A liquid rocket engine simulation test video noise reduction system and method

Abstract

This invention provides a noise reduction system and method for liquid rocket engine simulation test videos, addressing the technical problem that existing image noise reduction methods are insufficient to meet the practical engineering requirements of noise reduction in liquid rocket engine simulation test videos. The noise reduction system for liquid rocket engine simulation test videos provided by this invention employs a color space conversion module to convert the input original RGB format image frames into YUV format, achieving separation of luminance and chrominance components. The Y channel, containing the main noise, undergoes multi-step processing through a Y channel processing unit, forming a spatial-temporal joint adaptive mechanism. This enables spatial adaptation of edge weights and local variance to accommodate scenes with rapid motion, responding more quickly to changes in the current frame, and combating trailing phenomena. For the U and V channels, which contain only color information, bilateral filtering is used to perform lightweight noise reduction, balancing effectiveness and efficiency, significantly improving noise reduction performance and video stability.

Description

Technical Field

[0001] This invention relates to video noise reduction systems and methods, specifically to a video noise reduction system and method for a liquid rocket engine simulation test. Background Technology

[0002] Launch vehicles are the cornerstone of the space economy, and liquid rocket engines, as a core subsystem, directly determine the effectiveness of the launch vehicle through their performance and technological level. During the operation and testing of liquid rocket engines, malfunctions often manifest as abnormal flames (including extreme forms such as explosions), liquid leaks, and sensor detachment, many of which can be observed with the naked eye, thus providing possibilities for deep learning based on machine vision. Simultaneously, to ensure reliability and stability, high-speed cameras are required for video monitoring of liquid rocket engines, capturing images at thousands of frames per second, resulting in extremely redundant video data.

[0003] Currently, existing image denoising methods typically include traditional filtering methods, classical Kalman filtering methods, and deep learning-based image denoising methods. However, these methods each have different shortcomings in the denoising scenarios of liquid rocket engine simulation test videos. While traditional filtering methods can suppress noise to some extent, nonlocal mean denoising has high computational complexity, and temporal averaging denoising is prone to motion blur and trailing phenomena. Furthermore, both are difficult to adapt to the complex degradation models in test videos, such as local overexposure, high-energy particle trajectories, curtain-shock coupling stripes, and transient thermal wave distortion. Classical Kalman filtering methods rely on single-frame prediction, resulting in low accuracy in complex video motion, a tendency to produce ghosting, and insufficient protection of edges and details due to a lack of spatial adaptive adjustment. Deep learning-based image denoising methods, such as convolutional neural network-based denoising models, can recover some details, but the models are complex, have many parameters, high computational cost, limited generalization ability, and poor adaptability to the real-time processing requirements of test videos. There are other noise reduction methods, such as wavelet transform, which is sensitive to threshold selection, and median filtering, which has difficulty in pixel determination under high noise density. These methods cannot simultaneously meet the requirements of real-time performance, detail preservation, and noise suppression balance in liquid rocket engine simulation test video noise reduction.

[0004] In summary, existing image denoising methods generally suffer from three major problems: insufficient real-time performance (high computational complexity), limited adaptability (unable to adapt to complex degradation models), and an imbalance between denoising and detail preservation (easily resulting in blurring and trailing). These problems make it difficult to meet the practical engineering requirements for denoising liquid rocket engine simulation test videos. There is an urgent need for an image denoising method that balances efficient denoising, detail preservation, and real-time performance. Summary of the Invention

[0005] The purpose of this invention is to solve the technical problem that existing image denoising methods are difficult to meet the actual engineering needs of denoising videos for simulated test runs of liquid rocket engines, and to provide a denoising system and method for denoising videos for simulated test runs of liquid rocket engines.

[0006] To achieve the above objectives, the technical solution provided by this invention is as follows:

[0007] A video noise reduction system for a liquid rocket engine simulation test is characterized by including an input module, a color space conversion module, a Y channel processing unit, a UV channel processing module, a channel fusion module, and an output module.

[0008] The input terminal of the input module is used to receive the original frame sequence of the simulated test video of the liquid rocket engine and output the original RGB format image frames frame by frame.

[0009] The input end of the color space conversion module is connected to the output end of the input module. It is used to convert the original RGB format image frame into YUV format and output it in three channels. The Y channel outputs the luminance component image, the U channel outputs the blue chrominance component image, and the V channel outputs the red chrominance component image.

[0010] The Y-channel processing unit includes a pre-denoising module, an edge weight calculation module, a local variance calculation module, a spatial adaptive Q-map construction module, a multi-frame weighted memory prediction model, an improved Kalman filter, and a high-frequency compensation module.

[0011] The input of the pre-denoising module is connected to the output of the Y channel in the color space conversion module, and is used to perform pre-denoising processing on the luminance component image to obtain the pre-denoised Y channel image. The inputs of the edge weight calculation module and the local variance calculation module are respectively connected to the output of the pre-denoising module. The edge weight calculation module is used to detect the edge regions of the pre-denoised Y channel image and obtain the edge weights. The local variance calculation module is used to estimate the local noise level of each pixel in the pre-denoised Y channel image and obtain the local variance. The two inputs of the spatial adaptive Q_map construction module are respectively connected to the outputs of the edge weight calculation module and the local variance calculation module, and are used to combine the edge weights and local variances to construct the pixel-level spatial adaptive process noise covariance Q_map.

