Image turbulence removal method and system based on feature fusion and adaptive filtering
Through the method of feature fusion and adaptive filtering, combined with local block division, multi-scale image fusion and low-rank decomposition, the problem of image deturbulence in the existing technology under complex meteorological conditions is solved, and the high-quality deturbulence effect is achieved. In various application scenarios.
Patent Information
- Application Number
- CN202411297040.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-18
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2044-09-18
AI Technical Summary
Existing image deturbulence methods are difficult to effectively remove image blur and distortion under complex meteorological conditions, especially in different scenarios and meteorological conditions, with limited generalization capabilities, high computing costs and poor real-time performance, making it difficult to use on resource-constrained devices. deploy.
Using image deturbulence method based on feature fusion and adaptive filtering, image noise is gradually removed to obtain high-quality de-removal through local block division, multi-scale image fusion, low rank and sparse decomposition, and adaptive weight adjustment based on gradient amplitude. Turbulence image.
Effectively eliminate image distortion caused by turbulence, improve image quality and clarity, is suitable for various complex meteorological conditions, is highly robust and adaptable, and is suitable for different application scenarios.
Smart Images

Figure CN118822895B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image turbulence removal, and in particular to an image turbulence removal method and system based on feature fusion and adaptive filtering. Background Art
[0002] Under complex meteorological conditions, image turbulence is caused by unstable flows in the atmosphere. This phenomenon can cause image distortion such as blurring, distortion, and jitter, which seriously affects image quality and further affects downstream visual analysis tasks such as target detection, tracking, and recognition. In practical application scenarios such as aerial photography, remote sensing, and ground monitoring, de-turbulence processing has become a key step in improving image quality and subsequent analysis accuracy.
[0003] At present, image de-turbulence methods are mainly divided into two categories: physical model-based methods and data-driven methods. Physical model-based methods attempt to restore images by simulating the physical process of atmospheric turbulence effects. Such methods usually require accurate modeling of atmospheric conditions, which is computationally complex and poorly adaptable to uncertainties in practical applications. In addition, physical models rely on a large amount of prior knowledge and are difficult to process in real time for dynamically changing meteorological conditions.
[0004] Data-driven methods, especially the deep learning methods that have become popular in recent years, build complex neural network models to directly learn the mapping relationship of turbulence removal from data. Such methods rely on a large amount of labeled data for training, and have limited generalization capabilities in different scenarios and meteorological conditions. At the same time, deep learning methods have high computational costs, poor real-time performance, and are difficult to deploy on resource-constrained devices. Summary of the invention
[0005] In view of this, the present invention provides an image de-turbulence method based on feature fusion and adaptive filtering, aiming at providing high-quality de-turbulence images under different scenes and meteorological conditions.
[0006] To achieve the above object, the present invention provides an image turbulence removal method based on feature fusion and adaptive filtering, comprising the following steps:
[0007] S1: Divide the acquired image into local blocks, calculate the mean and variance of each local block, then calculate the weight of the adaptive filter and perform weighted average filtering to obtain the denoised image;
[0008] S2: extracting a multi-scale image of the denoised image, normalizing the multi-scale image, and obtaining a normalized multi-scale image;
[0009] S3: Calculate the local variance of the normalized multi-scale image, calculate the weight based on the local variance and fuse the multi-scale images to obtain a fused image;
[0010] S4: Perform low-rank and sparse decomposition on the fused image, construct and optimize the low-rank and sparse decomposition optimization target, and obtain the optimized low-rank image;
[0011] S5: Use adaptive weights based on gradient amplitude to enhance the details of the optimized low-rank image and obtain a de-turbulent image.
[0012] As a further improvement method of the present invention:
[0013] Optionally, in step S1, the acquired image is divided into local blocks, the mean and variance of each local block are calculated, and then the weight of the adaptive filter is calculated and weighted average filtering is performed to obtain a denoised image, including:
[0014] S11: Divide local blocks:
[0015] The acquired image Divide into multiple non-overlapping local blocks, each local block The size is ,in For the The index of a local block, the local block Specifically:
[0016] ;
[0017] in, is the pixel position; To obtain the image At pixel position The pixel value at ; , , To round down, For local blocks The pixel position of the upper left pixel within;
[0018] S12: Calculate the mean and variance of the local block:
[0019] For each local block , calculate the local block The mean and variance , specifically:
[0020] ;
[0021] ;
[0022] in, For local blocks The set of all pixel positions within ; is included in The pixel position in ; To obtain the image At pixel position The pixel value at ;
[0023] S13: Calculate the weight of the adaptive filter and perform weighted average filtering to obtain a denoised image.
