions and algorithmic complexity.

[0003] In traditional MEF algorithms, methods based on quality metrics and color correction (such as contrast, saturation, and brightness metrics combined with Laplacian pyramid fusion) can improve image visual quality, but they are highly dependent on the number of source images involved in the fusion and have difficulty addressing information loss in extreme exposure scenarios. Furthermore, some methods reconstruct images by decomposing image blocks into independent components such as signal intensity and structure. While this improves fusion quality, it is also limited by the number of source images and cannot effectively address the dynamic range compression requirements under extreme exposure conditions. Other studies have proposed using intensity mapping to generate virtual transition images to assist in fusion, but these methods still require additional images and cannot achieve true double-exposure fusion, limiting the algorithm's efficiency and scope of application.

[0004] In recent years, deep learning techniques (such as convolutional neural networks and generative adversarial networks) have been introduced into the MEF field, achieving image enhancement through supervised training. Although these methods have performed well in some scenarios, they rely on artificially constructed "ground truth" data, which lacks authenticity. Furthermore, model training is complex and computationally expensive, making them unable to meet the real-time and universal requirements of practical applications.

[0005] The above analysis reveals that existing low-light image enhancement technologies urgently need improvement in the following areas: 1) Traditional fusion algorithms lack adaptability to extreme exposure scenarios and rely on multi-source input; 2) Deep learning-based methods require complex training and expensive labeled data, limiting their practicality; and 3) Existing methods lack the ability to balance brightness balance, detail preservation, and color naturalness. Therefore, a low-light image enhancement algorithm that requires no complex training, is highly adaptable, and can effectively fuse dual-channel information is urgently needed to address these issues. Summary of the Invention

[0006] The purpose of the present invention is to provide a low-light image enhancement method based on an information fusion strategy to address the above problems, so as to overcome the traditional algorithm's dependence on multi-source input, avoid the complex training of deep learning, and achieve optimization of brightness balance, detail preservation and color naturalness.

[0007] The present invention discloses a low-light image enhancement method based on an information fusion strategy, which comprises the following steps:

[0008] Step S100: Processing an input low-light image using an image enhancement algorithm based on adaptive gamma correction to generate a first set of multi-exposure image sequences. Step S200: Processing the same input image using an improved quadratic function enhancement algorithm to generate a second set of multi-exposure image sequences. Step S300: Multi-scale fusion of the first and second sets of multi-exposure image sequences, including: assigning weights to each image using a weighting strategy, performing multi-scale decomposition of the image using a Gaussian pyramid and a Laplacian pyramid, and reconstructing the final enhanced image after layer-by-layer fusion.

[0009] Furthermore, the image enhancement algorithm based on adaptive gamma correction in step S100 specifically includes:

[0010] Step S110: convert the input image into a grayscale image and perform noise detection;

[0011] Step S120: performing noise reduction processing on the detected noise area;

[0012] Step S130: Convert the image to the HSV color space, and dynamically adjust the Gamma value based on the mean and standard deviation of the brightness component.

[0013] Furthermore, the step S110 includes:

[0014] Step S111: Convert the input RGB image into a grayscale image. The conversion formula is:

[0015] ,

[0016] in, are the red, green, and blue channels of the image respectively;

[0017] Step S112: Decompose the grayscale image using a two-dimensional wavelet transform to divide the image into components of different frequencies, where the high-frequency components contain edge information and noise of the image. The wavelet transform calculation formula is:

[0018] ,

[0019] Where C is the coefficient after wavelet transform, which contains information of different decomposition levels; S is the size matrix of wavelet transform, which represents the image size of each level of decomposition;

[0020] Step S113: Calculate the absolute value of the wavelet coefficient and compare it with the set threshold T to determine whether there is noise. If the noise mask If some positions in the is 1, it means that noise is detected at that position. The formula is

[0021] ,

[0022] in and are the mean and standard deviation of the pixels in the window centered at pixel (x, y), respectively.