[0012] The multi-frame weighted memory prediction model includes a frame buffer for storing historical frame images of the luminance component. The input of the model is connected to the output of the Y channel in the color space conversion module. This allows the model to add the luminance component image of the current frame output from the Y channel of the color space conversion module to the frame buffer, and then perform prediction based on a weighted average of the multiple frames in the frame buffer to obtain the predicted value. ;

[0013] The three inputs of the improved Kalman filter are respectively connected to the outputs of the spatial adaptive Q_map construction module, the multi-frame weighted memory prediction model, and the pre-denoising module, and are used to calculate the noise covariance Q_map of the spatial adaptive process and the predicted value. Kalman gain is calculated pixel by pixel to fuse the predicted values. The Kalman-filtered Y-channel image is obtained by comparing the pre-denoised Y-channel image with the Y-channel image.

[0014] The two input terminals of the high-frequency compensation module are respectively connected to the output terminals of the improved Kalman filter and the local variance calculation module. It is used to receive the Y-channel image and local variance after Kalman filtering, and to repair the high-frequency details that were smoothed during the Kalman filtering process to obtain the Y-channel image after high-frequency enhancement.

[0015] The two input terminals of the UV channel processing module are respectively connected to the output terminals of the U channel and V channel of the color space conversion module. It is used to perform lightweight denoising on the chromaticity component images output by the U channel and V channel using a bilateral filtering algorithm to obtain the bilateral filtered U channel image and V channel image.

[0016] The three input terminals of the channel fusion module are respectively connected to the output terminal of the high-frequency compensation module and the two output terminals of the UV channel processing module. It is used to merge the high-frequency enhanced Y channel image, the bilaterally filtered U channel image and V channel image back into a complete image, and convert it back to RGB format to obtain a denoised RGB format image frame.

[0017] One input terminal of the output module is connected to the output terminal of the channel fusion module, and the other input terminal is used to receive the original video frame rate so as to stitch the denoised RGB format image frames into a complete denoised video according to the original video frame rate.

[0018] Furthermore, the spatial adaptive Q_map construction module constructs the pixel-level spatial adaptive process noise covariance Q_map using the following formula:

[0019] ;

[0020] Where Q_base is the baseline of the process noise, For frame-level adaptive coefficients, For local variance, For edge weights.

[0021] Furthermore, the multi-frame weighted memory prediction model calculates the predicted value using the following formula. :

[0022] ;

[0023] in, Let be the exponential decay weighting coefficient of the i-th frame in the frame buffer. Let M be the brightness value of the i-th frame preceding time t in the frame buffer, M be the current frame number in the frame buffer, and i be the frame sequence number.

[0024] Furthermore, the improved Kalman filter calculates the Kalman-filtered Y-channel image using the following formula:

[0025] ;

[0026] in, To update the denoised values ​​pixel by pixel, that is, the pixel values ​​of the Y channel image after Kalman filtering; This is the predicted value for the current frame. This refers to the observation value of the current frame, i.e., the Y-channel image of the current frame after pre-denoising. For Kalman gain, , ∈[0,1], To predict covariance, To measure the noise covariance.

[0027] Furthermore, the high-frequency compensation module calculates the enhanced Y-channel image using the following formula:

[0028] Y_enh=clip(Y_blur+λ×ΔY,0,255);

[0029] Where Y_enh is the Y channel image after high-frequency enhancement, Y_blur is the Gaussian blurred image, ΔY is the high-frequency component, λ is the high-frequency enhancement gain, and clip(·,0,255) is the clipping function used to ensure that the pixel value is within the range of [0,255].

[0030] Furthermore, the UV channel processing module calculates the bilaterally filtered U-channel image using the following formula:

[0031] ;

[0032] in, This is the U-channel image after bilateral filtering, where W is the normalization constant. For the current pixel, For neighboring pixels, This is the chroma component image of the current pixel output by the U channel. σ is the chromaticity component image of the neighboring pixels output by the U channel; σ is the spatial kernel. For the range kernel;

[0033] The calculation method for the bilaterally filtered V channel image and the bilaterally filtered U channel image in the UV channel processing module is the same.

[0034] Furthermore, the color space conversion module converts the input original RGB format image frame into YUV format through linear matrix operations, the expression of which is:

[0035] ,

[0036] Where Y is the luminance component, U is the blue chromaticity component, V is the red chromaticity component, and R, G, and B are the red, green, and blue primary color components, respectively.

[0037] Furthermore, while performing pre-denoising processing on the luminance component image, the pre-denoising module utilizes the weighted average characteristic of similar patches to initially retain key details and obtain the pre-denoised Y channel image.

[0038] Furthermore, the edge weight calculation module calculates the edge weights based on the Laplace operator and the exponential decay function;

[0039] The local variance calculation module calculates the local variance based on the box filter of the local variance calculation window.

[0040] In addition, the present invention also provides a method for noise reduction of liquid rocket engine simulation test videos, based on the above-mentioned liquid rocket engine simulation test video noise reduction system, characterized by the following steps:

[0041] Step 1: The input module reads the original frame sequence of the liquid rocket engine simulation test video and outputs the original RGB format image frames frame by frame, which are then transmitted to the color space conversion module.

[0042] Step 2: The color space conversion module converts the input original RGB format image frame into YUV format, and outputs the luminance component image, blue chrominance component image and red chrominance component image through the Y channel, U channel and V channel respectively.