[0024] S131: Calculate adaptive filter weights:
[0025] Calculate the acquired image Variance , and then calculate each local block The weights of the adaptive filter , specifically:
[0026] ;
[0027] in, is a natural constant; For local blocks The adaptive filter is at pixel position The weight value at ;
[0028] S132: Perform weighted average filtering:
[0029] For each local block The pixel positions within are weighted average filtered, specifically:
[0030] ;
[0031] in, For local blocks The result of weighted average filtering At pixel position The pixel value at ;
[0032] Calculate each local block The pixel positions within the image are weighted averaged and filtered to obtain the denoised image. .
[0033] Optionally, extracting a multi-scale image of the denoised image in step S2, normalizing the multi-scale image to obtain a normalized multi-scale image includes:
[0034] S21: Constructing multi-scale images:
[0035] For the denoised image Use multi-scale Gaussian filtering to generate Layer multi-scale image , specifically:
[0036] ;
[0037] ;
[0038] in, For the Layer multi-scale image At pixel position The pixel value at , is the total number of layers of multi-scale images; For Centered The set of all pixel positions within the range; is included in The pixel position in ; The denoised image At pixel position The pixel value at ; To generate the The Gaussian kernel used in the multi-scale image layer is The weight value at ; To generate the The scale parameter of the Gaussian kernel used for the multi-scale image layer; is pi;
[0039] S22: Perform multi-scale image normalization:
[0040] The first Layer multi-scale image Normalized to interval, obtain the normalized Layer multi-scale image , the normalization is maximum and minimum value normalization.
[0041] Optionally, in step S3, calculating the local variance of the normalized multi-scale image, calculating the weight based on the local variance and fusing the multi-scale images to obtain a fused image includes:
[0042] Calculate the normalized Layer multi-scale image The local variance at each pixel position , specifically:
[0043] ;
[0044] in, For normalization Layer multi-scale image At pixel position The pixel value at ; For normalization Layer multi-scale image The mean of
[0045] Calculate the Layer multi-scale image Weight , specifically:
[0046] ;
[0047] in, For the Layer multi-scale image Weight At pixel position The weight of For normalization Layer multi-scale image The local variance at each pixel location, ; ;
[0048] Fuse different layers of multi-scale images to obtain the fused image , specifically:
[0049] ;
[0050] in, The fused image At pixel position The pixel value at .
[0051] Optionally, in step S4, the fused image is subjected to low-rank and sparse decomposition, a low-rank and sparse decomposition optimization target is constructed and optimized, and an optimized low-rank image is obtained, including:
[0052] The fused image Treated as a two-dimensional matrix, the fused image is decomposed using low rank and sparse decomposition Decompose into a low-rank matrix With sparse matrices ;
[0053] Construct low-rank and sparse decomposition optimization objectives, specifically:
[0054] ;
[0055] in, is the nuclear norm of the low-rank matrix, is a sparse matrix 1-norm of ; is the balance parameter;
[0056] Use the alternating direction multiplier method to solve the optimization objective and obtain the optimized low-rank image .
[0057] Optionally, in step S5, using adaptive weights based on gradient amplitude to enhance details of the optimized low-rank image to obtain a de-turbulent image includes:
[0058] S51: Constructing detail enhancement filter:
[0059] Using the Laplacian operator For the optimized low-rank image Perform convolution to extract detailed information image , specifically:
[0060] ;
[0061] in, For detailed information image At pixel position The pixel value at ; is the convolution operation;
[0062] S52: Calculate the optimized low-rank image Adaptive weights for the gradient magnitude ;
[0063] S521: Calculate the gradient amplitude:
[0064] Use the Sobel operator to calculate the optimized low-rank image Gradients in horizontal and vertical directions:
[0065] ;
[0066] ;
[0067] in, and They are the horizontal gradients and the vertical gradient At pixel position The amplitude at ; and They are the Sobel operators in the horizontal and vertical directions respectively;
[0068] Calculate the gradient magnitude :
[0069] ;
[0070] in, is the gradient amplitude At pixel position The amplitude at ;
[0071] S522: Calculate adaptive weight:
[0072] Based on the gradient magnitude Calculating adaptive weights , specifically:
[0073] ;
[0074] in, is the steepness parameter; is the threshold parameter.
