A method for denoising continuous frames of high-speed motion

Abstract

This invention relates to the field of video image processing technology and discloses a method for denoising high-speed motion continuous frames. The invention receives a raw video sequence in Bayer format and establishes a multi-scale image pyramid with a variance-stabilized domain. At each level of the pyramid, based on local brightness statistics, local texture contrast information, and estimated motion amplitude, a motion texture discriminability index representing texture change relative to noise intensity is calculated. This index is used to dynamically adjust the data fidelity term and smoothing constraint term in the variational optical flow estimation process to obtain a motion-compensated image. Subsequently, this index is mapped to an alignment deviation estimate and a similarity measurement range, multi-frame spatiotemporal fusion weights are calculated, and a primary fusion image is generated. Finally, an adaptive filtering threshold is calculated based on this index, high-frequency texture details are extracted from the differential data and superimposed back onto the primary fusion image, and a denoised video sequence is output after inverse transformation processing.

Description

Technical Field

[0001] This invention relates to the field of video image processing technology, and more specifically, to a method for denoising continuous frames of high-speed motion. Background Technology

[0002] High-speed video acquisition systems are commonly used to record and analyze fast-moving targets. To effectively freeze motion and avoid image blur, imaging devices typically require extremely short exposure times. However, extremely short exposure times significantly reduce the number of photons received by the image sensor, resulting in captured images filled with severe shot noise that is highly correlated with signal strength. Simultaneously, high frame rates lead to extremely short time intervals between adjacent frames, causing the target's displacement within adjacent frames to be very small, typically only at the sub-pixel level.

[0003] To improve the signal-to-noise ratio and visual quality of such noisy videos, most existing video denoising techniques employ multi-frame spatiotemporal fusion methods based on motion compensation. These methods typically first utilize registration techniques such as multi-scale optical flow estimation to calculate motion vectors from adjacent frames to the reference frame, spatially aligning the images, and then perform multi-frame fusion using methods such as weighted averaging.

[0004] However, when faced with complex scenes involving high frame rates and short exposures, existing multi-scale motion compensation and fusion techniques reveal significant technical shortcomings. Specifically, when the inter-frame displacement of the target is at the sub-pixel level and the dark areas of the image are dominated by strong shot noise, the motion estimation process is easily influenced by noise, leading to severe sub-pixel drift in the estimated motion vectors. Especially in areas where the target contains high-frequency fine textures, the existing coarse-to-fine multi-scale transfer mechanism not only fails to effectively correct the deviation but also amplifies these minute alignment errors. Ultimately, during the multi-frame spatiotemporal fusion stage, due to the inaccurate alignment of the images, the fusion system misjudges these high-frequency fine textures as random noise and forcibly averages them out, resulting in irreversible texture collapse and detail loss in the denoised image, severely reducing the noise reduction quality and clarity of high-speed video. Summary of the Invention

[0005] This invention provides a method for denoising continuous frames of high-speed motion, which solves the technical problems mentioned in the background art.

[0006] This invention provides a method for denoising continuous frames of high-speed motion, comprising: The system receives raw video sequences in Bayer format, performs linearization and noise parameter calibration on them, and transforms them to the variance stable domain to build a multi-scale image pyramid. At each level of the multi-scale image pyramid, a motion texture distinguishability index, which characterizes texture changes relative to noise intensity, is calculated based on local brightness statistics, local texture contrast information, and estimated motion amplitude. The weighting coefficients are determined using the motion texture resolution index, and the data fidelity term and smoothing constraint term in the variational optical flow estimation process based on the current reference frame are dynamically adjusted to obtain a motion-compensated image. The motion texture discriminability index is mapped to the alignment deviation estimate and similarity measurement range. Based on this, the multi-frame spatiotemporal fusion weight is calculated, and the motion-compensated image is weighted and fused to generate a primary fused image. An adaptive filtering threshold is calculated based on the motion texture discriminability index. High-frequency texture details are extracted from the differential data between the current reference frame and the primary fused image using the adaptive filtering threshold and superimposed back onto the primary fused image. After inverse transformation processing, a denoised video sequence is output.

[0007] The beneficial effects of this invention are as follows: By explicitly calculating the motion texture resolvability index, which characterizes texture changes relative to noise intensity, this invention overcomes the technical challenge of motion vectors being severely dragged by shot noise due to extremely short exposures and the superposition of minute displacements in high-speed acquisition scenarios. This invention utilizes this motion texture resolvability index to dynamically adjust the entire link in optical flow estimation, spatiotemporal fusion, and residual filtering. Under harsh conditions where strong noise and sub-pixel motion coexist, it can accurately distinguish between real high-frequency details and random noise interference, effectively avoiding texture collapse and detail extinguishing caused by blind alignment and fusion of multiple frames. Ultimately, it achieves high-speed video denoising output with both high signal-to-noise ratio and extremely high definition. Attached Figure Description

[0008] Figure 1 This is a flowchart of a high-speed motion continuous frame denoising method according to the present invention. Detailed Implementation

[0009] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0010] It should be noted that, unless otherwise defined, the technical or scientific terms used in one or more embodiments of the present invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in one or more embodiments of the present invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" indicate that the element or object preceding the term encompasses the elements or objects listed following the term and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0011] like Figure 1 As shown, a method for denoising continuous frames of high-speed motion includes: The system receives raw video sequences in Bayer format, performs linearization and noise parameter calibration on them, and transforms them to the variance stable domain to build a multi-scale image pyramid. At each level of the multi-scale image pyramid, a motion texture distinguishability index, which characterizes texture changes relative to noise intensity, is calculated based on local brightness statistics, local texture contrast information, and estimated motion amplitude. The weighting coefficients are determined using the motion texture resolution index, and the data fidelity term and smoothing constraint term in the variational optical flow estimation process based on the current reference frame are dynamically adjusted to obtain a motion-compensated image. The motion texture discriminability index is mapped to the alignment deviation estimate and similarity measurement range. Based on this, the multi-frame spatiotemporal fusion weight is calculated, and the motion-compensated image is weighted and fused to generate a primary fused image. An adaptive filtering threshold is calculated based on the motion texture discriminability index. High-frequency texture details are extracted from the differential data between the current reference frame and the primary fused image using the adaptive filtering threshold and superimposed back onto the primary fused image. After inverse transformation processing, a denoised video sequence is output.