[0043] Step 3: The UV channel processing module uses a bilateral filtering algorithm to perform lightweight denoising on the blue chroma component image output by the U channel and the red chroma component image output by the V channel, respectively, to obtain the bilaterally filtered U channel image and V channel image.

[0044] Simultaneously, the brightness component image output from the Y channel undergoes adaptive Kalman filtering via the Y channel processing unit, specifically including:

[0045] The first step involves the pre-denoising module performing pre-denoising on the luminance component of the Y channel output to obtain the pre-denoised Y channel image; the edge weight calculation module then detects the edge regions of the pre-denoised Y channel image and calculates the edge weights. Simultaneously, the local variance calculation module estimates the local noise level of each pixel in the pre-denoised Y-channel image and calculates the local variance σ²(x); the spatial adaptive Q-map construction module combines edge weights... And the local variance σ²(x) are used to construct the pixel-level spatial adaptive process noise covariance Q_map;

[0046] The second step involves the multi-frame weighted memory prediction model adding the current frame's luminance component image output from the Y channel to the frame buffer, and then performing predictions based on a weighted average of multiple frames in the frame buffer to obtain pixel-level predictions without residual noise. ;

[0047] The third step involves the improved Kalman filter adapting to the spatial adaptive process noise covariance Q_map and the predicted value. Kalman gain is calculated pixel by pixel to fuse the predicted values. The Kalman-filtered Y-channel image is obtained by comparing the pre-denoised Y-channel image with the Y-channel image.

[0048] The fourth step involves the high-frequency compensation module receiving the Kalman-filtered Y-channel image and local variance σ²(x), repairing the high-frequency details smoothed during the Kalman filtering process, and avoiding high-frequency enhancement in edge regions and noise-dense areas, while outputting the high-frequency enhanced Y-channel image.

[0049] Step 4: The channel fusion module merges the high-frequency enhanced Y channel image, the bilaterally filtered U channel image, and the V channel image back into a complete image, and converts it back to RGB format to obtain a denoised RGB format image frame.

[0050] Step 5: The output module stitches the denoised RGB format image frames together into a complete denoised video at the original video frame rate.

[0051] Compared with the prior art, the present invention has the following beneficial technical effects:

[0052] 1. Compared to existing image denoising methods that often process the RGB channels uniformly, this invention provides a liquid rocket engine simulation test video denoising system. It uses a color space conversion module to convert the input original RGB format image frames to YUV format, achieving separation of luminance and chrominance components. The Y channel, which contains the main noise, undergoes multi-step processing through seven sub-modules of the Y channel processing unit, focusing on balancing noise suppression and detail preservation. For the U and V channels, which contain only color information, a computationally efficient bilateral filter is used for lightweight denoising, balancing effectiveness and efficiency. This invention employs a channel-specific differentiated processing strategy, significantly improving denoising performance and video stability. Compared to existing image denoising methods, this invention improves the peak signal-to-noise ratio (PSNR) by an average of 2.57 dB, increases the structural similarity index (SSIM) by 67.9%, and improves the signal-to-noise ratio (SNR) by 12.43 dB, effectively suppressing complex noise such as local overexposure and high-energy particle trajectories.

[0053] 2. The present invention provides a video noise reduction system for simulated liquid rocket engine test. Through an edge weight calculation module, a local variance calculation module, and a spatial adaptive Q-map construction module, a spatial-temporal joint adaptive mechanism is formed, enabling the edge weights and local variance to achieve spatial adaptation. This adapts to scenes with violent motion (such as image jitter caused by the vibration of a liquid rocket engine). The inter-frame difference index is superior to traditional methods, responding faster to changes in the current frame, combating trailing phenomena, and further improving video stability. At the same time, compared with conventional Kalman filtering systems, which use a globally fixed Q value in video denoising, the present invention achieves temporal adaptation through a multi-frame weighted memory prediction model and a weighted average of multiple frames in the frame buffer. This avoids ghosting caused by relying on the denoising result of the previous frame, adapting to the complex dynamic scenes of simulated liquid rocket engine test videos.

[0054] 3. The present invention provides a noise reduction system for simulated test videos of liquid rocket engines. By using an edge weight calculation module to reduce the filtering update intensity in the edge region to avoid edge smoothing, and to enhance the filtering intensity in the flat region to improve the noise reduction effect, the system also uses a high-frequency compensation module to compensate the flat region with high frequency to restore weak textures. This effectively preserves key visual features such as engine flame shape, plume characteristics, and component status. The edge variance and high-frequency energy indicators are reasonable, which greatly improves the ability to preserve details and edges.

[0055] 4. Compared to the high complexity O(N×K) of existing deep learning-based image denoising methods, where N is the number of pixels and K is the model parameters, the denoising method for simulated test videos of liquid rocket engines provided in this invention significantly reduces the filtering complexity. The overall single-frame processing complexity is O(2^1615N), the processing complexity of the improved Kalman filter is only O(6N), and the processing complexity of the UV channel processing module is O(2N). The computational efficiency is close to that of mean filtering, and it can run efficiently on a 4090D-24G GPU, meeting the real-time monitoring requirements of test runs (abnormal response time <50ms).