[0075] S53: Improving the details of the optimized low-rank image:
[0076] Using adaptive weights Improve the details of the optimized low-rank image and obtain the de-turbulenced image :
[0077] ;
[0078] in, This is the image after removing turbulence. At pixel position The pixel value at ; is the optimized low-rank image at pixel position The pixel value at ; is the adaptive weight At pixel position The weight value at .
[0079] The present invention also discloses an image turbulence removal system based on feature fusion and adaptive filtering, comprising:
[0080] Denoising module: divide the acquired image into local blocks, calculate the mean and variance of each local block, then calculate the weight of the adaptive filter and perform weighted average filtering;
[0081] Normalization module: extract the multi-scale image of the denoised image and normalize the multi-scale image;
[0082] Fusion module: calculates the local variance of the normalized multi-scale image, calculates the weight based on the local variance and fuses the multi-scale image;
[0083] Low-rank and sparse decomposition module: performs low-rank and sparse decomposition on the fused image, constructs low-rank and sparse decomposition optimization targets and performs optimization;
[0084] Enhancement module: Use adaptive weights based on gradient magnitude to enhance the details of the optimized low-rank image;
[0085] Beneficial Effects
[0086] The present invention can effectively eliminate the image distortion problem caused by turbulence by combining multi-scale image fusion with adaptive filtering. Multi-scale image fusion can retain image details and edge information of different scales, thereby enhancing the clarity and structural information of the image. At the same time, adaptive filtering can adaptively adjust the weight of the filter according to local image features, effectively removing noise while retaining image details. This combined method can significantly improve image quality while removing turbulence, making the removed image closer to the real scene, and is suitable for various complex meteorological conditions.
[0087] The adaptive filtering and gradient-based adaptive weight adjustment strategy adopted in the present invention enable the algorithm to adaptively adjust the parameters of the de-turbulence processing according to different image features and turbulence intensity. This adaptability enables the algorithm to perform better in dealing with various dynamically changing meteorological conditions and has higher robustness. Whether it is a slight or severe turbulence effect, the method of the present invention can be adaptively adjusted to ensure that the ideal de-turbulence effect can be obtained in different application scenarios, effectively expanding the application scope of the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS
[0088] Figure 1 A schematic diagram of a process of an image turbulence removal method based on feature fusion and adaptive filtering according to an embodiment of the present invention;
[0089] Figure 2 Schematic diagram of the turbulence removal result according to an embodiment of the present invention, wherein FIG (a) is an image affected by turbulence, and FIG (b) is an image after turbulence removal. DETAILED DESCRIPTION
[0090] The present invention is further described below in conjunction with the accompanying drawings, but the present invention is not limited in any way. Any changes or substitutions made based on the teachings of the present invention belong to the protection scope of the present invention.
[0091] Embodiment 1: An image de-turbulence method based on feature fusion and adaptive filtering, such as Figure 1 As shown, the following steps are included:
[0092] S1: Divide the acquired image into local blocks, calculate the mean and variance of each local block, then calculate the weight of the adaptive filter and perform weighted average filtering to obtain the denoised image:
[0093] S11: Divide local blocks:
[0094] The acquired image Divide into multiple non-overlapping local blocks, each local block The size is ,in For the The index of a local block, the local block Specifically:
[0095] ;
[0096] in, is the pixel position; To obtain the image At pixel position The pixel value at ; , , To round down, For local blocks The pixel position of the upper left pixel within;
[0097] S12: Calculate the mean and variance of the local block:
[0098] For each local block , calculate the local block The mean and variance , specifically:
[0099] ;
[0100] ;
[0101] in, For local blocks The set of all pixel positions within ; is included in The pixel position in ; To obtain the image At pixel position The pixel value at ;
[0102] S13: Calculate the weight of the adaptive filter and perform weighted average filtering to obtain a denoised image.
[0103] S131: Calculate adaptive filter weights:
[0104] Calculate the acquired image Variance , and then calculate each local block The weights of the adaptive filter , specifically:
[0105] ;
[0106] in, is a natural constant; For local blocks The adaptive filter is at pixel position The weight value at ;
[0107] S132: Perform weighted average filtering:
[0108] For each local block The pixel positions within are weighted average filtered, specifically:
[0109] ;
[0110] in, For local blocks The result of weighted average filtering At pixel position The pixel value at ;
[0111] Calculate each local block The pixel positions within the image are weighted averaged and filtered to obtain the denoised image. .