[0012] Preferably, the process involves receiving the original video sequence in Bayer format, linearizing it, calibrating its noise parameters, and transforming it to the variance-stable domain to establish a multi-scale image pyramid. Specifically, this includes: Step A: For the original video sequence in Bayer format For each sub-sampling channel of each frame of the image, the linear electronic domain image is calculated using the following formula. : ; in, For pixel coordinates, The preset black level value. The system gain coefficient, This is a function to find the maximum value. Step B: Image the linear electronic domain Divide the image into patches, filter flat region samples and bin them according to the mean, and calculate the first... Sample mean of each bin With sample variance The shot noise ratio coefficient is estimated by minimizing the weighted error function shown in the following equation. With readout noise variance : ; in, For the first Number of samples in each bin; Step C: Utilize the shot noise scaling factor With readout noise variance The linear electronic domain image is obtained by the following formula. Perform a generalized Anscomb transform to generate a variance stability region image. : ; Step D: The variance stability region image is calculated using the following formula. Perform recursive processing to construct a hierarchy. The multi-scale image pyramid: ; in, The variance stability domain image at the original resolution. This indicates a 2x downsampling operation. The standard deviation is expressed as Gaussian smoothing kernel, This represents the convolution operation.

[0013] The Bayer format original video sequence is the input data for the entire denoising method and can be acquired using a high-speed camera. The preset black level value is a parameter, preferably 64, which is the black level calibration range for common 16-bit Bayer format images and conforms to general industrial camera settings. The system gain coefficient is an acquisition parameter that can be obtained from the calibration documentation provided by the camera manufacturer or calculated by comparing grayscale values ​​using a standard calibration board. The linear electronic domain image is intermediate data obtained after the original video sequence has undergone black level and non-negative truncation processing and multiplied by the system gain coefficient. The bin index is an identifier used to distinguish different mean bins. The sample mean of the k-th bin is the arithmetic mean of all pixel values ​​within that bin. The sample variance of the k-th bin is the squared average of the differences between the pixel values ​​within that bin and the sample mean. The shot noise scaling factor is a Poisson-Gaussian noise model parameter obtained through weighted least squares fitting. The readout noise variance is a Poisson-Gaussian noise model parameter obtained through weighted least squares fitting. The variance stability domain image is intermediate data obtained after the linear electronic domain image undergoes a generalized Anscombe transform. The standard deviation of the Gaussian smoothing kernel is a parameter, preferably 1.2. This value effectively balances smoothing effect and detail preservation, conforming to the general optimization range of multi-scale pyramid construction. The Gaussian smoothing kernel is a convolution kernel with a set standard deviation used for image smoothing. The 2x downsampling operation halves the resolution of the image in both the horizontal and vertical directions. The multi-scale image pyramid levels are used to distinguish images of different resolutions; a value of 0 represents the original resolution, and a larger value indicates a lower resolution.

[0014] The specific method for selecting flat region samples involves dividing the linear electronic domain image into non-overlapping image blocks. The horizontal and vertical gradients of each image block are calculated using the Sobel operator, and the gradient energy is obtained by summing them. The top 20% of image blocks with the smallest gradient energy are then selected as flat region samples. This selection method avoids interference from texture or motion on noise parameter calibration, ensuring accurate calibration results. For example, for videos containing high-speed rotating blades, this method can exclude blade texture areas, retaining only flat background areas for noise calibration.

[0015] The generalized Anscom transform is based on the input linear electronic domain image and the fitted shot noise scaling factor and readout noise variance. It transforms the signal intensity-related Poisson Gaussian noise into approximately unit Gaussian white noise, thus solving the problem of increased processing difficulty caused by noise variation with signal intensity in high-speed video.

[0016] The construction process of the multi-scale image pyramid is as follows: first, Gaussian smoothing filtering is performed on the variance-stabilized domain image, and then a 2x downsampling operation is performed. This process is repeated until the preset coarsest level is reached. The image of each level is obtained based on the processing of the previous level, ensuring the correlation and consistency between levels.

[0017] The specific size of the non-overlapping image blocks is set to eight by eight pixels. This size can ensure the number of samples while avoiding excessively large image blocks containing too much texture information, which would affect the accuracy of noise parameter calibration.

[0018] The specific number of bins for binning statistics is set at fifty. Fifty bins can completely cover the brightness range of the linear electronic domain image, ensuring that the fitted noise parameters are more accurate.

[0019] The Gaussian smoothing kernel is set to a size of 7 by 7 pixels. This size can effectively achieve a smoothing effect without excessively blurring image details.

[0020] The maximum number of levels in the multi-scale image pyramid is set to four, meaning the coarsest level is four. This number of levels balances processing efficiency and multi-scale optimization effects, avoiding excessive computation due to too many levels or insufficient optimization due to too few levels.

[0021] In weighted least squares fitting, the effective bin number threshold is set to at least 500 pixels per bin. This threshold avoids fitting errors caused by small-sample binning and ensures the stability and reliability of noise parameter estimation.

[0022] Preferably, at each level of the multi-scale image pyramid, a motion texture discriminability index characterizing texture change relative to noise intensity is calculated based on local brightness statistics, local texture contrast information, and estimated motion amplitude. Specifically, this includes: Step A: In the multi-scale image pyramid... Hierarchy, utilizing the linear electronic domain image Calculated by the following formula in pixels local window centered The local brightness statistics within : ; in, This represents the total number of pixels within a local window. The pixel index within the window; Step B: Utilize the variance stability region image The local texture contrast information is calculated using the following formula. : ; in, Represents the gradient operator, This indicates taking the square modulo 1. To prevent tiny constants with a denominator of zero; Step C: Utilize the variance stability region image In the local window Windowed power spectrum The dominant spatial frequency is calculated using the following formula. : ; in, Frequency domain coordinate index; Step D: At the coarsest level of the multi-scale image pyramid The phase correlation method is used to calculate the integer pixel displacement, and the sub-pixel displacement is obtained by fitting a quadratic surface in the peak neighborhood. The two are then combined to obtain the displacement amplitude of the coarsest layer. Then, by iteratively applying the following formula layer by layer downwards, the estimated motion amplitude of the current level is obtained. : ; Step E: Combine the readout noise variance The motion texture resolvability index is calculated using the following formula. : ; The local statistical window is a fixed-range region centered on a pixel, used to calculate local features. The local brightness statistics are the arithmetic mean of the linear electronic domain pixel values ​​within the local statistical window at the current level of the multi-scale image pyramid. The local texture contrast information is the ratio of the gradient magnitude energy to the pixel magnitude energy of the variance-stable domain image within the local statistical window at the current level of the multi-scale image pyramid. The dominant spatial frequency is the centroid frequency of the power spectrum distribution of the variance-stable domain image within the local statistical window at the current level of the multi-scale image pyramid. The coarsest scale level is the lowest resolution level of the multi-scale image pyramid, a parameter, preferably 4. This number of levels balances the accuracy and computational efficiency of motion amplitude prediction, meeting the practical needs of high-speed video processing. The coarsest layer displacement amplitude is the value obtained by merging the integer pixel displacement obtained through phase correlation and the sub-pixel displacement obtained through quadratic surface fitting at the coarsest scale level. The current level predicted motion amplitude is obtained by sampling and multiplying the displacement amplitude of the coarsest scale level layer by layer upwards. The motion texture discriminability index is the ratio of texture change intensity to noise intensity, used to characterize the discriminability of texture changes relative to noise. The zero-prevention constant is an extremely small value used to avoid zero denominators during calculations; it is a parameter, preferably one part per million. This value effectively prevents zero denominators without substantially affecting the calculation results. The frequency domain coordinate index is used to identify different frequency positions in the frequency domain. The windowed power spectrum is the power spectrum distribution obtained by frequency domain transformation of the variance stability domain image within a local statistical window after window function processing.

