Image denoising and inpainting method based on adaptive kernel estimation
By adopting an adaptive kernel field generation mechanism and an iterative filling strategy, the problem of kernel scale adaptation in image processing is solved, achieving high-quality image denoising and restoration under multi-scale characteristics, thus improving image quality and reliability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHANGSHU INSTITUTE OF TECHNOLOGY
- Filing Date
- 2026-02-15
- Publication Date
- 2026-06-02
Smart Images
Figure CN122134584A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer and image filtering technology, specifically relating to an image denoising and restoration method based on adaptive kernel estimation. Background Technology
[0002] Image denoising and restoration are crucial tasks in digital image processing, widely applied in fields such as medical imaging, remote sensing, security monitoring, and consumer electronics. The goal is to suppress noise interference while preserving or restoring as much structural detail and texture information as possible. Traditional methods are mostly based on local statistical models or filter designs. Techniques such as nonlocal mean (NLM) and bilateral filtering, by introducing spatial or gray-level similarity constraints, achieve a balance between noise suppression and edge preservation to some extent. However, these methods rely on preset or globally uniform kernel function scales, making it difficult to handle the multi-scale characteristics of image content differences between local regions.
[0003] Image restoration methods based on kernel estimation reconstruct signals by constructing pixel-centered weighted neighborhood models, and their performance is highly dependent on the choice of the kernel function's support domain. High-frequency, detail-rich regions require small-scale kernels to avoid structural blurring caused by excessive smoothing, while low-frequency, flat regions can use large-scale kernels to improve noise suppression efficiency. Existing techniques typically employ fixed-scale or coarse-grained partitioning strategies based solely on simple gradient thresholds, failing to provide fine-grained perception and response to local image "activity."
[0004] In existing technologies, static kernel scale configuration or single-feature-based partitioning mechanisms inevitably lead to oversmoothing or undersmoothing in complex scenes: the former causes decreased edge sharpness and loss of subtle textures, while the latter leaves behind significant noise interference, reducing visual quality and the reliability of subsequent analysis. This problem is more pronounced in areas with high noise intensity or dense structures. Although some methods attempt to introduce adaptive mechanisms, the lack of systematic modeling of the multi-resolution characteristics of images makes it difficult to achieve a dynamic optimal trade-off between global noise suppression and local detail preservation. An adaptive image denoising and restoration method that can deeply integrate multi-scale signal analysis and statistical kernel estimation is needed to address the rigidity of scale selection in existing technologies. Summary of the Invention
[0005] This invention provides an image denoising and restoration method based on adaptive kernel estimation, aiming to solve the over-smoothing or under-smoothing problems caused by the inability of existing image processing techniques to effectively adapt to the multi-scale characteristics of image content due to the use of fixed or simply partitioned kernel scales. This method constructs a locally structure-aware adaptive kernel field generation mechanism, combining multi-scale gradient response analysis and non-local similarity constraints, to dynamically determine the optimal kernel function support domain and shape parameters at each pixel. This effectively suppresses noise interference while preserving image detail structure and performs high-fidelity restoration of damaged areas.
[0006] This invention provides an image denoising and restoration method based on adaptive kernel estimation, comprising: Acquire raw image data that is noisy or partially missing; Multi-scale gradient magnitude calculation is performed on the original image data to generate a gradient response map characterizing the local structural complexity of the image; Based on the gradient response map, a local directional covariance matrix is constructed at each pixel location, and the local principal direction and degree of anisotropy are determined according to the eigenvalue decomposition results of the local directional covariance matrix. Using the local principal direction and the degree of anisotropy, a spatially varying elliptical kernel function is constructed, with the major axis extending along the local edge direction and the minor axis contracting perpendicular to the local edge direction, and the ratio of the major axis to the minor axis dynamically adjusted by the local gradient intensity. The elliptical kernel function is used as a weighted window to perform weighted average filtering on the pixels in the neighborhood. At the same time, a non-local mean prior constraint is introduced to search for image patches with similar local structural features from the entire image range and include them in the reconstruction weight calculation of the current pixel. For missing or degraded regions in the image, the repair boundary is expanded outward layer by layer based on the iterative filling strategy guided by the elliptical kernel function. Each iteration updates the kernel function parameters according to the structural information of the repaired region until the entire missing region is completely reconstructed. Output a high-quality image after noise reduction and restoration.