[0112] By dividing the image into local blocks, this step can be processed adaptively for different image regions. Since each local block may contain different texture and noise characteristics, the locally calculated mean and variance enable the weights of the adaptive filter to be dynamically adjusted to adapt to different noise levels and image characteristics.
[0113] The weight calculation of the adaptive filter not only depends on the global variance of the image, but also combines the mean and variance of the local block. This dual standard enables the algorithm to better preserve the details and edges of the local image while removing noise. Especially in the case of uneven noise intensity, adaptive filtering can avoid over-smoothing and effectively protect the details of the image.
[0114] S2: extracting a multi-scale image of the denoised image, normalizing the multi-scale image, and obtaining a normalized multi-scale image;
[0115] S21: Constructing multi-scale images:
[0116] For the denoised image Use multi-scale Gaussian filtering to generate Layer multi-scale image , specifically:
[0117] ;
[0118] ;
[0119] in, For the Layer multi-scale image At pixel position The pixel value at , is the total number of layers of multi-scale images; For Centered The set of all pixel positions within the range; is included in The pixel position in ; The denoised image At pixel position The pixel value at ; To generate the The Gaussian kernel used in the multi-scale image layer is The weight value at ; To generate the The scale parameter of the Gaussian kernel used in the multi-scale image layer. , , , ; is pi;
[0120] S22: Perform multi-scale image normalization:
[0121] The first Layer multi-scale image Normalized to interval, obtain the normalized Layer multi-scale image , specifically:
[0122] ;
[0123] in, For normalization Layer multi-scale image At pixel position The pixel value at ; and They are Layer multi-scale image The minimum and maximum values of
[0124] By using multi-scale Gaussian filtering, it is possible to extract features of images at different scales. This method can capture structural information of different sizes in the image, including fine textures and larger regional features. Images of different scales help to identify and analyze various features in the image, from details to the whole.
[0125] The scale parameter in the Gaussian filtering process can adjust the smoothness of the filter, which effectively enhances the contrast of image features at different scales. By performing multi-scale processing on the image, features at different levels can be highlighted, making important information in the image more obvious at a specific scale.
[0126] S3: Calculate the local variance of the normalized multi-scale image, calculate the weight based on the local variance and fuse the multi-scale images to obtain the fused image:
[0127] Calculate the normalized Layer multi-scale image The local variance at each pixel position , specifically:
[0128] ;
[0129] in, For normalization Layer multi-scale image At pixel position The pixel value at ; For normalization Layer multi-scale image The mean of
[0130] Calculate the Layer multi-scale image Weight , specifically:
[0131] ;
[0132] in, For the Layer multi-scale image Weight At pixel position The weight of For normalization Layer multi-scale image The local variance at each pixel location, ; ;
[0133] Fuse different layers of multi-scale images to obtain the fused image , specifically:
[0134] ;
[0135] in, The fused image At pixel position The pixel value at .
[0136] By calculating the local variance of the normalized multi-scale images and calculating the weights based on the variance, the important information in the images of different scales can be effectively combined. The local variance reflects the texture details and contrast of the image in a specific area. Using these variances to weight the fusion of images of different scales can enhance the significant features and details in the image, thereby improving the quality and clarity of the final fused image.
[0137] S4: Perform low-rank and sparse decomposition on the fused image, construct and optimize the low-rank and sparse decomposition optimization objectives, and obtain the optimized low-rank image:
[0138] The fused image Treated as a two-dimensional matrix, the fused image is decomposed using low rank and sparse decomposition Decompose into a low-rank matrix With sparse matrices , specifically:
[0139] ;
[0140] Among them, the low-rank matrix With sparse matrices The matrix size and fused image Consistency;
[0141] Construct low-rank and sparse decomposition optimization objectives, specifically:
[0142] ;
[0143] in, is the nuclear norm of the low-rank matrix, is a sparse matrix 1-norm of ; is a balance parameter. ;
[0144] Use the alternating direction multiplier method to solve the optimization objective and obtain the optimized low-rank image .
[0145] Low-rank and sparse decomposition can effectively separate the main structure in the image from noise and other interference. The low-rank matrix represents the main structure and background information in the image, while the sparse matrix captures the noise and irregular interference in the image. Through this decomposition, a clearer and less noisy low-rank image can be obtained.