[0023] Furthermore, the noise reduction process in step 120 includes performing noise reduction on the image using bilateral filtering, and the specific steps are as follows:

[0024] Step S121: Based on the noise mask Determine the pixels that need to be filtered and define the neighborhood of each noise pixel (r, c) as

[0025] ,

[0026] Make sure the neighborhood is within the image dimensions

[0027] ;

[0028] Step S122: Generate neighborhood coordinates through the grid

[0029] ,

[0030] Calculating spatial weights

[0031] ,

[0032] Calculating range weights

[0033] ,

[0034] in, Represents the pixel value matrix of the current neighborhood;

[0035] Step S123: noise pixels For each neighborhood of , calculate the weighted average as the denoised pixel value, and update the pixel value to

[0036] ,

[0037] in, is the current pixel Neighborhood, is the spatial weight, is the range weight.

[0038] Furthermore, the step 130 includes:

[0039] Convert the image from RGB space to HSV space and process the brightness component separately. The brightness component V corresponds to the third channel in the HSV image. The formula for image conversion and brightness component extraction is

[0040] ,

[0041] ,

[0042] According to the average brightness of the brightness component and contrast , select the gamma value range

[0043] ,

[0044] According to the average brightness, select the appropriate gamma value range:

[0045] ,

[0046] Apply gamma correction to the luminance component V

[0047] ,

[0048] The generated brightness value is further improved

[0049] ,

[0050] Make sure the result is in the range [0, 1]:

[0051] .

[0052] Furthermore, the improved quadratic function enhancement algorithm in step 200 includes:

[0053] Step S210: Convert the input image to the HSV color space and extract the brightness channel;

[0054] Step S220: classify the image into three types: underexposure, normal exposure, and overexposure based on the global mean and standard deviation of the brightness channel;

[0055] Step S230: dynamically generate an adaptive enhancement coefficient according to the exposure type, and combine the saturation channel to suppress the enhancement amplitude of the high saturation area;

[0056] Step S240: Generate output images with multiple levels of different enhancement intensities through nonlinear mapping.

[0057] Furthermore, in step 220:

[0058] The calculation formula for brightness mean and standard deviation is:

[0059]

[0060]

[0061] in, is the total number of pixels;

[0062] The exposure type classification rules are: For underexposure, if For normal exposure, if For overexposure.

[0063] Furthermore, in step S230:

[0064] Self-adaptive enhancement factor The calculation formula is

[0065] ,

[0066] in, is the basic enhancement coefficient, and To control the parameters, the sensitivity of the enhancement degree and standard deviation were adjusted respectively.

[0067] Furthermore, in step S240:

[0068] The multi-level enhancement result is generated by the following formula

[0069] ,

[0070] in, They correspond to five different levels of adaptive enhancement coefficients, from low to high.

[0071] Furthermore, the step S300 includes:

[0072] Step S310: Acquire size parameters of the first and second groups of multi-exposure image sequences, and select a number of pyramid layers suitable for the current image size according to the input parameters;

[0073] Step S320: normalize the weights of each group of input images at the same pixel position so that the sum of the weight contributions of all input images is 1;

[0074] Step S330: construct a Gaussian pyramid and a Laplacian pyramid for each set of input images, and perform weighted fusion on multiple input images at each pyramid level;

[0075] Step S340: Restore the fused multi-scale pyramid to the final enhanced image through an inverse operation.

[0076] Furthermore, the step S310 includes:

[0077] Get the size parameters of the first and second multi-exposure image sequences, where [H, W, C, N] represent the image height, width, number of channels, and number of input images, respectively.

[0078] By checking the input parameter lev, select a pyramid level suitable for the current image size

[0079] ,

[0080] The pyramid is decomposed at most to a scale logarithmic of the minimum dimension of the input image.

[0081] Furthermore, the step S320 includes:

[0082] Normalize the weight of each input image so that the sum of the weights of all input images at the same pixel position is 1, making the fused weight smooth and distortion-free. Suppose the size of the input image is , weight Indicates that the kth image is at position The initial weight distribution is , the normalization step is expressed as

[0083] .

[0084] Furthermore, the step S330 includes:

[0085] The Gaussian pyramid generates images of different resolutions through multiple Gaussian filtering and downsampling. The Laplacian pyramid is used to store the differences between images of different scales.

[0086] ,

[0087] For each input image , calculate its Gaussian pyramid and Laplace pyramid After pyramid decomposition, the images at each level are fused by multi-scale weighted fusion. Layer, the fused image is

[0088] ,

[0089] in, The nth image is in the The weight of the layer;

[0090] The fused image is restored using the Laplacian pyramid reconstruction process.