[0056] 5. The present invention provides a noise reduction method for simulated test videos of liquid rocket engines, which has good scalability and adaptability, supports the fusion of subsequent neural network models (such as CNN-based anomaly detection models), and can be integrated into visualization software for engineering applications; the algorithm parameters can be adjusted within a reasonable range to adapt to simulated test video scenarios with different noise intensities. Attached Figure Description

[0057] Figure 1 This is a structural block diagram of an embodiment of a video noise reduction system for simulated test of a liquid rocket engine according to the present invention;

[0058] Figure 2 This is a structural block diagram of the Y-channel processing unit in an embodiment of the present invention. Detailed Implementation

[0059] To make the objectives, advantages, and features of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Those skilled in the art should understand that these embodiments are merely used to explain the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0060] like Figure 1 As shown, a video noise reduction system for a liquid rocket engine simulation test includes an input module, a color space conversion module, a Y-channel processing unit, a UV-channel processing module, a channel fusion module, and an output module. The specific functions, input / output signals, and relationships of each module are as follows:

[0061] I. Input Module.

[0062] The input module receives the raw frame sequence of liquid rocket engine simulation test video and outputs raw RGB format image frames frame by frame for transmission to the color space conversion module. The raw RGB format image frames contain pixel data for the R (red), G (green), and B (blue) channels respectively. The input module supports high frame rate (thousands of frames per second) and high resolution video data acquired by high-speed cameras, ensuring complete and lossless frame data to provide raw input for subsequent processing. The format of the raw frame sequence of liquid rocket engine simulation test video is not limited, supporting mainstream video encoding, and includes video frame sequences with complex noise such as white fog interference, overexposure, and vibration. Each frame is in RGB format (uint8 type, pixel value range [0,255]).

[0063] II. Color Space Conversion Module.

[0064] The input end of the color space conversion module is connected to the output end of the input module. It is used to perform color space conversion according to the BT.601 standard, that is, to convert the input original RGB format image frame into YUV format, realizing the separation of luminance component and chrominance component, laying the foundation for subsequent channel-specific differential processing. The output end of the color space conversion module is divided into Y channel, U channel and V channel. Among them, the Y channel is used to output the luminance component image, the U channel outputs the blue chrominance component image, and the V channel outputs the red chrominance component image. All three channels are single channels of uint8 type.

[0065] Considering the human visual system's sensitivity to luminance information but relative insensitivity to chromaticity information, this invention employs a channel-specific differential processing strategy within the YUV color space. Since the Y channel contains image structure and edge information, it is the primary carrier of noise interference and requires focused denoising. The U and V channels, on the other hand, contain only color information, have lower requirements for detail resolution, and are less affected by noise; therefore, only computationally efficient bilateral filtering is needed for lightweight denoising.

[0066] The core conversion logic of the color space conversion module is as follows: color space conversion is achieved through linear matrix operations, and its expression is as follows:

[0067] ;

[0068] Where Y is the luminance component, U is the blue chromaticity component, V is the red chromaticity component, and R, G, and B are the red, green, and blue primary color components, respectively.

[0069] III. Y-channel processing unit.

[0070] The Y-channel processing unit is the core processing module of the system of this invention. It adopts adaptive Kalman filtering and specifically includes seven sub-modules: a pre-denoising module, an edge weight calculation module, a local variance calculation module, a spatial adaptive Q-map construction module, a multi-frame weighted memory prediction model, an improved Kalman filter, and a high-frequency compensation module. It is used to perform multi-step processing on the Y-channel containing the main noise to achieve a balance between noise suppression and detail preservation.

[0071] (1) Pre-denoising module.

[0072] The input of the pre-denoising module is connected to the Y-channel output of the color space conversion module. It performs pre-denoising processing on the luminance component image output from the Y-channel, obtaining a pre-denoised Y-channel image. This suppresses some Gaussian noise and high-frequency interference, providing a cleaner input for subsequent Kalman filtering. Simultaneously, it utilizes the weighted averaging characteristic of similar patches to initially preserve key details such as flame edges and feather textures. The output of the pre-denoising module outputs the pre-denoised Y-channel image, which is then transmitted to the edge weight calculation module and the local variance calculation module, respectively.

[0073] The core processing logic of the pre-denoising module is as follows: The filter intensity parameter h is set to 8.0; experiments have verified that this value strikes the optimal balance between noise suppression and detail preservation. The search window Ω is set to 21×21 to cover a sufficiently large neighborhood to find similar patches. For each pixel x, the similarity weights of all pixels y within the search window Ω are calculated. P(x) and P(y) are image patch vectors centered at x and y, respectively. Let be the Euclidean distance. The brightness value Y(x) of the pre-denoised pixel x is calculated using the normalization constant C(x), where C(x) = Σ_{y∈Ω}w(x,y) and Y(x) = 1 / C(x) × Σ_{y∈Ω}w(x,y)Y(y), where Pixels in the luminance component image output for the Y channel The brightness value.

[0074] (2) Edge weight calculation module.

[0075] The input of the edge weight calculation module is connected to the output of the pre-denoising module. It detects edge regions in the pre-denoised Y-channel image, such as flame boundaries and engine component outlines, and calculates edge weights to protect important edge structures and provide a spatial adaptive basis for subsequent Kalman filtering. The edge weight calculation module reduces the filter update intensity in edge regions to avoid edge smoothing and increases the filter intensity in flat regions to improve denoising performance. The output of the edge weight calculation module outputs the edge weights and transmits them to the spatial adaptive Q_map construction module. The pixel value range of the edge weights is (0,1).