[0007] Preferably, performing multi-scale gradient magnitude calculation on the original image data to generate a gradient response map characterizing the local structural complexity of the image includes: Gaussian pyramid downsampling is performed on the original image to obtain sub-images at multiple scale levels; At each scale level, calculate the first-order partial derivatives in the horizontal and vertical directions respectively; The first-order partial derivatives at each scale are combined into a multi-channel gradient tensor, and a unified gradient response intensity map is obtained by normalizing with the L2 norm. Gradient response intensity maps at different scales are upsampled to the original resolution and then weighted and fused to form the final multi-scale gradient response map. The weight coefficients of each scale are dynamically allocated according to their signal-to-noise ratio estimates.
[0008] Preferably, based on the gradient response map, a local directional covariance matrix is constructed at each pixel location, including: A square neighborhood window of fixed size is selected with the current pixel as the center; Within this square neighborhood window, extract the gradient vectors of all pixels; Calculate the weighted covariance matrix of these gradient vectors relative to the center pixel, where the weights are determined by the corresponding values in the multi-scale gradient response map; The weighted covariance matrix is decomposed into eigenvalues to obtain two orthogonal eigenvectors and their corresponding eigenvalues. The eigenvector corresponding to the eigenvalue with the larger value among two eigenvalues is defined as the local principal direction, and the ratio of the two eigenvalues is used to quantify the degree of anisotropy of the local structure.
[0009] Preferably, constructing a spatially varying elliptical kernel function using the local principal directions and the degree of anisotropy includes: Elliptic kernel function The mathematical expression is: ; Indicates the current pixel coordinates. Represents the coordinates of any pixel within the neighborhood. For The local covariance matrix centered at the center is constructed as follows: ; This is a rotation matrix, where the elements are determined by the local principal direction angle. Decide, It is calculated from the direction angle of the eigenvector; and Let these represent the standard deviations along the principal direction and the perpendicular direction, respectively. They satisfy the following relationship: , and Let be two eigenvalues of the covariance matrix, and .
[0010] Preferably, the elliptical kernel function is used as a weighted window to perform weighted average filtering on pixels in the neighborhood, while introducing a non-local mean prior constraint, including: Calculate the structural similarity measure between the image patch containing the current pixel and other candidate image patches in the entire image. The structural similarity measure is based on the normalized cross-correlation coefficient and combined with the gradient direction consistency correction term. The center pixel of the corresponding candidate block is included in the reconstruction set of the current pixel only when the structural similarity metric is greater than a preset threshold. The reconstruction weights are determined by the product of the weight value of the elliptical kernel function at the center pixel position and the structural similarity measure, and are then normalized and used for weighted summation.
[0011] Preferably, the normalized cross-correlation coefficient is obtained by calculating the dot product of the two image blocks after removing the mean and dividing it by the product of their respective standard deviations; the gradient direction consistency correction term is obtained by calculating the mean cosine similarity of the gradient directions of corresponding pixels in the two image blocks.
[0012] Preferably, for missing or degraded regions in the image, the iterative filling strategy guided by the elliptical kernel function includes: Identify the boundaries of missing regions in an image. A boundary pixel is a point in the missing region that has at least one eight-neighbor pixel that is a valid pixel. Initialize the interpolation results where the pixel value inside the missing region is zero or the value of the nearest valid pixel; In each iteration, only the outermost boundary pixels of the current missing region are processed; For each boundary pixel to be repaired, the local covariance matrix and the corresponding elliptical kernel function are recalculated based on the structural information of the surrounding known pixels. The neighborhood window contains only valid pixels. The effective pixel is reconstructed using the elliptical kernel function and non-local similarity block information. The search range of the non-local similarity block is limited to the repaired area and the original effective area. Update the state of the valid pixel to "repaired" and include it in the valid reference set for the next iteration; Repeat all steps until all missing pixels are repaired.
[0013] Preferably, the missing region is identified by a pixel mask, where a mask value of zero indicates a valid pixel and a non-zero value indicates a missing pixel; the pixel mask is updated synchronously after each iteration.
[0014] Preferably, the global noise level is estimated by subtracting the structural variance from the global variance of the image; if the calculated... If the noise level exceeds the threshold corresponding to the global noise level, it will be truncated to that threshold.
[0015] Preferably, the Gaussian pyramid is constructed by convolving the current-scale image with a standard Gaussian low-pass filter and then downsampling it by half its size; the first-order partial derivatives are calculated using the Sobel operator or the Scharr operator.