[0091] Furthermore, the step S340 includes:

[0092] The final fused image is obtained by reconstructing the weighted fused images of all levels

[0093] ,

[0094] in, represents the inverse operation of the Gaussian pyramid, is the fused image layer, and the final fused image Preserve dominant regions from multiple input images.

[0095] The present invention's beneficial effects lie in its proposed low-light image enhancement method, based on an information fusion strategy. This method overcomes the reliance of traditional algorithms on multiple input sources and avoids the complex training required for deep learning. By combining the advantages of brightness enhancement and color optimization, it can produce images with balanced brightness, rich detail, and natural color in low-light scenarios. Compared to traditional algorithms, the method provided by the invention achieves significant improvements in brightness distribution, detail preservation, and color naturalness. BRIEF DESCRIPTION OF THE DRAWINGS

[0096] Figure 1 It is a flow chart of the image enhancement algorithm based on adaptive gamma correction in the present invention;

[0097] Figure 2 It uses an image enhancement algorithm based on adaptive gamma correction to generate a comparison chart of the image and the original input image;

[0098] Figure 3 is a flow chart of the improved quadratic function enhancement algorithm of the present invention;

[0099] Figure 4 It is a comparison chart of the image generated by the improved quadratic function enhancement algorithm and the original input image;

[0100] Figure 5 is a comparison diagram of the image processed by the method of the present invention and the original input image;

[0101] Figure 6 These are three sets of comparison results of images processed using five different processing methods;

[0102] Figure 7This is a flow chart of the low-illumination image enhancement method based on the information fusion strategy provided by the present invention. DETAILED DESCRIPTION

[0103] The core of the present invention is to construct two different feature extraction enhancement methods to generate two sets of virtual multi-exposure images respectively, and then use the image fusion strategy to generate a new image from these two sets of images to achieve the purpose of image enhancement.

[0104] Reference Figure 7 The low-light image enhancement method based on the information fusion strategy provided by the present invention includes the following steps:

[0105] Step S100: Processing an input low-light image using an image enhancement algorithm based on adaptive gamma correction to generate a first set of multi-exposure image sequences. Step S200: Processing the same input image using an improved quadratic function enhancement algorithm to generate a second set of multi-exposure image sequences. Step S300: Multi-scale fusion of the first and second sets of multi-exposure image sequences, including: assigning weights to each image using a weighting strategy, performing multi-scale decomposition of the image using a Gaussian pyramid and a Laplacian pyramid, and reconstructing the final enhanced image after layer-by-layer fusion.

[0106] In some embodiments of the present invention, the image enhancement algorithm of adaptive gamma correction, the image enhancement algorithm based on the improved quadratic function method, and the image fusion are specifically as follows:

[0107] 1. Image enhancement algorithm based on adaptive gamma correction

[0108] In order to obtain a virtually exposed image, this paper proposes an image enhancement algorithm based on adaptive gamma correction. By combining noise detection and noise reduction, adaptive gamma correction, and saturation and hue enhancement, it generates image results with different exposure levels for subsequent applications such as high dynamic range imaging.

[0109] Reference Figure 1 The image enhancement algorithm based on adaptive gamma correction includes: image input; image preprocessing, conversion to grayscale image; noise detection, wavelet transform; noise reduction, bilateral filtering; HSV space conversion and brightness component extraction; adaptive gamma value adjustment; gamma correction and brightness enhancement; color enhancement, saturation and hue adjustment; and output image. The details of the image enhancement algorithm based on adaptive gamma correction are described below.

[0110] 1.1 Noise Detection

[0111] Noise usually manifests itself as brightness changes, so converting the image to grayscale can simplify the noise detection process by removing color information and retaining only brightness information. The calculation formula is as follows:

[0112] (1)

[0113] in, are the red, green, and blue channels of the image, respectively.

[0114] Using two-dimensional wavelet transform to decompose grayscale images can effectively divide the image into components of different frequencies, where high-frequency components usually contain edge information and noise of the image. The calculation formula of wavelet transform is as follows:

[0115] (2)

[0116] Among them, C is the coefficient after wavelet transform, which contains the information of different decomposition levels; S is the size matrix of wavelet transform, which represents the image size of each level of decomposition.