[0023] Local texture contrast information is obtained by first calculating the gradient magnitude of each pixel within a local statistical window of the variance-stabilized domain image and summing the gradient magnitude energy, and simultaneously calculating the magnitude of each pixel and summing the pixel magnitude energy. The ratio of the two is then used as the local texture contrast information. This calculation method eliminates the interference of pure brightness scale on contrast, making it suitable for high-speed video scenes with low light and strong noise. For example, in videos containing fast-moving targets like thin rods, it can accurately depict the relative undulation intensity of the target texture, unaffected by the overall low brightness.

[0024] The dominant spatial frequency is determined by processing the variance stability domain image within a local statistical window using a window function to reduce spectral leakage. A frequency domain transformation is then performed to obtain the power spectrum distribution, and finally, the centroid frequency of the power spectrum distribution is calculated as the dominant spatial frequency. This method can accurately capture the core frequency features of common high-frequency thin structures or periodic textures found in high-speed moving targets.

[0025] Cross-scale propagation of motion amplitude involves first obtaining an accurate displacement amplitude reference at the coarsest scale level, and then sampling and propagating it layer by layer upwards according to the multiplication factor corresponding to the level difference. This ensures that the motion amplitude at each level is adapted to the resolution of that level, maintaining consistency in motion representation. For example, if the coarsest level is 4 and the current level is 2, the multiplication factor is 4, meaning the motion amplitude at the current level is 4 times that of the coarsest level, which conforms to the scaling rules of multi-scale image resolution.

[0026] The motion texture resolvability index comprehensively characterizes the intensity of texture changes by multiplying four parameters: local brightness statistics, local texture contrast information, dominant spatial frequency, and estimated motion amplitude; it characterizes the intensity of noise by adding the local brightness statistics to the readout noise variance; and it obtains the index by the ratio of the two, thereby quantifying the relative intensity of texture changes and noise.

[0027] The local statistical window is nine by nine pixels. This size can include enough pixel samples to ensure statistical reliability, while avoiding an excessively large window that contains too much irrelevant texture or motion information, which would affect the accuracy of local feature calculation.

[0028] The gradient operator is the Sobel operator. The gradient magnitude is calculated by using the Sobel operator to calculate the gradient values ​​in the horizontal and vertical directions respectively, and then obtaining the gradient magnitude by taking the square root of the sum of the squares. This method is a mature scheme for image gradient calculation, which can balance computational efficiency and gradient representation accuracy.

[0029] The frequency domain transform is set to Fast Fourier Transform.

[0030] The window function type for the windowed power spectrum is set to Hann window. This window can effectively suppress spectral leakage and smooth the transition of edge regions of local images, avoiding spectral distortion caused by window truncation.

[0031] The peak neighborhood range of the quadratic surface fitting is set to three by three pixels. This range can accurately capture the sub-pixel information near the integer pixel peak obtained by the phase correlation method, and the sub-pixel displacement deviation obtained by fitting is more accurate.

[0032] The mapping relationship between frequency coordinates and period per pixel is set as follows: if the local statistical window size is N times N pixels, the range of the frequency domain coordinate index is from negative N divided by two to positive N divided by two minus one, and the corresponding period per pixel value is the frequency domain coordinate index divided by N.

[0033] The specific value of the zero-prevention constant is set to one part per million.

[0034] Preferably, determining the weighting coefficients using the motion texture discriminability index specifically includes: Set preset data item transition control constants With smoothing term transition control constant ; For each pixel in the current level of the multi-scale image pyramid Using the motion texture resolution index at that pixel location The weighting coefficient for adjusting the data fidelity item is calculated using the following formula. : ; Using the motion texture resolution index at this pixel location The weighting coefficient of the smoothing constraint term used to adjust the smoothing constraint term is calculated using the following formula. : ; The data term transition control constant is a parameter used to adjust the weighting coefficient of the data fidelity term, preferably 0.5. This value balances the dynamic adjustment range of data fidelity and smoothing constraints, adapting to scenarios where high-speed video subpixel motion and strong noise coexist. The smoothing term transition control constant is a parameter used to adjust the weighting coefficient of the smoothing constraint term, preferably 0.5. This value makes the smoothing constraint strength change more gradually with the motion texture resolvability index, avoiding over-smoothing or data distortion. The motion texture resolvability index is a quantitative indicator characterizing texture change relative to noise intensity, calculated from local brightness statistics, local texture contrast information, dominant spatial frequency, and estimated motion amplitude. The data fidelity term weighting coefficient is a coefficient used to adjust the weight of the data fidelity term in variational optical flow estimation, calculated from the motion texture resolvability index and the data term transition control constant. The smoothing constraint term weighting coefficient is a coefficient used to adjust the weight of the smoothing constraint term in variational optical flow estimation, calculated from the smoothing term transition control constant and the motion texture resolvability index.

[0035] The mapping of weighting coefficients is continuously adjusted through fractional calculation. The weighting coefficient of the data fidelity term increases with the motion texture resolution index, which means that the larger the index, the higher the data reliability and the more it relies on the image brightness difference for motion estimation; the weighting coefficient of the smoothness constraint term decreases with the increase of the index, which means that the larger the index, the more reliable the motion information and the weaker the spatial smoothness constraint.

[0036] The transition control constants for the data term and the smoothing term are the same and fixed at 0.5, making the changing trends of the two weighting coefficients complementary and ensuring the coordinated adaptation of data fidelity and structure regularization during motion estimation. For example, when the motion texture resolution index is 0.2, the weighting coefficient of the data fidelity term is approximately 0.29, and the weighting coefficient of the smoothing constraint term is approximately 0.71. In this case, the smoothing constraint is prioritized to suppress noise interference. When the index is 1.0, both weighting coefficients are 0.5, achieving a balance between data fidelity and smoothing constraints.

[0037] The pixel-wise calculation of weighting coefficients enables variational optical flow estimation to adapt to the differences in texture and noise distribution in different regions of an image, avoiding local estimation bias caused by globally uniform weights, and is especially suitable for the characteristic differences between target and background regions in high-speed videos.

[0038] The preferred value range for the transition control constant of the data item is 0.3 to 0.7. A value of 0.5 is suitable for most high-speed video scenarios. If the noise is extremely strong, a value of 0.3 can be used. If the motion information is relatively reliable, a value of 0.7 can be used.