[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention solves the inherent problems of traditional fixed kernels or simple partitioned kernels when processing images with complex multi-scale structures by constructing an adaptive kernel field generation mechanism based on local structure awareness.
[0017] 2. This method can dynamically adjust the spatial support domain and shape of the kernel function according to the local gradient response of the image, so that a larger kernel scale is used in smooth areas to suppress noise, while the kernel scale is automatically shrunk and extended along the structural direction in edge or textured areas, thereby avoiding loss of detail and artifact generation.
[0018] 3. By fusing nonlocal similarity priors, the ability to preserve repetitive structures and textures is enhanced, thereby improving the overall quality of noise reduction and restoration.
[0019] 4. For missing areas in an image, the iterative repair strategy proposed in this invention can gradually restore details under structural guidance, ensuring the visual coherence and geometric consistency of the repair results. Attached Figure Description
[0020] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the adaptive nuclear field generation mechanism with local structure awareness in this invention; Figure 3 This is a flowchart illustrating the logical process of generating multi-scale gradient response maps and constructing local covariance matrices in this invention. Figure 4 This is a flowchart illustrating the logical flow of the adaptive elliptic kernel function construction and weighted filtering fusion of nonlocal similarity constraints in this invention. Figure 5 This is a flowchart illustrating the logical flow of the missing region iterative repair strategy in this invention. Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between the terminal and the cloud in this invention. Detailed Implementation
[0021] refer to Figures 1 to 6 This invention provides an image denoising and restoration method based on adaptive kernel estimation. It dynamically generates spatially varying elliptical kernel functions through a local structure-aware mechanism and combines nonlocal similarity constraints and an iterative filling strategy to achieve high-quality reconstruction of noisy or partially missing images. The following detailed explanation of the specific implementation steps of this method ensures full disclosure and feasibility of the technical solution.
[0022] The method first performs the acquisition of raw image data. The raw image data is a two-dimensional digital image to be processed, with its pixel values stored in a memory buffer in grayscale or color channel form. This two-dimensional digital image may contain random noise points caused by sensor noise, transmission interference, or physical damage, and may also contain continuous missing regions due to occlusion, scratches, or data loss. After the raw image data is completely loaded into the processing unit, it serves as the input basis for all subsequent calculation steps.
[0023] After acquiring the original image data, multi-scale gradient magnitude calculation is performed to generate a gradient response map characterizing the local structural complexity of the image. This step specifically includes: sequentially performing Gaussian pyramid downsampling on the original image to generate sub-images at multiple scale levels. The Gaussian pyramid is constructed by convolving the image at the current scale using a standard Gaussian low-pass filter, followed by downsampling to half its original size, repeating this process until the preset maximum number of scale levels is reached. Each sub-image at each scale level retains the smooth structural information of the original image at that maximum scale.
[0024] At each scale level, the first-order partial derivatives in the horizontal and vertical directions are calculated. The calculation of the first-order partial derivatives is implemented using the Sobel or Scharr operator, with a 3×3 matrix in the horizontal direction (±1 in the center column and 0 elsewhere) and a transposed matrix in the vertical direction. This yields the horizontal and vertical gradient components at each scale level. The first-order partial derivatives at each scale are combined into a multi-channel gradient tensor, with the number of channels equal to twice the number of scale levels, and each channel corresponding to a gradient component in one direction. Subsequently, the L2 norm is calculated along the channel dimension at each pixel location of this multi-channel gradient tensor, resulting in a unified gradient response intensity map. This multi-channel gradient response intensity map reflects the combined activity of local edges and textures in the image at different scales.
[0025] Gradient response intensity maps at different scales are upsampled to the original image resolution using bilinear interpolation, and weight coefficients are assigned based on the signal-to-noise ratio (SNR) estimates for each scale. The SNR estimates are obtained through local variance analysis: in smooth regions without significant edges, pixel value variance is mainly contributed by noise; this allows for the inference of noise power and subsequent SNR estimation. Scales with higher SNRs are assigned larger weights because their gradient responses are more reliable. All upsampled gradient response intensity maps are then weighted and fused to form the final multi-scale gradient response map. This multi-scale gradient response map serves as the weighting basis for subsequent local covariance matrix construction.