[0117] By calculating the absolute value of the wavelet coefficient and comparing it with the set threshold T, we can determine whether there is noise. Some positions in are 1, indicating that noise is detected at that position.

[0118] (3)

[0119] in and are the mean and standard deviation of the pixels in the window centered at pixel (x, y), respectively.

[0120] 1.2 Noise Reduction Processing

[0121] If noise is detected, bilateral filtering is used to reduce the noise of the image. The advantage of bilateral filtering is that it can preserve the edge information of the image while removing the noise.

[0122] According to the noise mask To determine which pixels need to be filtered. For each noise pixel (r, c), its neighborhood is defined as:

[0123] (4)

[0124] Make sure the neighborhood fits within the image dimensions:

[0125] (5)

[0126] Generate neighborhood coordinates using a grid:

[0127] (6)

[0128] Calculate spatial weights:

[0129] (7)

[0130] Calculate range weight:

[0131] (8)

[0132] in, A matrix representing the pixel values ​​of the current neighborhood.

[0133] For noisy pixels For each neighborhood of , bilateral filtering calculates the weighted average as the denoised pixel value. Update pixel value:

[0134] (9)

[0135] in, is the current pixel Neighborhood, is the spatial weight, is the range weight.

[0136] 1.3 Dynamic gamma adjustment

[0137] Convert the image from RGB space to HSV space so that the brightness component can be processed separately. The brightness component V corresponds to the third channel in the HSV image. The formula for image conversion and brightness component extraction is:

[0138] (10)

[0139] (11)

[0140] According to the average brightness of the brightness component and contrast , select the gamma value range:

[0141] (12)

[0142] According to the average brightness, select the appropriate gamma value range:

[0143] (13)

[0144] Apply gamma correction to the luminance component V:

[0145] (14)

[0146] The resulting brightness value is further improved:

[0147] (15)

[0148] Make sure the result is in the range [0, 1]:

[0149] (16)

[0150] Figure 2 It is a comparison chart of the image generated by the image enhancement algorithm based on adaptive gamma correction and the original input image. The ten pictures are divided into two groups according to the rows. Each group corresponds to an input image and four output images generated by the image enhancement algorithm based on adaptive gamma correction from left to right.

[0151] pass Figure 2 As can be seen from the comparison in Figure 2, the adaptive gamma correction-based image enhancement algorithm proposed in this section addresses the shortcomings of traditional methods for low-light image enhancement by dynamically adjusting gamma values ​​and enhancing saturation. This method excels in brightness adjustment and detail preservation, and can also significantly improve the brightness, contrast, and color saturation of low-light images.

[0152] 2. Improved quadratic function enhancement algorithm

[0153] The improved quadratic function enhancement algorithm aims to improve the brightness, contrast and detail performance of low-light images through adaptive parameter adjustment and multi-level enhancement strategy.

[0154] Reference Figure 3 The improved quadratic function enhancement algorithm includes: inputting an image; converting the image to the HSV color space; extracting the luminance and saturation channels; calculating the luminance mean and standard deviation; adjusting the enhancement parameters based on the exposure type; performing multi-level enhancement on the luminance channel using an improved quadratic method; optimizing the enhancement coefficients based on the saturation channel; combining the enhanced luminance channel with the original hue and saturation channels; converting back to the RGB color space; and outputting the image. The details of the image enhancement algorithm based on the improved quadratic function method are described below.

[0155] 2.1 Theoretical basis

[0156] The traditional quadratic function method is a nonlinear enhancement technology based on brightness adjustment. Its core idea is to enhance the image by using the nonlinear characteristics of the quadratic curve. The formula is as follows:

[0157] (17)

[0158] in, is the original brightness channel, is the enhancement factor, by adjusting The improved method introduces an adaptive parameter adjustment mechanism based on this, making the enhancement process more targeted.

[0159] To improve the robustness of the algorithm, the algorithm dynamically adjusts the enhancement parameters based on the overall brightness of the image and the brightness standard deviation. The specific steps are as follows:

[0160] a. Calculate the brightness mean and standard deviation:

[0161] (18)

[0162] (19)

[0163] in, is the total number of pixels.