[0039] The preferred range of the transition control constant for the smoothing term is consistent with that for the transition control constant for the data term, which is 0.3 to 0.7, to ensure that the adjustment ranges of the two weighting coefficients are matched.

[0040] The common range of motion texture resolution index is 0.1 to 2.0. A value less than 0.5 is considered a low resolution scene, and a value greater than 1.0 is considered a high resolution scene. This range is statistically derived based on the characteristics of high-speed video with short exposure, strong noise, and subpixel motion.

[0041] During the calculation of weighting coefficients, if the motion texture resolution index approaches 0, the weighting coefficient of the data fidelity term approaches 0, and the weighting coefficient of the smoothing constraint term approaches 1. At this time, motion estimation depends entirely on spatial smoothing constraints. If the index is much larger than the transition control constant, the weighting coefficient of the data fidelity term approaches 1, and the weighting coefficient of the smoothing constraint term approaches 0. At this time, motion estimation depends entirely on the image brightness difference.

[0042] Preferably, the data fidelity term and smoothing constraint term in the variational optical flow estimation process based on the current reference frame are dynamically adjusted by combining weighting coefficients to obtain a motion-compensated image, specifically including: Step A: At the current level of the multi-scale image pyramid With the current reference frame Variance stability region image Based on this, construct a framework for adjacent frames. motion vector field variational energy objective function : ; in, Weighting coefficients for the data fidelity items, The weighting coefficients for the smoothing constraint term, The regularization coefficient for the smoothing term. The gradient of the motion vector. Let be the Chabonniere bar penalty function. It is a tiny constant; Step B: Adopt a strategy from coarse to fine scale, moving the upper level... Estimated motion vector field Perform a 2x upsampling as the current level. Initial motion vector: ; Step C: Minimize the variational energy objective function using the successive over-relaxation iterative method, at the th... In the next iteration, the current motion vector is first used to... Resampling And calculate the spatial derivative. With time derivative Then update the pixels using the following formula. Horizontal motion component at the location With vertical motion component : ; ; in, For pixels The components of the average motion vector in the four neighborhoods, To prevent zero constant; Step D: Based on the final motion vector field obtained through iterative solution The motion-compensated image is generated using the following formula. : ; The variational energy objective function is used to solve for the motion vector field, consisting of a data fidelity term and a smoothing constraint term. The robust penalty function handles image brightness differences, employing the Charbonnier robust penalty form. The smoothing term regularization coefficient is a parameter that adjusts the strength of the smoothing constraint term, preferably 15, to adapt to the smoothing requirements of sub-pixel motion estimation in high-speed video, balancing motion detail preservation and estimation stability. The motion vector gradient is the rate of change of the motion vector in space, characterizing the spatial smoothness of the motion vector. The small constant of the robust penalty function is a parameter used to avoid computational anomalies in the robust penalty function, preferably one part per million. This value effectively prevents abrupt changes in function values ​​without affecting the penalty effect. The coarse-scale motion vector field is the motion vector field of the coarse level in the multi-scale image pyramid, providing the initial basis for motion vector estimation at the current level. The 2x upsampling operation doubles the resolution of the image or motion vector field in both the horizontal and vertical directions. The current-scale initial motion vector field is the initial value of the motion vector at the current level of the multi-scale image pyramid, obtained by 2x upsampling the coarse-scale motion vector field. The outer iteration count is the number of outer iterations in minimizing the variational energy objective function. It is a parameter, preferably 5, which balances solution accuracy and computational efficiency, avoiding increased time consumption due to excessive iteration. The inner iteration count of the successive over-relaxation iteration is the number of inner iterations of the successive over-relaxation iteration. It is a parameter, preferably 40, which ensures iteration convergence and meets the accuracy requirements of motion vector estimation. The resampled reference frame variance stable domain image is the image obtained by resampling the variance stable domain images of adjacent frames based on the current motion vector. The spatial gradient in the x-direction is the gradient value of the resampled image in the horizontal direction, used to characterize the horizontal change in image brightness. The spatial gradient in the y-direction is the gradient value of the resampled image in the vertical direction, used to characterize the vertical change in image brightness. The temporal gradient is the brightness difference between adjacent frames and the resampled reference frame, used to characterize the temporal change in image brightness. The four-neighbor average motion vector horizontal component is the average of the horizontal motion vector components of the four neighboring pixels above, below, left, and right of the current pixel. The four-neighbor average motion vector vertical component is the average of the vertical motion vector components of the four neighboring pixels above, below, left, and right of the current pixel. The horizontal motion vector component is the horizontal component of the motion vector, used to describe the horizontal displacement of a pixel. The vertical motion vector component is the vertical component of the motion vector, used to describe the vertical displacement of a pixel. The motion-compensated image is a variance-stabilized domain image obtained by resampling adjacent frames using motion vectors, and it has a spatial alignment with the reference frame.

[0043] The data fidelity term consists of the weighted coefficients of the data fidelity term and the difference in image brightness after robust penalty, while the smoothing constraint term consists of the weighted coefficients of the smoothing constraint term and the squared magnitude of the motion vector gradient. This enables dynamic adjustment of the motion texture resolvability index to the motion estimation. For example, when the motion texture resolvability index is small, the weighted coefficients of the data fidelity term are small while the weighted coefficients of the smoothing constraint term are large. The objective function focuses more on the spatial smoothness of the motion vector, avoiding noise interference.

[0044] The initial values ​​for the coarse-to-fine solution strategy are set as follows: the motion vector of the previous level is upsampled by a factor of two and used as the initial motion vector of the current level. This ensures the continuity of motion vectors across different levels and reduces the propagation of multi-scale errors. For example, the coarse-level motion vector has low resolution but good global consistency. Upsampling it and using it as the initial value for the fine-level level can quickly guide the fine-level motion vectors to converge to a reasonable range.

[0045] The successive over-relaxation iteration update mechanism combines the average motion vector of the four-neighborhood, spatial gradient, temporal gradient, and weighting coefficients to update the motion vector components pixel by pixel, adapting to the fine estimation requirements of sub-pixel motion. This update method can fully utilize local motion information and image brightness change information, making motion vector estimation more accurate.

[0046] The motion compensation implementation includes: resampling adjacent frame variance-stabilized domain images using bilinear interpolation, combined with reflection boundary processing to ensure effective compensation of pixels in boundary regions and avoid fusion errors caused by missing boundary pixels. For example, for high-speed moving targets at image edges, reflection boundary processing can provide a reasonable interpolation basis, ensuring the integrity of motion compensation.

[0047] The specific value of the tiny constant in the robust penalty function is one part per million.

[0048] The preferred range for the number of outer iterations is 3 to 7. A value of 5 is suitable for most high-speed video scenarios. If higher estimation accuracy is required, a value of 7 can be used. If computational efficiency is desired, a value of 3 can be used.