[0026] Based on the multi-scale gradient response map, a local directional covariance matrix is constructed at each pixel location. This step involves selecting a square neighborhood window of a fixed size, typically 7×7 or 9×9 pixels, centered on the current pixel. Within this square neighborhood window, the gradient vectors of all pixels are extracted. Each gradient vector consists of a horizontal gradient component and a vertical gradient component, forming a two-dimensional column vector. Subsequently, the weighted covariance matrix of these gradient vectors relative to the center pixel is calculated. The formula for calculating the weighted covariance matrix is as follows: for all pixels within the window except the center pixel, the corresponding values in the multi-scale gradient response map are used as weights to calculate the weighted average gradient vector. Then, the cross product of each gradient vector and the weighted average gradient vector is calculated, and all cross product results are weighted and summed. Finally, the result is divided by the sum of the weights to obtain a 2×2 covariance matrix.
[0027] The covariance matrix captures the distribution characteristics of gradient directions within a local region. Eigenvalue decomposition of this covariance matrix yields two orthogonal eigenvectors and their corresponding non-negative real eigenvalues. The eigenvector corresponding to the larger eigenvalue is defined as the local principal direction, which aligns with the local edge or texture direction of the image. The ratio of the two eigenvalues quantifies the anisotropy of the local structure: a larger ratio indicates a more directional local structure; a ratio close to 1 indicates a smooth, isotropic region.
[0028] By utilizing the aforementioned local principal directions and degree of anisotropy, a spatially varying elliptical kernel function is constructed. This elliptical kernel function... The mathematical expression for is: ; Indicates the current pixel coordinates. Represents the coordinates of any pixel within the neighborhood. For The local covariance matrix centered at the center is constructed as follows: ; This is a rotation matrix, where the elements are determined by the local principal direction angle. Decide, It is calculated from the direction angle of the eigenvector. and Let these represent the standard deviations along the principal direction and the perpendicular direction, respectively. They satisfy the following relationship: , and Let be two eigenvalues of the covariance matrix, and . The absolute value is limited by the estimated global noise level: the global noise level is estimated by subtracting the structural variance from the global variance of the image. If the calculated... If the noise level exceeds the threshold corresponding to the global noise level, it is truncated to that threshold to ensure that the kernel function does not over-expand due to local abnormal responses, thereby avoiding the introduction of artifacts.
[0029] The elliptical kernel function is used as a weighted window to perform weighted average filtering on pixels within the neighborhood. For each pixel... The filtered output value is a normalized weighted sum of all pixel values in its neighborhood multiplied by the corresponding kernel function weights. Weight normalization is achieved by summing all kernel function values and taking the reciprocal. Based on this, a non-local mean prior constraint is introduced. This non-local mean prior constraint is introduced by additionally calculating the structural similarity measure between the current pixel's image patch and other candidate image patches in the entire image when performing weighted average filtering.
[0030] Image patch sizes are typically set to 5×5 or 7×7 pixels. Structural similarity is measured based on a normalized cross-correlation coefficient, combined with a gradient orientation consistency correction term. The normalized cross-correlation coefficient is calculated by dividing the dot product of the two image patches (after removing the mean) by the product of their respective standard deviations. The gradient orientation consistency correction term is obtained by calculating the mean cosine similarity of the gradient directions of corresponding pixels within the two image patches, and is used to penalize patch matching with inconsistent orientations.
[0031] The center pixel of the corresponding candidate block is included in the reconstruction set of the current pixel only when the structural similarity metric is greater than a preset threshold. The reconstruction weights are jointly determined by the adaptive elliptic kernel weights and the structural similarity metric, and are fused in a product form: that is, the final weight is equal to the weight value of the elliptic kernel at the center pixel position multiplied by the structural similarity metric value of the candidate block to which the center pixel belongs. All effective weights are normalized again and then used for weighted summation to obtain the final filtered output.
[0032] For missing or degraded regions in the image, an iterative filling strategy guided by an adaptive kernel field is implemented. This iterative filling strategy first identifies the boundaries of missing regions in the image. Missing regions are identified by pixel masks, where a mask value of zero indicates a valid pixel, and a non-zero value indicates a missing pixel. Boundary pixels are defined as points in the missing region that have at least one valid eight-neighbor pixel. The pixels inside the missing region are initialized to zero or the interpolation results of neighboring valid pixels obtained through bilinear interpolation. In each iteration, only the outermost boundary pixels of the current missing region are processed. For each boundary pixel to be repaired, the local covariance matrix and the corresponding adaptive kernel function are recalculated based on the structural information of its surrounding known pixels (including repaired valid pixels).