[0164] b. Define the enhancement interval:

[0165] According to the mean brightness The value range defines three exposure types: underexposure , requiring high-intensity enhancement; normal exposure , needs moderate enhancement; overexposed , requiring low-intensity reinforcement The actual value of is 0, 1, 2, 3, 4. Different exposure images are obtained by taking different values. is the mean of the image, ranging from [0,1]. Pure black is 0 and pure white is 1.

[0166] c. Adaptive adjustment formula

[0167] (20)

[0168] in, is the basic enhancement coefficient, and To control the parameters, the sensitivity of the enhancement degree and standard deviation were adjusted respectively.

[0169] 2.2 Saturation Perception Enhancement Strategy

[0170] The algorithm generates five different levels of enhancement results based on the exposure type of the image. The specific calculation formula is:

[0171] (twenty two)

[0172] in They correspond to five different levels of adaptive enhancement coefficients, from low to high.

[0173] Figure 4Comparison diagram of the image generated by the improved quadratic function enhancement algorithm and the original input image, including ten images. The ten images are divided into two groups according to the rows. Each group corresponds to an input image and four output images generated by the image enhancement algorithm based on the improved quadratic function method from left to right.

[0174] Combining the above discussion and Figure 4 The results demonstrate the theoretical and experimental effectiveness of the improved quadratic function enhancement algorithm proposed in this paper. Through adaptive parameter adjustment, saturation-aware enhancement, and a multi-level enhancement strategy, the improved quadratic function enhancement algorithm can meet image enhancement requirements in various scenarios, providing a solid foundation for subsequent image analysis and processing tasks.

[0175] 3. Image Fusion

[0176] 3.1 Parameter initialization and size confirmation

[0177] First, obtain the size of the multi-exposure image sequence generated by the adaptive gamma correction image enhancement algorithm and the improved quadratic function enhancement algorithm. [H, W, C, N] represents the image height, width, number of channels (usually 3 channels of RGB), and the number of input images. At the same time, by checking the input parameter lev, a pyramid level number suitable for the current image size is selected:

[0178] (twenty three)

[0179] The pyramid is decomposed at most to a scale logarithmic of the minimum dimension of the input image.

[0180] 3.2 Weight Normalization

[0181] The first step of image fusion is to assign a corresponding weight to each input image. In order to ensure the rationality of global image fusion, the weights need to be normalized so that the sum of all weights at the same pixel position is equal to 1. Suppose the size of the input image is , weight Indicates that the kth image is at position The initial weight distribution is , then the normalization step can be expressed as:

[0182] (twenty four)

[0183] This ensures that at each pixel location, the sum of the weight contributions of all input images is always 1, which makes the fused weights smooth and distortion-free.

[0184] 3.3 Multi-scale Pyramid Decomposition

[0185] During image fusion, we use Gaussian and Laplacian pyramids to process image information at different scales. The Gaussian pyramid generates images of different resolutions through multiple Gaussian filtering and downsampling, while the Laplacian pyramid is used to store the differences between images of different scales.

[0186] (25)

[0187] For each input image (from different exposure sequences), we first calculate its Gaussian pyramid and Laplace pyramid ; After pyramid decomposition, we fuse the images of each level through multi-scale weighted fusion; Layer, the fused image is:

[0188] (26)

[0189] in, The nth image is in the The weights of the layers are obtained; finally, the fused image is restored using the Laplacian pyramid reconstruction process.

[0190] 3.4 Image Reconstruction

[0191] By reconstructing the weighted fusion images of all levels, we obtain the final fusion image

[0192] (27)

[0193] in, represents the inverse operation of the Gaussian pyramid, is the fused image layer, and the final fused image The dominant regions in multiple input images are retained. In the experiment, the number of layers L is set to 6 based on experience. The ideal result can be reconstructed through 7 layers from 0 to 6.

[0194] 4. Experiment and Analysis

[0195] In order to verify the effectiveness of the method proposed in this paper, this paper uses multiple public datasets such as DICM, LIME, LOL, and FUSION to test the algorithm, and experimentally compares and analyzes the algorithm in this paper with the DONG algorithm, IBOOST, LDR, and PLM.