[0146] S5: Use adaptive weights based on gradient magnitude to enhance the details of the optimized low-rank image and obtain a de-turbulent image:
[0147] S51: Constructing detail enhancement filter:
[0148] Using the Laplacian operator For the optimized low-rank image Perform convolution to extract detailed information image , specifically:
[0149] ;
[0150] ;
[0151] in, For detailed information image At pixel position The pixel value at ; is the convolution operation;
[0152] S52: Calculate the optimized low-rank image Adaptive weights for the gradient magnitude ;
[0153] S521: Calculate the gradient amplitude:
[0154] Use the Sobel operator to calculate the optimized low-rank image Gradients in horizontal and vertical directions:
[0155] ;
[0156] ;
[0157] in, and They are the horizontal gradients and the vertical gradient At pixel position The amplitude at ; and They are the Sobel operators in the horizontal and vertical directions, specifically:
[0158] ;
[0159] Calculate the gradient magnitude :
[0160] ;
[0161] in, is the gradient amplitude At pixel position The amplitude at ;
[0162] S522: Calculate adaptive weight:
[0163] Based on the gradient magnitude Calculating adaptive weights , specifically:
[0164] ;
[0165] in, is the adaptive weight At pixel position The weight value at ; is a steep parameter. In this embodiment, ; is a threshold parameter. In this embodiment, ;
[0166] S53: Improving the details of the optimized low-rank image:
[0167] Using adaptive weights Improve the details of the optimized low-rank image and obtain the de-turbulenced image ,like Figure 2 As shown:
[0168] ;
[0169] in, This is the image after removing turbulence. At pixel position The pixel value at ; is the optimized low-rank image at pixel position The pixel value at ;
[0170] By using gradient-based adaptive weights to enhance the details of the optimized low-rank image, the details and clarity of the image can be significantly enhanced. The detail enhancement filter extracts the detail information in the image, and combined with the adaptive weights, these detail information can be accurately enhanced, thereby improving the visual effect and detail performance of the image.
[0171] Embodiment 2: The present invention also discloses an image turbulence removal system based on feature fusion and adaptive filtering, comprising the following five modules:
[0172] Denoising module: divide the acquired image into local blocks, calculate the mean and variance of each local block, then calculate the weight of the adaptive filter and perform weighted average filtering;
[0173] Normalization module: extract the multi-scale image of the denoised image and normalize the multi-scale image;
[0174] Fusion module: calculates the local variance of the normalized multi-scale image, calculates the weight based on the local variance and fuses the multi-scale image;
[0175] Low-rank and sparse decomposition module: performs low-rank and sparse decomposition on the fused image, constructs low-rank and sparse decomposition optimization targets and performs optimization;
[0176] Enhancement module: Use adaptive weights based on gradient magnitude to enhance the details of the optimized low-rank image.
[0177] It should be noted that the serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages and disadvantages of the embodiments. And the terms "including", "comprising" or any other variants thereof in this article are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "including a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.
[0178] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention.
[0179] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.
Claims
1. An image turbulence removal method based on feature fusion and adaptive filtering, characterized in that: The following steps are involved: S1: Divide the acquired image into local blocks, calculate the mean and variance of each local block, then calculate the weight of the adaptive filter and perform weighted average filtering to obtain the denoised image; S2: extracting a multi-scale image of the denoised image, normalizing the multi-scale image, and obtaining a normalized multi-scale image; S3: Calculate the local variance of the normalized multi-scale image, calculate the weight based on the local variance and fuse the multi-scale images to obtain a fused image; S4: Perform low-rank and sparse decomposition on the fused image, construct and optimize the low-rank and sparse decomposition optimization target, and obtain the optimized low-rank image; S5: Use adaptive weights based on gradient amplitude to enhance the details of the optimized low-rank image and obtain a de-turbulent image; The step S5 includes the following steps: S51: Constructing detail enhancement filter: Using the Laplacian operator For the optimized low-rank image Perform convolution to extract detailed information image , specifically: ; in, For detailed information image At pixel position The pixel value at ; is the convolution operation; S52: Calculate the optimized low-rank image Adaptive weights for the gradient magnitude ; The step S52 includes the following steps: S521: Calculate the gradient amplitude: Use the Sobel operator to calculate the optimized low-rank image Gradients in horizontal and vertical directions: ; ; in, and They are the horizontal gradients and the vertical gradient At pixel position The amplitude at ; and They are the Sobel operators in the horizontal and vertical directions respectively; Calculate the gradient magnitude : ; in, is the gradient amplitude At pixel position The amplitude at ; S522: Calculate adaptive weight: Based on the gradient magnitude Calculating adaptive weights , specifically: ; in, is the steepness parameter; is the threshold parameter; S53: Improving the details of the optimized low-rank image: Using adaptive weights Improve the details of the optimized low-rank image and obtain the de-turbulenced image : ; in, This is the image after removing turbulence. At pixel position The pixel value at ; is the optimized low-rank image at pixel position The pixel value at ; is the adaptive weight At pixel position The weight value at .