[0076] The core processing logic of the edge weight calculation module is as follows: ① Calculate the edge intensity L(x) using the Laplacian operator to capture regions of sudden gray-level changes (edges) in the image. , ① Use the Laplace operator; ② Normalize the absolute value of the edge intensity L(x) to obtain the normalized edge intensity value. , e(x)∈[0,1], where 0 represents a flat region and 1 represents a strong edge region; ③ Set the edge penalty coefficient α=2.0 to control the degree of edge protection. Experiments have verified that this edge penalty coefficient can effectively avoid edge blurring. Then, the edge weight at pixel x is calculated by the exponential decay function. , =exp(-α× Edge strength value The closer to 1, the higher the edge weight at pixel x. The smaller the value, the weaker the Kalman filter's update capability in that region, thus preserving the edges; edge strength value When the value is close to 0, it represents a flat region, and the edge weights are... ≈1, which allows for strong noise reduction.

[0077] (3) Local variance calculation module.

[0078] The input of the local variance calculation module is also connected to the output of the pre-denoising module. It estimates the local noise level of each pixel in the pre-denoised Y-channel image and calculates the local variance to reflect the intensity of noise or texture within the region. High variance regions are typically noise-dense or texture-rich areas, while low variance regions are flat areas, providing a basis for adaptive noise adjustment in Kalman filtering. The output of the local variance calculation module outputs the local variance, which is then transmitted to the spatial adaptive Q-map construction module and the high-frequency compensation module, respectively. The unit of local variance is the square of gray levels.

[0079] The core processing logic of the local variance calculation module is as follows: The local variance calculation window size is set to k=5. Experiments have verified that a 5×5 local variance calculation window with box filtering can balance the sensitivity to small structures and the smoothness of variance estimation, avoiding missed detections of small edges due to an excessively large local variance calculation window, and avoiding excessive variance fluctuations due to an excessively small local variance calculation window. For each pixel x, ① calculate the local mean at pixel x.

[0080] , Let be the brightness value at the i-th position. ① Calculate the local neighborhood of pixel x; ② Calculate the local mean of the squared luminance value at pixel x. , ²; ③ Calculate the local variance σ²(x) at pixel x using the variance formula, which is: The local variance σ²(x) is clipped to make σ²(x)≥1e-3 to avoid numerical anomalies. The larger σ²(x) is, the higher the noise or texture intensity in that area.

[0081] (4) Spatial adaptive Q_map building module.

[0082] The two inputs of the spatial adaptive Q_map construction module are connected to the outputs of the edge weight calculation module and the local variance calculation module, respectively. This module combines edge weights and local variance to construct a pixel-level spatial adaptive process noise covariance Q_map, replacing the globally fixed Q value of traditional Kalman filtering. This allows the filtering process to adaptively adjust based on local image features (edge / flatness, noise intensity), achieving the dual goals of "edge protection + dynamic noise reduction." The output of the spatial adaptive Q_map construction module outputs the spatial adaptive process noise covariance Q_map and transmits it to the improved Kalman filter.

[0083] The core processing logic of the spatial adaptive Q_map building module is as follows: ① Set the baseline of process noise, Q_base=5.0, to provide the minimum noise reduction strength and ensure basic noise reduction effect in flat areas; ② Calculate the frame-level adaptive coefficients. It is used to scale the spatial adaptive process noise covariance Q_map according to the overall noise level. The frame-level adaptive coefficient γ is increased in high-noise scenes to improve the filtering response speed. c is the noise scale normalization factor, used to map the global noise intensity estimate to a stable adaptive adjustment range, avoiding Kalman update instability caused by excessive noise estimation. c=10.0. ③ Construct the pixel-level spatial adaptive process noise covariance Q_map, which is the median of the local variance matrix; Edge weights A large spatial adaptive noise covariance Q_map results in a large Kalman gain K, making the filter more dependent on the current observations, responding faster to changes in the current frame, and combating trailing phenomena; edge weights Small spatial domain adaptive process noise covariance Q_map reduces updates to protect edges.

[0084] (5) Multi-frame weighted memory prediction model.

[0085] The input of the multi-frame weighted memory prediction model is connected to the Y-channel output of the color space conversion module to acquire the luminance component image output by the Y-channel of the color space conversion module in order to calculate the predicted value. Specifically, the multi-frame weighted memory prediction model of this invention replaces the single-frame prediction of traditional Kalman filtering. It incorporates a frame buffer to store historical frame luminance component images. Each new frame's luminance component image is added to the frame buffer upon input. Prediction is then performed based on the weighted average of the multiple frames in the frame buffer, yielding pixel-level predictions without residual noise. To avoid ghosting caused by relying on the denoising results of the previous frame, and to improve prediction stability by smoothing short-term fluctuations through weighted averaging, the multi-frame weighted memory prediction model is adapted to scenes with intense motion caused by engine vibration. The output of the multi-frame weighted memory prediction model is used to output the predicted values. The signal is then transmitted to an improved Kalman filter.