[0196] 4.1 Subjective Evaluation

[0197] First, we conduct a subjective perception analysis of the algorithm results from the perspective of visual perception. The following are some of the results obtained by running the algorithm in this paper on a self-built dataset, such as Figure 5As shown in the figure, there are four rows of twenty images in total, which are divided into four groups according to the rows. The first group is five input images, the second group is five output images after the first group of images is processed by the method of the present invention, the third group is another five input images, and the fourth group is five output images after the third group of images is processed by the method of the invention.

[0198] In order to verify the effectiveness of the algorithm in this paper, a variety of advanced low-light image enhancement algorithms were run on the DICM dataset, LIME dataset, LOL dataset, and FUSION dataset, and comparative experiments were conducted. Some experimental results were selected, such as Figure 6 As shown in the figure, there are three rows of eighteen images in total. The images are divided into three groups according to the number of rows. The six images in each group represent the input image, the image processed by the DONG algorithm, the image processed by the IBOOST algorithm, the image processed by the LDR algorithm, the image processed by the PLM algorithm, and the image processed by the method provided by the present invention in order from left to right.

[0199] The low-light image enhancement method based on the information fusion strategy provided by the present invention adopts a dual-channel fusion algorithm based on adaptive gamma correction and quadratic function, combines the advantages of brightness enhancement and color optimization, and can generate images with balanced brightness, rich details and natural colors in low-light scenes. Compared with traditional algorithms, the method provided by the present invention has significant improvements in brightness distribution, detail retention and color naturalness. In summary, the experimental results of the low-light image enhancement method based on the information fusion strategy provided by the present invention on multiple data sets show that it performs well in brightness distribution, detail retention, color naturalness and edge feature enhancement. Through the comprehensive analysis of subjective evaluation and objective evaluation, it can be seen that the method provided by the present invention has significant advantages and broad application prospects in the field of low-light image enhancement.

Claims

1. A low-light image enhancement method based on information fusion strategy, characterized in that: The method comprises the following steps: Step S100: Processing the input low-light image by an image enhancement algorithm based on adaptive gamma correction to generate a first set of multi-exposure image sequences; Step S200: Processing the same input image using an improved quadratic function enhancement algorithm to generate a second set of multi-exposure image sequences; Step S300 , performing multi-scale fusion on the first set of multi-exposure image sequences and the second set of multi-exposure image sequences, including: assigning weights to each image using a weighting strategy, performing multi-scale decomposition of the images using a Gaussian pyramid and a Laplacian pyramid, and reconstructing and generating a final enhanced image after layer-by-layer fusion.

2. The low-light image enhancement method based on information fusion strategy according to claim 1, characterized in that: The image enhancement algorithm based on adaptive gamma correction in step S100 specifically includes: Step S110: convert the input image into a grayscale image and perform noise detection; Step S120: performing noise reduction processing on the detected noise area; Step S130: Convert the image to the HSV color space, and dynamically adjust the Gamma value based on the mean and standard deviation of the brightness component.

3. The low-light image enhancement method based on information fusion strategy according to claim 2, characterized in that: The step S110 includes: Step S111: Convert the input RGB image into a grayscale image. The conversion formula is: ， in, are the red, green, and blue channels of the image respectively; Step S112: Decompose the grayscale image using a two-dimensional wavelet transform to divide the image into components of different frequencies, where the high-frequency components contain edge information and noise of the image. The wavelet transform calculation formula is: ， Where C is the coefficient after wavelet transform, which contains information of different decomposition levels; S is the size matrix of wavelet transform, which represents the image size of each level of decomposition; Step S113: Calculate the absolute value of the wavelet coefficient and compare it with the set threshold T to determine whether there is noise. If the noise mask If some positions in the is 1, it means that noise is detected at that position. The formula is ， in and are the mean and standard deviation of the pixels in the window centered at pixel (x, y), respectively.

4. The low-light image enhancement method based on information fusion strategy according to claim 2, characterized in that: The noise reduction process in step 120 includes performing noise reduction on the image using bilateral filtering, and the specific steps are as follows: Step S121: Based on the noise mask Determine the pixels that need to be filtered and define the neighborhood of each noise pixel (r, c) as ， Make sure the neighborhood is within the image dimensions ； Step S122: Generate neighborhood coordinates through the grid ， Calculating spatial weights ， Calculating range weights ， in, Represents the pixel value matrix of the current neighborhood; Step S123: noise pixels For each neighborhood of , calculate the weighted average as the denoised pixel value, and update the pixel value to ， in, is the current pixel Neighborhood, is the spatial weight, is the range weight.