2. The image turbulence removal method based on feature fusion and adaptive filtering according to claim 1 is characterized in that: The step S1 includes the following steps: S11: Divide local blocks: The acquired image Divide into multiple non-overlapping local blocks, each local block The size is ,in For the The index of a local block, the local block Specifically: ; in, is the pixel position; To obtain the image At pixel position The pixel value at ; , , To round down, For local blocks The pixel position of the upper left pixel within; S12: Calculate the mean and variance of the local block: For each local block , calculate the local block The mean and variance , specifically: ; ; in, For local blocks The set of all pixel positions within ; is included in The pixel position in ; To obtain the image At pixel position The pixel value at ; S13: Calculate the weight of the adaptive filter and perform weighted average filtering to obtain a denoised image.
3. The image turbulence removal method based on feature fusion and adaptive filtering according to claim 2 is characterized in that: The step S13 includes the following steps: S131: Calculate adaptive filter weights: Calculate the acquired image Variance , then calculate each local block The weights of the adaptive filter , specifically: ; in, is a natural constant; For local blocks The adaptive filter is at pixel position The weight value at ; S132: Perform weighted average filtering: For each local block The pixel positions within are weighted average filtered, specifically: ; in, For local blocks The result of weighted average filtering At pixel position The pixel value at ; Calculate each local block The pixel positions within the image are weighted averaged and filtered to obtain the denoised image. .
4. The image turbulence removal method based on feature fusion and adaptive filtering according to claim 2 is characterized in that: The step S2 includes the following steps: S21: Constructing multi-scale images: For the denoised image Use multi-scale Gaussian filtering to generate Layer multi-scale image , specifically: ; ; in, For the Layer multi-scale image At pixel position The pixel value at , is the total number of layers of multi-scale images; For Centered The set of all pixel positions within the range; is included in The pixel position in ; The denoised image At pixel position The pixel value at ; To generate the The Gaussian kernel used in the multi-scale image layer is The weight value at ; To generate the The scale parameter of the Gaussian kernel used for the multi-scale image layer; is the ratio of pi; S22: Perform multi-scale image normalization: The first Layer multi-scale image Normalized to interval, obtain the normalized Layer multi-scale image , the normalization is maximum and minimum value normalization.
5. The image turbulence removal method based on feature fusion and adaptive filtering according to claim 4 is characterized in that: The step S3 includes the following steps: Calculate the normalized Layer multi-scale image The local variance at each pixel position , specifically: ; in, For normalization Layer multi-scale image At pixel position The pixel value at ; For normalization Layer multi-scale image The mean of Calculate the Layer multi-scale image Weight , specifically: ; in, For the Layer multi-scale image Weight At pixel position The weight of For normalization Layer multi-scale image The local variance at each pixel location, ; ; Fuse different layers of multi-scale images to obtain the fused image , specifically: ; in, The fused image At pixel position The pixel value at .
6. The image turbulence removal method based on feature fusion and adaptive filtering according to claim 5 is characterized in that: The step S4 includes the following steps: The fused image Treated as a two-dimensional matrix, the fused image is decomposed using low rank and sparse decomposition Decompose into a low-rank matrix With sparse matrices ; Construct low-rank and sparse decomposition optimization objectives, specifically: ; in, is the nuclear norm of the low-rank matrix, is a sparse matrix 1-norm of ; is the balance parameter; Use the alternating direction multiplier method to solve the optimization objective and obtain the optimized low-rank image .
7. An image turbulence removal system based on feature fusion and adaptive filtering, characterized in that: include: Denoising module: divide the acquired image into local blocks, calculate the mean and variance of each local block, then calculate the weight of the adaptive filter and perform weighted average filtering; Normalization module: extract the multi-scale image of the denoised image and normalize the multi-scale image; Fusion module: calculates the local variance of the normalized multi-scale image, calculates the weight based on the local variance and fuses the multi-scale image; Low-rank and sparse decomposition module: performs low-rank and sparse decomposition on the fused image, constructs low-rank and sparse decomposition optimization targets and performs optimization; Enhancement module: Use adaptive weights based on gradient magnitude to enhance the details of the optimized low-rank image; To realize an image de-turbulence method based on feature fusion and adaptive filtering as described in any one of claims 1-6.