[0086] The core processing logic of the multi-frame weighted memory prediction model is as follows: ① Set the frame buffer window size to window_size=3 to control the number of historical frames involved in the prediction. Experiments have verified that storing 3 frames in the frame buffer can balance prediction accuracy and computational complexity; too many frames will increase latency. ② Add the current frame luminance component image output from the Y channel to the frame buffer. If the frame buffer length exceeds window_size, the oldest frame is automatically removed. ③ Calculate the exponential decay weight of the i-th frame in the frame buffer. β is the attenuation coefficient, β>0, M is the current frame number in the frame buffer, and i is the frame number; the latest frame image has the largest weight, and the weight of historical frame images decreases with the sequence number, ensuring that the sum of the weights is 1. ); ④ Calculate the predicted value using a weighted average. , , Let be the brightness value of the i-th frame preceding time t in the frame buffer. This prediction is based on the original frame rather than the denoised frame, avoiding the propagation of residual errors (ghosting) from the previous frame. At the same time, the weighted average smooths out short-term fluctuations, improves prediction stability, and eliminates ghosting at its source.

[0087] (6) Improved Kalman filter.

[0088] The three inputs of the improved Kalman filter are connected to the outputs of the spatial adaptive Q_map construction module, the multi-frame weighted memory prediction model, and the pre-denoising module, respectively, to calculate the noise covariance Q_map of the spatial adaptive process and the predicted value. Kalman gain is calculated pixel by pixel to fuse the predicted values. The Kalman-filtered Y-channel image is obtained by comparing the pre-denoised Y-channel image (current observation value) with the pre-denoised Y-channel image, achieving pixel-level accurate noise reduction while maintaining low computational complexity (O(6N)) to meet real-time requirements. The output of the improved Kalman filter is used to output the preliminary noise reduction result, i.e., the Kalman-filtered Y-channel image, and transmit it to the high-frequency compensation module.

[0089] The core processing logic of the improved Kalman filter is as follows: ① Set the measurement noise covariance R=10.0, which represents the unreliability of the observation. Experiments have verified that a measurement noise covariance of R=10.0 can balance observation dependence and prediction dependence, avoiding over-reliance on noisy observations; ② Simplify the prediction covariance. The calculation, let =Q_map, to replace complex matrix operations and reduce computational complexity; ③ Calculate Kalman gain pixel by pixel. , , ∈[0,1], The larger the value, the more dependent it is on the current observation value; ④ Update the denoising value pixel by pixel. , , This is the predicted value for the current frame. The predicted value is the observation value of the current frame, i.e., the Y-channel image of the current frame after pre-denoising. Weighted fusion with the pre-denoised Y-channel image.

[0090] (7) High-frequency compensation module.

[0091] The two inputs of the high-frequency compensation module are connected to the outputs of the improved Kalman filter and the local variance calculation module, respectively. It receives the Kalman-filtered Y-channel image and local variance, and repairs high-frequency details smoothed during the Kalman filtering process, such as faint flame textures and subtle changes in plumes. Simultaneously, it avoids enhancing high frequencies in edge regions and noise-dense areas, thus preventing noise amplification, improving image visual clarity, and preserving key details for subsequent anomaly detection. The output of the high-frequency compensation module outputs the enhanced Y-channel image and transmits it to the channel fusion module; the enhanced Y-channel image is the final processing result for the Y-channel.

[0092] Core processing logic of the high-frequency compensation module: ① Perform 3×3 Gaussian blur on the Y-channel image after Kalman filtering to obtain the Gaussian-blurred image Y_blur, extract the high-frequency component ΔY, where ΔY = Y - Y_blur, Y is the Y-channel image after Kalman filtering, and the high-frequency component ΔY contains smoothed details and a small amount of noise; ② Generate a binary edge mask edge_mask through Canny edge detection, where 1 represents the edge area and 0 represents the non-edge area; ③ Generate a smoothing mask smooth_mask, where smooth_mask = NOT(edge_mask) AND (σ²(x) < median(σ²)), where 1 represents a flat area with low variance and non-edge, allowing high-frequency enhancement, and 0 represents the edge or high-variance area, prohibiting high-frequency enhancement; ④ Set the high-frequency enhancement gain λ = 0.3 to control the intensity of high-frequency detail enhancement and avoid noise amplification caused by over-enhancement. For pixels with a value of 1 in the smoothing mask smooth_mask, calculate the high-frequency-enhanced Y-channel image Y_enh = clip(Y_blur + λ×ΔY, 0, 255), and clip(·, 0, 255) is a clipping function used to ensure that pixel values are within the range of [0, 255].

[0093] IV. UV channel processing module.

[0094] The two input terminals of the UV channel processing module are respectively connected to the output terminals of the U channel and V channel of the color space conversion module, and are used to perform lightweight denoising on the chrominance component images output by the U channel and V channel using the bilateral filtering algorithm, remove chrominance noise (such as color shift, color band, etc.), and at the same time protect chrominance edges (such as the flame color gradient boundary) to avoid affecting the overall image quality due to chrominance noise, and use a low-complexity algorithm (O(2N)) to ensure that the overall system delay is not increased. After performing lightweight denoising on the chrominance components output by the U channel and V channel using the bilateral filtering algorithm, the bilaterally filtered U-channel image and V-channel image are obtained. The two output terminals of the UV channel processing module are respectively used to output the bilaterally filtered U-channel image and V-channel image and transmit them to the channel fusion module.