[0049] The optimal range for the number of iterations in successive over-relaxation iterations is 30 to 50. A value of 40 can ensure iteration convergence and balance convergence speed and computational cost.

[0050] The calculation of the four-neighbor average motion vector is set to equal weight average, that is, the sum of the motion vector components of four adjacent pixels is divided by 4. This method is simple to calculate and can effectively represent the local motion trend.

[0051] The convergence criterion is set to stop once a preset number of external iterations is reached.

[0052] When mapping pixel coordinates using bilinear interpolation, the pixel center coordinate system is used, and reflection boundary processing is applied to coordinates that are outside the image range.

[0053] The spatial gradient is calculated using the Sobel operator, while the temporal gradient is calculated directly using the brightness difference between adjacent frames and the resampled reference frame.

[0054] Preferably, mapping the motion texture discriminability index to an alignment deviation estimate and a similarity measurement range specifically includes: Set a preset uncertainty scaling constant Reference bandwidth constant and zero-prevention small constants ; For the current reference frame At the original resolution level Each pixel Using the motion texture resolution index at that pixel location The estimated alignment deviation is calculated using the following formula. : ; Using the motion texture resolution index at this pixel location The similarity metric range is calculated using the following formula. : ; The uncertainty scaling constant is a parameter used to adjust the magnitude of the alignment deviation estimate, preferably 0.25. This value matches the alignment uncertainty magnitude of sub-pixel motion in high-speed video, making the estimate more closely match the actual error range. The baseline bandwidth constant is a parameter used to set the basic range of the similarity measurement, preferably 1.0. This value balances the sensitivity and stability of similarity determination and is suitable for most high-speed video scenarios. The zero-prevention constant is a parameter used to avoid zero denominators in calculations, preferably one part per million. This value effectively prevents calculation anomalies without substantially affecting the results. The alignment deviation estimate is a parameter characterizing the reliability of motion compensation alignment, calculated from the uncertainty scaling constant, the square of the motion texture resolvability index, and the zero-prevention constant. The similarity measurement range is an effective interval parameter used to determine the similarity of inter-frame patches, calculated from the baseline bandwidth constant, the motion texture resolvability index, and the zero-prevention constant. The current reference frame is a single-frame image used as a baseline during multi-frame fusion. The original resolution level is the level in the multi-scale image pyramid where the resolution has not been downsampled. A pixel is the basic unit of an image. Motion texture discriminability index is a quantitative parameter that characterizes texture changes relative to noise intensity. It is calculated from local brightness statistics, local texture contrast information, dominant spatial frequency, and estimated motion amplitude.

[0055] The mapping of the alignment deviation estimate is calculated using the squared term of the motion texture resolution index. Because the index is positively correlated with alignment reliability, the squared term amplifies the alignment deviation in areas with low index values, making the estimate more accurately reflect the differences in alignment quality. For example, when the motion texture resolution index is 0.2, the estimated alignment deviation is approximately 6.25, indicating extremely low alignment reliability in that area; when the index is 1.0, the estimated value is 0.25, indicating relatively reliable alignment.

[0056] The mapping of the similarity measurement range involves multiplying the square root of the motion texture distinguishability index by the baseline bandwidth constant. This makes the range change more smoothly with the index, avoiding abrupt changes in the similarity judgment standard caused by small fluctuations in the index. For example, when the index increases from 0.2 to 1.0, the similarity measurement range increases from approximately 0.45 to 1.0, which reflects the improvement in the sensitivity of similarity judgment without excessively relaxing the standard.

[0057] Both mapping relationships are built around the motion texture distinguishability index, realizing the collaborative logic that the lower the index, the greater the alignment deviation and the smaller the similarity range, thus avoiding blind fusion.

[0058] The preferred value range for the uncertainty scaling constant is 0.1 to 0.5. A value of 0.25 is suitable for most high-speed video scenarios. If the noise is extremely strong, a value of 0.1 can be used. If the motion is relatively stable, a value of 0.5 can be used.

[0059] The preferred value range for the reference bandwidth constant is 0.8 to 1.2, with 1.0 being the most versatile. If the texture details are rich, a value of 1.2 can be used, while if there is a lot of noise, a value of 0.8 can be used.

[0060] The common numerical range of motion texture resolution index in mapping calculation is 0.1 to 2.0. This range is based on the statistical characteristics of high-speed video with short exposure, strong noise, and subpixel motion, and covers most practical application scenarios.

[0061] The original resolution level is specifically identified as level zero.

[0062] Preferably, the multi-frame spatiotemporal fusion weights are calculated based on the alignment deviation estimate and the similarity metric range, and the motion-compensated images are weighted and fused to generate a primary fused image, specifically including: Step A: Set the time fusion window radius With the stability constant of the fusion denominator ; Step B: For the time fusion window range For each neighboring image within a frame, the pixel value is calculated using the following formula. Centered local patch window Local patch difference values ​​within : ; in, For the motion-compensated image, The variance-stabilized domain image of the current reference frame at the original resolution level. This represents the total number of pixels within the local patch window. Step C: Combine the aforementioned similarity measurement range The alignment deviation estimate The frame corresponding to the following formula is calculated. The multi-frame spatiotemporal fusion weights : ; in, The zero-prevention small constant; Step D: Perform weighted fusion on all motion-compensated images within the time fusion window to generate the primary fused image using the following formula. : ; For time indexes that exceed the range of the video sequence, reflection boundary rules are used for processing.

[0063] The temporal fusion window radius is a parameter used to limit the time range of multi-frame fusion, preferably 3. This value balances the fusion noise reduction effect with computational efficiency, covers a sufficient number of adjacent frames before and after the reference frame, and does not introduce too much redundant data. The fusion denominator stability constant is a parameter used to avoid the denominator being too small after summing the fusion weights, preferably one millionth. This value effectively prevents computational anomalies caused by the denominator approaching zero and does not affect the accuracy of the fusion result. The local patch window is a fixed range area centered on a pixel used to calculate inter-frame similarity. The local patch difference value is the squared mean of the pixel differences between the motion-compensated image and the current reference frame within the local patch window, used to characterize the similarity of local regions in the two frames. The exponential decay term is a weight factor calculated based on the normalized local patch difference value, used to reflect the impact of inter-frame structural similarity on the fusion weights. The uncertainty penalty term is the reciprocal of the sum of the alignment deviation estimate and the fusion denominator stability constant, used to reflect the impact of motion compensation alignment reliability on the fusion weights. The multi-frame spatiotemporal fusion weights are the fusion weights of the corresponding neighborhood images, obtained by multiplying the exponential decay term and the uncertainty penalty term. The primary fused image is obtained by weighted summation and normalization of all motion-compensated images within the temporal fusion window. The current reference frame is the single-frame image used as the baseline during multi-frame fusion. The original resolution level is the unsampled level in the multi-scale image pyramid. The neighborhood image is the adjacent frame image surrounding the current reference frame within the temporal fusion window. The motion-compensated image is the neighborhood image resampled by motion vectors and spatially aligned with the current reference frame. The zero-prevention constant is an extremely small value used to avoid zero denominators in the calculation.