5. The low-light image enhancement method based on information fusion strategy according to claim 2, characterized in that: The step 130 includes: Convert the image from RGB space to HSV space and process the brightness component separately. The brightness component V corresponds to the third channel in the HSV image. The formula for image conversion and brightness component extraction is ， ， According to the average brightness of the brightness component and contrast , select the gamma value range ， According to the average brightness, select the appropriate gamma value range: ， Apply gamma correction to the luminance component V ， The generated brightness value is further improved ， Make sure the result is in the range [0, 1]: 。 6. The low-light image enhancement method based on information fusion strategy according to claim 1, characterized in that: The improved quadratic function enhancement algorithm in step 200 includes: Step S210: Convert the input image to the HSV color space and extract the brightness channel; Step S220: classify the image into three types: underexposure, normal exposure, and overexposure based on the global mean and standard deviation of the brightness channel; Step S230: dynamically generate an adaptive enhancement coefficient according to the exposure type, and combine the saturation channel to suppress the enhancement amplitude of the high saturation area; Step S240: Generate output images with multiple levels of different enhancement intensities through nonlinear mapping.

7. The low-light image enhancement method based on information fusion strategy according to claim 6, characterized in that: In step 220: The calculation formula for brightness mean and standard deviation is: ； ； in, is the total number of pixels; The exposure type classification rules are: For underexposure, if For normal exposure, if For overexposure.

8. The low-light image enhancement method based on information fusion strategy according to claim 6, characterized in that: In step S230: Self-adaptive enhancement factor The calculation formula is ， in, is the basic enhancement coefficient, and To control the parameters, the sensitivity of the enhancement degree and standard deviation were adjusted respectively.

9. The low-light image enhancement method based on information fusion strategy according to claim 6, characterized in that: In step S240: The multi-level enhancement result is generated by the following formula ， in, They correspond to five different levels of adaptive enhancement coefficients, from low to high.

10. The low-light image enhancement method based on information fusion strategy according to claim 1, characterized in that: The step S300 includes: Step S310: Acquire size parameters of the first and second groups of multi-exposure image sequences, and select a number of pyramid layers suitable for the current image size according to the input parameters; Step S320: normalize the weights of each group of input images at the same pixel position so that the sum of the weight contributions of all input images is 1; Step S330: construct a Gaussian pyramid and a Laplacian pyramid for each set of input images, and perform weighted fusion on multiple input images at each pyramid level; Step S340: Restore the fused multi-scale pyramid to the final enhanced image through an inverse operation.

11. The low-light image enhancement method based on information fusion strategy according to claim 10, characterized in that: The step S310 includes: Get the size parameters of the first and second multi-exposure image sequences, where [H, W, C, N] represent the image height, width, number of channels, and number of input images, respectively. By checking the input parameter lev, select a pyramid level suitable for the current image size ， The pyramid is decomposed at most to a scale logarithmic of the minimum dimension of the input image.

12. The low-light image enhancement method based on information fusion strategy according to claim 10, characterized in that: The step S320 includes: Normalize the weight of each input image so that the sum of the weights of all input images at the same pixel position is 1, making the fused weight smooth and distortion-free. Suppose the size of the input image is , weight Indicates that the kth image is at position The initial weight distribution is , the normalization step is expressed as 。 13. The low-light image enhancement method based on information fusion strategy according to claim 10, characterized in that: The step S330 includes: The Gaussian pyramid generates images of different resolutions through multiple Gaussian filtering and downsampling. The Laplacian pyramid is used to store the differences between images of different scales. ， For each input image , calculate its Gaussian pyramid and Laplace pyramid After pyramid decomposition, the images at each level are fused by multi-scale weighted fusion. Layer, the fused image is ， in, The nth image is in the The weight of the layer; The fused image is restored using the Laplacian pyramid reconstruction process.

14. The low-light image enhancement method based on information fusion strategy according to claim 10, characterized in that: The step S340 includes: The final fused image is obtained by reconstructing the weighted fused images of all levels ， in, represents the inverse operation of the Gaussian pyramid, is the fused image layer, and the final fused image Preserve dominant regions from multiple input images.

Images