[0095] Core processing logic of the UV channel processing module: Use the bilateral filtering algorithm, considering both the spatial domain neighborhood distance and chrominance value similarity: ① Set the spatial kernel σ = 10.0 to control the neighborhood size and cover a sufficient range to suppress noise, and set the range kernel = 50.0 to control the chrominance similarity weight. The larger the value of the range kernel , the higher the tolerance for chrominance differences; ② For each pixel x, calculate the filtered output , , W is the normalization constant, is the neighborhood pixel, is the current pixel, For neighboring pixels, This is the chroma component image of the current pixel output by the U channel. This is the chroma component image of the neighboring pixels output for the U channel. The processing logic for the V channel and U channel is completely identical.

[0096] V. Channel Fusion Module.

[0097] The three inputs of the channel fusion module are connected to the output of the high-frequency compensation module and the two outputs of the UV channel processing module, respectively. This allows the module to re-merge the high-frequency enhanced Y-channel image and the bilaterally filtered U-channel and V-channel images into a complete image, which is then converted back to RGB format to obtain a denoised RGB image frame. This ensures the output image format is compatible with mainstream vision processing systems, providing a standard input for subsequent anomaly detection (such as flame anomaly recognition). The output of the channel fusion module is used to output the denoised RGB image frame and transmit it to the output module.

[0098] The core processing logic of the channel fusion module is as follows: ① Merging step: Merge the three single-channel images of Y channel, U channel and V channel according to the channel dimension to generate a YUV format image; ② Format conversion step: Convert the YUV format to RGB format through the inverse transformation of the BT.601 standard to ensure accurate color reproduction and no color distortion.

[0099] VI. Output Module.

[0100] One input of the output module connects to the output of the channel fusion module, while the other input receives the original video frame rate. This allows for the stitching of the denoised RGB format image frames into a complete denoised video at the original video frame rate. Simultaneously, the processing time for each frame is recorded, and the average processing time and frame rate are calculated to verify whether the real-time performance meets engineering requirements. Typically, the average processing time should be <1 / FPS (frame rate) to ensure smooth operation without stuttering. The output of the output module serves as the output of the entire liquid rocket engine simulation test video denoising system, outputting the denoised liquid rocket engine simulation test video file, the average processing time (avg_time_per_frame), and the average frame rate (avg_FPS = 1 / avg_time_per_frame).

[0101] The advantages of the liquid rocket engine simulation test video noise reduction system of the present invention are as follows:

[0102] (1) The Y-channel processing unit of this invention introduces a pixel-level spatial adaptive process noise covariance Q_map. Traditional Kalman filtering uses a globally fixed Q value in video denoising, ignoring the spatial heterogeneity of image content, resulting in over-smoothing in edge regions and insufficient noise suppression in flat regions. The spatial adaptive mechanism of this invention analyzes the local statistical characteristics of the current frame image and independently calculates the optimal Q value for each pixel, enabling it to trust the current observation value in flat regions (increasing the Q value to respond quickly to changes) and trust the historical prediction value in edge regions (decreasing the Q value to protect the structure), thereby achieving an optimal balance between noise suppression and detail preservation. This innovation elevates Kalman filtering from a simple temporal recursive estimator to an intelligent filter capable of perceiving image content.

[0103] (2) Traditional Kalman filtering state prediction is based on the first-order Markov assumption, relying solely on the estimate from the previous moment for linear extrapolation, which makes it difficult to accurately describe the complex motion in the original frame sequence of a liquid rocket engine simulation test video. To overcome this limitation, this invention designs a multi-frame weighted memory prediction model based on a frame buffer, using exponentially decaying weights to perform a weighted average of historical frames to generate a more accurate prior estimate. This multi-frame weighted memory mechanism has a dual advantage: on the one hand, the weighted average smooths out short-term noise fluctuations, providing a more stable state estimate than single-frame prediction; on the other hand, the exponentially decaying weights give higher importance to recent frames, enabling the prediction to better track motion trends. Therefore, the multi-frame image weighted memory prediction model based on a frame buffer significantly reduces the trailing phenomenon common in traditional Kalman filtering, with particularly noticeable performance improvement in fast-moving scenarios.

[0104] (3) This invention proposes a cascaded hybrid denoising architecture that organically combines a pre-denoising module with a multi-frame weighted memory prediction model to form a collaboratively enhanced noise suppression pipeline. The specific process is as follows: the brightness component image output by the Y channel is first processed by the pre-denoising module to generate a pre-denoised Y channel image; then, the pre-denoised Y channel image is used as the current observation value to input the improved Kalman filter for temporal filtering; finally, the Kalman-filtered Y channel image is enhanced with high frequency to restore lost details. The core advantage of this cascaded hybrid denoising architecture design is that the pre-denoising module utilizes the non-local self-similarity of the image to effectively suppress random noise within a single frame, providing high-quality observation input for subsequent temporal filtering; the improved Kalman filter utilizes the temporal correlation between frames to further smooth temporal fluctuations and maintain motion consistency. Compared with a single filtering method, this invention improves the PSNR index by an average of 2.57dB and the SSIM index by 67.9%, while keeping the computational complexity within an acceptable range.

[0105] (4) The original RGB format image frame is converted to YUV format by using a color space conversion module to achieve separation of luminance and chrominance components. This channel-based processing strategy has three advantages: First, it conforms to the characteristics of human visual perception and allocates the main computing resources to the luminance channel, which has the greatest impact on visual quality; second, bilateral filtering can effectively preserve color edges while smoothing chrominance noise and avoid color penetration; finally, this strategy significantly improves computing efficiency.