[0064] The local patch difference values ​​are calculated using the squared mean, which is more sensitive to capturing minute alignment deviations than the ordinary mean, making it suitable for high-speed video subpixel motion scenes. For example, when there is a tiny deviation of 0.2 pixels in motion compensation, the squared mean can amplify the difference caused by this deviation, avoiding fusion blurring caused by the deviation being too small to be ignored.

[0065] The construction logic of the multi-frame spatiotemporal fusion weights is a dual weighting: an exponential decay term reflects the structural similarity between frames, with higher similarity resulting in greater weights; an uncertainty penalty term reflects the alignment reliability, with more reliable alignment resulting in greater weights. Multiplying these two terms ensures that neighboring frames with similar structures and reliable alignment receive higher fusion weights, avoiding blind weighted averaging. For example, if a neighboring frame has small local patch differences with the reference frame and a low estimated alignment deviation, its fusion weight will be significantly higher than that of neighboring frames with large differences or unreliable alignment.

[0066] The time window endpoint processing uses reflection boundary rules to ensure that even when the reference frame is at the end of the time series, a fixed number of neighboring frames can still be fused, avoiding uneven noise reduction effects caused by insufficient fusion of boundary frames. For example, when the reference frame is the first frame of the video and there are no earlier neighboring frames, the reflection boundary rules use the second frame as the preceding neighboring frame of the first frame to ensure that the number of fused frames is 7.

[0067] The optimal value range for the time fusion window radius is 2 to 5. A value of 3 is suitable for most high-speed video scenarios. If higher noise reduction effect is required, a value of 5 can be used. If real-time performance is required, a value of 2 can be used.

[0068] The local patch window is set to 7 by 7 pixels. This size can contain enough local structural information to ensure the accuracy of similarity determination, while avoiding the increase in computation caused by an excessively large window.

[0069] The preferred value range for the stability constant of the fusion denominator is one part per ten million to one part per ten thousand.

[0070] The reflection boundary rule is as follows: if the time index of the neighboring frame is less than 1, then 2 minus that index is used as the index after reflection; if the time index of the neighboring frame is greater than the total number of video frames, then twice the total number of frames minus that index is used as the index after reflection.

[0071] The normalization method for the local patch difference values ​​in the exponential decay term is as follows: divide the local patch difference values ​​by the sum of the square of the similarity measurement range and the zero-prevention constant, so that the normalized value range matches the decay characteristics of the exponential function.

[0072] The calculation order for the local patch difference value is as follows: first, calculate the difference between the corresponding pixel of the motion-compensated image and the current reference frame, square the difference, and then average all squared differences within the local patch window to obtain the final local patch difference value.

[0073] Preferably, an adaptive filtering threshold is calculated based on the motion texture discriminability index. High-frequency texture details are extracted from the difference data between the current reference frame and the primary fused image using the adaptive filtering threshold and superimposed back onto the primary fused image. After inverse transform processing, a denoised video sequence is output. Specifically, this includes: Step A: Set the preset threshold reference constant ; Step B: For the current reference frame At the original resolution level Each pixel Using the motion texture resolution index at that pixel location The adaptive filtering threshold is calculated using the following formula. : ; in, The zero-prevention small constant; Step C: Calculate the current reference frame With the primary fused image Image residual The image residual is then subjected to soft thresholding shrinkage using the following formula to obtain high-frequency texture details. : ; ; Step D: The high-frequency texture details are superimposed onto the primary fused image and numerically truncated to obtain the variance stability domain image after detail enhancement. : ; in, This is the preset maximum value limit; Step E: Utilize the shot noise scaling factor With the aforementioned readout noise variance The linear electronic domain denoised image is obtained by performing the generalized Anscomb inverse transform using the following formula. And perform non-negative truncation: ; Step F: Utilize the system gain coefficient With respect to the preset black level value The numerical domain image is generated using the following formula. The denoised video sequence is then combined and output as follows: ; in, This is the maximum value corresponding to the output bit depth.

[0074] The threshold baseline constant is a parameter used to calculate the adaptive filtering threshold, preferably 2.0. This value is suitable for extracting texture details in high-speed video with strong noise, balancing the needs of noise filtering and detail preservation. The zero-prevention constant is a parameter used to avoid zero denominators during calculation, preferably one part per million. This value effectively prevents calculation anomalies and does not substantially affect the calculation results. The adaptive filtering threshold is a parameter used to filter high-frequency texture details in the image residual, calculated from the threshold baseline constant, motion texture discriminability index, and zero-prevention constant. The current reference frame is a single-frame image used as a reference during multi-frame fusion and residual processing. The original resolution level is the level in the multi-scale image pyramid that has not undergone downsampling. Pixels are the basic building blocks of an image. The motion texture discriminability index is a quantitative parameter characterizing texture changes relative to noise intensity, calculated from local brightness statistics, local texture contrast information, dominant spatial frequency, and estimated motion amplitude. The image residual is the numerical difference between the current reference frame and the primary fused image at corresponding pixel positions, obtained by subtracting the pixels of the two frames. High-frequency texture details are the parts of the image residual retained after soft thresholding and shrinking, representing the true image texture information. The primary fusion image is obtained by weighted fusion of motion-compensated images within the temporal fusion window. The variance-stabilized domain image after detail enhancement is obtained by superimposing high-frequency texture details onto the primary fusion image and then numerically truncating it. The preset maximum numerical upper limit is a parameter used for numerical truncating the variance-stabilized domain image after detail enhancement, preferably 65535. This value adapts to the dynamic range of 16-bit images and conforms to the general output standard of industrial cameras. The shot noise scaling factor is a Poisson-Gaussian noise model parameter obtained through weighted least squares fitting. The readout noise variance is a Poisson-Gaussian noise model parameter obtained through weighted least squares fitting. The generalized inverse Anscomb transform is a transformation method that restores the variance-stabilized domain image to the linear electronic domain image, which needs to be performed in combination with the shot noise scaling factor and the readout noise variance. The linear electronic domain denoised image is obtained by performing the generalized inverse Anscomb transform and non-negative truncation on the detail-enhanced variance-stabilized domain image. The system gain coefficient is the coefficient used to convert the Bayer format raw image into a linear electronic domain image. It is a parameter obtained from the camera, which can be obtained from the calibration documentation provided by the camera manufacturer, or calculated by comparing grayscale values ​​using a standard calibration board. The preset black level value is the basic dark current value that needs to be subtracted from the Bayer format raw image. It is a parameter obtained from the camera's factory calibration data, or calculated by calculating the mean value of images acquired in dark environments. The output numerical domain is the range of values ​​converted from the linear electronic domain denoised image to conform to the output standard. The Bayer format is the raw output format of the image sensor, containing multiple sub-sampling channels. Channel reassembly is the operation of rearranging the processed linear electronic domain denoised image according to the order of the Bayer format sub-sampling channels.A denoised video sequence is a noise-free video sequence that retains details after full-process processing. It is composed of denoised frames combined in chronological order. The maximum value corresponding to the output bit depth is a parameter that adapts to the bit depth of the output image, preferably 65535. This value is the maximum quantization value of a 16-bit image and is a commonly used bit depth standard for high-speed video acquisition.