[0033] The calculation process is the same as the previous steps, but the neighborhood window only includes valid pixels, and invalid pixels are excluded from the covariance calculation. Using the adaptive kernel function and non-local similarity block information, a weighted reconstruction is performed on the invalid pixel. The search range of the non-local similarity block is limited to the repaired region and the original valid region to ensure data reliability. After reconstruction, the invalid pixel's state is updated to "repaired," and its pixel value is written to the image buffer, while the mask state is updated. This invalid pixel then becomes the reference pixel for the next iteration. All steps are repeated, expanding the repair boundary layer by layer until all missing pixels are repaired.
[0034] The output is a high-quality image after noise reduction and restoration. This high-quality image exhibits good noise suppression in smooth areas, retains clear structural details in edge and textured areas, and the restoration results of missing areas maintain a high degree of consistency with the surrounding content in terms of geometric structure and visual style.
[0035] At the system level, the implementation of this invention relies on a complete image processing system. This image processing system includes a raw image input module, which receives and buffers noisy or partially missing raw image data. A multi-scale gradient response map generation module is responsible for performing Gaussian pyramid construction, multi-scale gradient calculation, L2 norm normalization, and weighted fusion, outputting a unified gradient response map.
[0036] The local covariance matrix construction module takes the gradient response map as input, calculates the weighted covariance matrix at each pixel location, performs eigenvalue decomposition, and outputs parameters for the local principal direction and anisotropy level. The adaptive elliptic kernel function generation module dynamically constructs the elliptic kernel function corresponding to each pixel based on the above parameters and manages the storage and retrieval of kernel function parameters. The weighted filtering and nonlocal constraint fusion module performs a dual-weighted fusion filtering operation, internally including an image patch matching engine, a structural similarity calculator, and a weighted fusion unit.
[0037] The missing region iterative repair module manages missing masks, boundary detection, iterative scheduling, and pixel state updates, coordinating the collaborative work of various modules during the repair process. The high-quality image output module is responsible for writing the final processing results to a specified storage location or display device. All modules exchange data via shared memory or message queues to ensure efficient and synchronized processing. The system runs on a hardware platform with sufficient computing resources, supporting parallel processing to accelerate the computation of large-scale images.
[0038] 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.
[0039] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. An image denoising and restoration method based on adaptive kernel estimation, characterized in that, include: Acquire raw image data that is noisy or partially missing; Multi-scale gradient magnitude calculation is performed on the original image data to generate a gradient response map characterizing the local structural complexity of the image; Based on the gradient response map, a local directional covariance matrix is constructed at each pixel location, and the local principal direction and degree of anisotropy are determined according to the eigenvalue decomposition results of the local directional covariance matrix. Using the local principal direction and the degree of anisotropy, a spatially varying elliptical kernel function is constructed, with the major axis extending along the local edge direction and the minor axis contracting perpendicular to the local edge direction, and the ratio of the major axis to the minor axis dynamically adjusted by the local gradient intensity. The elliptical kernel function is used as a weighted window to perform weighted average filtering on the pixels in the neighborhood. At the same time, a non-local mean prior constraint is introduced to search for image patches with similar local structural features from the entire image range and include them in the reconstruction weight calculation of the current pixel. For missing or degraded regions in the image, the repair boundary is expanded outward layer by layer based on the iterative filling strategy guided by the elliptical kernel function. Each iteration updates the kernel function parameters according to the structural information of the repaired region until the entire missing region is completely reconstructed. Output a high-quality image after noise reduction and restoration.
2. The image denoising and restoration method based on adaptive kernel estimation according to claim 1, characterized in that, Performing multi-scale gradient magnitude calculations on the original image data to generate a gradient response map characterizing the local structural complexity of the image includes: Gaussian pyramid downsampling is performed on the original image to obtain sub-images at multiple scale levels; At each scale level, calculate the first-order partial derivatives in the horizontal and vertical directions respectively; The first-order partial derivatives at each scale are combined into a multi-channel gradient tensor, and a unified gradient response intensity map is obtained by normalizing with the L2 norm. Gradient response intensity maps at different scales are upsampled to the original resolution and then weighted and fused to form the final multi-scale gradient response map. The weight coefficients of each scale are dynamically allocated according to their signal-to-noise ratio estimates.