[0075] The adaptive filtering threshold is calculated using a threshold baseline constant as the numerator and the sum of the motion texture resolvability index and the zero-prevention micro constant as the denominator, enabling the threshold to dynamically change with the motion texture resolvability index. A smaller index indicates stronger regional noise and weaker texture; a larger threshold results in more significant residual shrinkage, effectively filtering noise. Conversely, a larger index indicates stronger regional texture and weaker noise; a smaller threshold preserves more subtle textures. For example, a larger threshold in dark, low-resolvability areas filters noise residuals, while a smaller threshold in bright, high-resolvability areas preserves texture residuals.

[0076] Soft threshold shrinkage only retains the portion of the image residual whose absolute value is greater than the adaptive filtering threshold, and sets the residual whose absolute value is less than or equal to the threshold to zero. This method can separate noise residuals from real texture residuals, avoid image artifacts caused by hard thresholding, and adapt to the high-frequency texture detail extraction requirements of high-speed video.

[0077] The strategy for enhancing details involves overlaying the extracted high-frequency texture details onto the primary fused image and then performing numerical truncation. The overlay operation can restore the high-frequency textures lost during the fusion process, while the numerical truncation can prevent the values ​​after overlay from exceeding the dynamic range of the variance stability domain, thus ensuring the validity of the image data.

[0078] The output format restoration involves combining the system gain coefficient and black level value to convert the linear electronic domain denoised image into the output numerical domain, and then reconstructing the channels according to the Bayer format to completely restore the format characteristics of the original video, adapting to the original video output and subsequent processing requirements of industrial cameras.

[0079] The preferred value range for the threshold reference constant is 1.5 to 2.5. A value of 2.0 is suitable for most high-speed video scenarios. If the noise is extremely strong, a value of 2.5 can be used. If the texture details are extremely rich, a value of 1.5 can be used.

[0080] The specific calculation steps for soft threshold shrinkage processing are as follows: first, calculate the absolute value of the image residual; then, subtract the adaptive filtering threshold from the absolute value, retain the part of the result that is greater than 0, and then multiply it by the sign of the original image residual. The part of the result that is less than or equal to 0 is directly set to zero.

[0081] The preset maximum numerical upper limit is set according to the dynamic range of the numerical values ​​of the variance stability domain image. The value is 65535 for 16-bit images and 255 for 8-bit images. It can be adjusted according to the actual bit depth of the acquired image.

[0082] The specific implementation method of non-negative truncation is to set all pixel values ​​less than 0 in the linear electronic domain denoised image to 0, and retain values ​​greater than or equal to 0, so as to avoid image display abnormalities caused by negative values.

[0083] The generalized Anscom inverse transform is implemented by substituting the shot noise scaling factor and the readout noise variance, calculating step by step according to the transform formula, first calculating the squared term, then performing subtraction and division operations, and finally performing non-negative truncation.

[0084] The specific order of Bayer format channel reassembly is to fill the processed pixel values ​​back into the corresponding pixel positions of the channels according to the sub-sampling channel order of red, green, and blue.

[0085] The maximum value corresponding to the output bit depth can be adjusted according to actual needs. For eight-bit images, the preferred value is 255; for twelve-bit images, the preferred value is 4095; and for sixteen-bit images, the preferred value is 65535. These are all standard maximum quantization values ​​for the corresponding bit depth.

[0086] The specific method of numerical truncation is to set the pixel values ​​in the variance stabilization domain image after detail enhancement that are greater than the preset maximum value to the upper limit, set the values ​​that are less than 0 to 0, and leave the intermediate values ​​unchanged.

[0087] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0088] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A method for denoising continuous frames of high-speed motion, characterized in that, include: The system receives raw video sequences in Bayer format, performs linearization and noise parameter calibration on them, and transforms them to the variance stable domain to build a multi-scale image pyramid. At each level of the multi-scale image pyramid, a motion texture distinguishability index, which characterizes texture changes relative to noise intensity, is calculated based on local brightness statistics, local texture contrast information, and estimated motion amplitude. The weighting coefficients are determined using the motion texture resolution index, and the data fidelity term and smoothing constraint term in the variational optical flow estimation process based on the current reference frame are dynamically adjusted to obtain a motion-compensated image. The motion texture discriminability index is mapped to the alignment deviation estimate and similarity measurement range. Based on this, the multi-frame spatiotemporal fusion weight is calculated, and the motion-compensated image is weighted and fused to generate a primary fused image. An adaptive filtering threshold is calculated based on the motion texture discriminability index. High-frequency texture details are extracted from the differential data between the current reference frame and the primary fused image using the adaptive filtering threshold and superimposed back onto the primary fused image. After inverse transformation processing, a denoised video sequence is output.

2. The method for denoising continuous frames of high-speed motion according to claim 1, characterized in that, The process involves receiving raw video sequences in Bayer format, linearizing them, calibrating their noise parameters, and transforming them to the variance-stable domain to construct a multi-scale image pyramid, including: For each sub-sampling channel of each frame in the original video sequence in the Bayer format, a preset black level value is subtracted and non-negative truncation is performed, and then multiplied by the system gain coefficient to obtain a linear electronic domain image. The linear electronic domain image is divided into non-overlapping image blocks. Flat region samples are selected from the image blocks according to the gradient energy. The flat region samples are binned according to the mean size. The sample mean and sample variance in each bin are calculated. The weighted least squares method is used to perform linear regression fitting on the sample mean and sample variance to obtain the shot noise ratio coefficient and readout noise variance. Using the shot noise scaling factor and the readout noise variance, a generalized Anscom transform is performed on the linear electronic domain image to generate a variance stable domain image, wherein the generalized Anscom transform converts signal intensity-dependent Poisson Gaussian noise into approximately unit Gaussian white noise. Gaussian smoothing and downsampling operations are performed on the variance stability domain image to construct the multi-scale image pyramid step by step.

3. The method for denoising continuous frames of high-speed motion according to claim 2, characterized in that, At each level of the multi-scale image pyramid, based on local brightness statistics, local texture contrast information, and estimated motion amplitude, a motion texture discriminability index characterizing texture change relative to noise intensity is calculated, including: For the current level of the multi-scale image pyramid, a local statistical window is set on the linear electronic domain image, and the arithmetic mean of the pixel values ​​within the local statistical window is calculated to obtain the local brightness statistical value. A local statistical window is set on the variance stability domain image, and the ratio of gradient magnitude energy to pixel magnitude energy within the local statistical window is calculated to obtain the local texture contrast information. A frequency domain transformation is performed within the local statistical window of the variance stability domain image to calculate the centroid frequency of the power spectrum distribution and obtain the dominant spatial frequency. The global integer pixel displacement is calculated at the coarsest level of the multi-scale image pyramid using the phase correlation method. The sub-pixel displacement deviation is obtained by quadratic surface fitting. The merged displacement amplitude is then sampled up layer by layer and multiplied to obtain the estimated motion amplitude at the current level. The texture change intensity is obtained based on the local brightness statistics, the local texture contrast information, the dominant spatial frequency, and the estimated motion amplitude. The noise intensity is obtained based on the local brightness statistics and the readout noise variance. The ratio of the texture change intensity to the noise intensity is calculated to obtain the motion texture discriminability index.

4. The method for denoising continuous frames of high-speed motion according to claim 3, characterized in that, Determining the weighting coefficients using the motion texture discriminability index includes: Set the transition control constant for the data item and the transition control constant for the smoothing item; for each pixel in the current level of the multi-scale image pyramid, take the motion texture resolution index of the pixel as the first numerator, and take the sum of the motion texture resolution index of the pixel and the transition control constant for the data item as the first denominator. Calculate the ratio of the first numerator and the first denominator to obtain the weighting coefficient of the data fidelity item used to adjust the data fidelity item. Using the smoothing transition control constant as the second numerator, and the sum of the motion texture resolution index at the pixel location and the smoothing transition control constant as the second denominator, the ratio of the second numerator and the second denominator is calculated to obtain the smoothing constraint weighting coefficient used to adjust the smoothing constraint term.

5. A method for denoising continuous frames of high-speed motion according to claim 4, characterized in that, By combining weighting coefficients to dynamically adjust the data fidelity term and smoothing constraint term in the variational optical flow estimation process based on the current reference frame, a motion-compensated image is obtained, including: Construct a variational energy objective function that includes the data fidelity term and the smoothing constraint term, wherein the data fidelity term is composed of the weighting coefficient of the data fidelity term and the difference in image brightness after processing by the robust penalty function, and the smoothing constraint term is composed of the weighting coefficient of the smoothing constraint term and the square magnitude of the gradient of the motion vector. A strategy of solving layer by layer from coarse scale to fine scale is adopted. In the current level of the multi-scale image pyramid, the motion vector of the previous level is used as the initial motion vector of the current level after upsampling. The variational energy objective function is minimized using the successive over-relaxation iterative method. In each iteration, the variance stable domain images of adjacent frames are resampled and linearized based on the current initial motion vector. The spatial gradient and temporal gradient are calculated, and the average motion vector of the four neighborhoods of the current pixel position is calculated. Combining the weighting coefficients of the data fidelity term and the weighting coefficients of the smoothing constraint term, the horizontal and vertical motion vector components of the current pixel position are updated using the average motion vector of the four neighborhoods, the spatial gradient, and the temporal gradient until a preset number of iterations is reached; Based on the motion vector after iterative convergence, the variance stability domain images of the adjacent frames are resampled by bilinear interpolation to obtain the motion-compensated image.

6. The method for denoising continuous frames of high-speed motion according to claim 5, characterized in that, Mapping the motion texture discriminability index to an alignment deviation estimate and a similarity measurement range includes: Set the uncertainty scaling constant, the reference bandwidth constant, and the zero-prevention microconstant; For each pixel of the current reference frame at the original resolution level, the uncertainty scaling constant is used as the third numerator, and the sum of the square of the motion texture resolvability index at the pixel position and the zero-prevention small constant is used as the third denominator. The ratio of the third numerator and the third denominator is calculated to obtain the alignment deviation estimate. The similarity measurement range is obtained by adding the motion texture resolvability index of the pixel location to the anti-zero micro constant, taking the square root, and multiplying the square root result by the reference bandwidth constant.

7. A method for denoising continuous frames of high-speed motion according to claim 6, characterized in that, The multi-frame spatiotemporal fusion weights are calculated based on the alignment deviation prediction and similarity metric range, and the motion-compensated images are weighted and fused to generate a primary fused image, including: Set the temporal fusion window range and the fusion denominator stability constant; for each neighboring image within the temporal fusion window range, calculate the local patch difference value between the motion-compensated image and the current reference frame at the original resolution level, where the local patch difference value is the square mean of the pixel differences within the local patch window; The local patch difference values ​​are normalized using the similarity metric range and an exponential decay term is calculated. The alignment deviation estimate is added to the fusion denominator stability constant and the reciprocal is taken to obtain the uncertainty penalty term. The exponential decay term is multiplied by the uncertainty penalty term to obtain the multi-frame spatiotemporal fusion weight corresponding to the neighborhood image. The motion-compensated images within the time fusion window are weighted and summed using the multi-frame spatiotemporal fusion weights. The weighted sum is then divided by the sum of all the multi-frame spatiotemporal fusion weights and the zero-prevention constant to obtain the primary fused image.

8. A method for denoising continuous frames of high-speed motion according to claim 7, characterized in that, An adaptive filtering threshold is calculated based on the motion texture discriminability index. High-frequency texture details are extracted from the difference data between the current reference frame and the primary fused image using the adaptive filtering threshold and superimposed back onto the primary fused image. After inverse transform processing, a denoised video sequence is output, including: Set a threshold reference constant; for each pixel of the current reference frame at the original resolution level, use the threshold reference constant as the fourth numerator, and use the sum of the motion texture resolution index of the pixel position and the zero-prevention micro constant as the fourth denominator. Calculate the ratio of the fourth numerator and the fourth denominator to obtain the adaptive filtering threshold. Calculate the pixel difference between the current reference frame and the primary fused image to obtain the image residual; use the adaptive filtering threshold to perform soft threshold shrinkage on the image residual to obtain the shrunken high-frequency texture details, wherein the soft threshold shrinkage process retains the residual portion whose absolute value is greater than the adaptive filtering threshold; The high-frequency texture details are superimposed onto the primary fused image, and the superimposition result is numerically truncated according to a preset dynamic range to obtain a variance stable domain image with enhanced details. Using the shot noise scaling factor and the readout noise variance, a generalized inverse Anscomb transform is performed on the variance-stabilized domain image after detail enhancement to obtain a linear electronics domain denoised image, and the linear electronics domain denoised image is non-negatively truncated. Using the system gain coefficient and the preset black level value, the linear electronic domain denoised image is converted into the output numerical domain, and channel reconstruction is performed according to the Bayer format to obtain the denoised video sequence.

Images