3. The image denoising and restoration method based on adaptive kernel estimation according to claim 2, characterized in that, Based on the gradient response map, a local directional covariance matrix is constructed at each pixel location, including: A square neighborhood window of fixed size is selected with the current pixel as the center; Within this square neighborhood window, extract the gradient vectors of all pixels; Calculate the weighted covariance matrix of these gradient vectors relative to the center pixel, where the weights are determined by the corresponding values in the multi-scale gradient response map; The weighted covariance matrix is decomposed into eigenvalues to obtain two orthogonal eigenvectors and their corresponding eigenvalues. The eigenvector corresponding to the eigenvalue with the larger value among two eigenvalues is defined as the local principal direction, and the ratio of the two eigenvalues is used to quantify the degree of anisotropy of the local structure.
4. The image denoising and restoration method based on adaptive kernel estimation according to claim 3, characterized in that, Using the aforementioned local principal directions and degree of anisotropy, a spatially varying elliptic kernel function is constructed, including: Elliptic kernel function The mathematical expression is: ； Indicates the current pixel coordinates. Represents the coordinates of any pixel within the neighborhood. For The local covariance matrix centered at the center is constructed as follows: ； This is a rotation matrix, where the elements are determined by the local principal direction angle. Decide, It is calculated from the direction angle of the eigenvector; and Let these represent the standard deviations along the principal direction and the perpendicular direction, respectively. They satisfy the following relationship: , and Let be two eigenvalues of the covariance matrix, and .
5. The image denoising and restoration method based on adaptive kernel estimation according to claim 4, characterized in that, The elliptical kernel function is used as a weighted window to perform weighted average filtering on pixels in the neighborhood, while introducing a non-local mean prior constraint, including: Calculate the structural similarity measure between the image patch containing the current pixel and other candidate image patches in the entire image. The structural similarity measure is based on the normalized cross-correlation coefficient and combined with the gradient direction consistency correction term. The center pixel of the corresponding candidate block is included in the reconstruction set of the current pixel only when the structural similarity metric is greater than a preset threshold. The reconstruction weights are determined by the product of the weight value of the elliptical kernel function at the center pixel position and the structural similarity measure, and are then normalized and used for weighted summation.
6. The image denoising and restoration method based on adaptive kernel estimation according to claim 5, characterized in that, The normalized cross-correlation coefficient is obtained by calculating the dot product of the two image blocks after removing the mean and dividing it by the product of their respective standard deviations; the gradient direction consistency correction term is obtained by calculating the mean cosine similarity of the gradient directions of corresponding pixels in the two image blocks.
7. The image denoising and restoration method based on adaptive kernel estimation according to claim 6, characterized in that, For missing or degraded regions in the image, an iterative filling strategy guided by the elliptical kernel function includes: Identify the boundaries of missing regions in an image. A boundary pixel is a point in the missing region that has at least one eight-neighbor pixel that is a valid pixel. Initialize the interpolation results where the pixel value inside the missing region is zero or the value of the nearest valid pixel; In each iteration, only the outermost boundary pixels of the current missing region are processed; For each boundary pixel to be repaired, the local covariance matrix and the corresponding elliptical kernel function are recalculated based on the structural information of the surrounding known pixels. The neighborhood window contains only valid pixels. The effective pixel is reconstructed using the elliptical kernel function and non-local similarity block information. The search range of the non-local similarity block is limited to the repaired area and the original effective area. Update the state of the valid pixel to "repaired" and include it in the valid reference set for the next iteration; Repeat all steps until all missing pixels are repaired.
8. The image denoising and restoration method based on adaptive kernel estimation according to claim 7, characterized in that, The missing area is identified by a pixel mask, where a mask value of zero indicates a valid pixel and a non-zero value indicates a missing pixel. The pixel mask is updated synchronously after each iteration.
9. The image denoising and restoration method based on adaptive kernel estimation according to claim 8, characterized in that, The global noise level is estimated by subtracting the structural variance from the global variance of the image; if the calculated... If the noise level exceeds the threshold corresponding to the global noise level, it will be truncated to that threshold.
10. The image denoising and restoration method based on adaptive kernel estimation according to claim 9, characterized in that, The Gaussian pyramid is constructed by convolving the current-scale image with a standard Gaussian low-pass filter and then downsampling it by half its size; the first-order partial derivatives are calculated using the Sobel operator or the Scharr operator.