[0106] (5) In view of the unavoidable loss of detail in the time domain filtering process, the present invention designs a high-frequency compensation module for edge protection. Its core is the selective enhancement mechanism, that is, only the flat areas that are not edges are enhanced with high frequency, while important edge structures are left as is to avoid overshoot or halo effect.

[0107] Based on the aforementioned liquid rocket engine simulation test video noise reduction system, the liquid rocket engine simulation test video noise reduction method of the present invention specifically includes the following steps:

[0108] Step 1: The input module reads the original frame sequence of the liquid rocket engine simulation test video and outputs it frame by frame as an original RGB format image frame, which is then transmitted to the color space conversion module.

[0109] Step 2: The color space conversion module converts the input original RGB format image frame into YUV format, and outputs the luminance component image, blue chrominance component image, and red chrominance component image through the Y channel, U channel, and V channel, respectively.

[0110] Step 3: The UV channel processing module uses a bilateral filtering algorithm to perform lightweight denoising on the blue chroma component image output by the U channel and the red chroma component image output by the V channel, respectively, to obtain the bilaterally filtered U channel image and V channel image.

[0111] Simultaneously, the luminance component image output from the Y channel undergoes adaptive Kalman filtering via the Y channel processing unit, specifically including the following four steps:

[0112] The first step involves a pre-denoising module that performs pre-denoising on the luminance component of the Y channel output. This is achieved by calculating the denoised pixel value Y(x) using a weighted average of similar patches within a search window Ω (21×21), and then outputting the pre-denoised Y channel image based on this value. Next, the edge weight calculation module calls the EdgeWeightMap function to detect edge regions in the pre-denoised Y channel image and obtain edge weights. Simultaneously, the local variance calculation module calls the LocalVarMap function to estimate the local noise level of each pixel in the pre-denoised Y channel image and obtain the local variance. Finally, the spatial adaptive Q_map construction module combines the edge weights with the local variance calculation. We construct the pixel-level spatial adaptive process noise covariance Q_map using the local variance σ²(x) and the local variance σ²(x).

[0113] The second step involves the multi-frame weighted memory prediction model adding the current frame's luminance component image output from the Y channel to the frame buffer, and then performing predictions based on a weighted average of multiple frames in the frame buffer to obtain pixel-level predictions without residual noise. .

[0114] The third step involves the improved Kalman filter adapting to the spatial adaptive process noise covariance Q_map and the predicted value. Kalman gain is calculated pixel by pixel to fuse the predicted values. The Kalman-filtered Y-channel image is obtained by comparing the pre-denoised Y-channel image with the Y-channel image.

[0115] The fourth step involves the high-frequency compensation module receiving the Kalman-filtered Y-channel image and the local variance σ²(x), repairing the high-frequency details smoothed during the Kalman filtering process, and avoiding high-frequency enhancement in edge regions and noise-dense areas. The module then outputs the Y-channel image with enhanced high frequency, at which point the Y-channel denoising is complete.

[0116] Step 4: The channel fusion module merges the high-frequency enhanced Y channel image, the bilaterally filtered U channel image, and the V channel image back into a complete image, and converts it back to RGB format to obtain a denoised RGB format image frame.

[0117] Step 5: The output module stitches the denoised RGB format image frames together into a complete denoised video at the original video frame rate. Simultaneously, the output module records the processing time for each frame, and then calculates the average processing time and frame rate to verify whether the real-time performance meets engineering requirements.

[0118] It is worth noting that this invention allows for the adjustment of relevant parameters within a preset parameter tuning range according to actual application scenarios, thereby optimizing the noise reduction effect under different noise intensities. It also allows for local adjustments to parameters such as the frame buffer window size to adapt to video data at different frame rates, and the adjustment process does not change the core algorithm logic or overall architecture. Specifically, the adjustment range for the filter intensity parameter h is 3-15, the adjustment range for the local variance calculation window size k is 3-9, the adjustment range for the edge penalty coefficient α is 1-4, the adjustment range for the noise scale normalization factor c is 1-50, and the adjustment range for the spatial kernel σ is 5-20. The adjustment range is 30-100.

[0119] To comprehensively evaluate the performance of the liquid rocket engine simulation test video denoising method of this invention, five quality evaluation systems were used: 1) Peak Signal-to-Noise Ratio (PSNR) to evaluate pixel-level fidelity; 2) Structural Similarity (SSIM) to evaluate structural information preservation; 3) Mean Inter-Frame Difference to evaluate temporal smoothness; 4) Feature Point Stability to evaluate motion consistency; and 5) Edge Energy Variance to evaluate edge structure preservation. Based on these five quality evaluation systems, the liquid rocket engine simulation test video denoising method of this invention maintains high PSNR (average 21.3 dB) and SSIM (average 0.6861), while improving feature point stability from 0.9345 in the original video to 0.9559 and reducing the mean inter-frame difference from 21.8 to 14.6, verifying the superior performance of the algorithm in temporal consistency.

[0120] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the present invention.

Claims

1. A noise reduction system for simulated test videos of liquid rocket engines, characterized in that: It includes an input module, a color space conversion module, a Y channel processing unit, a UV channel processing module, a channel fusion module, and an output module; The input terminal of the input module is used to receive the original frame sequence of the simulated test video of the liquid rocket engine and output the original RGB format image frames frame by frame. The input end of the color space conversion module is connected to the output end of the input module. It is used to convert the original RGB format image frame into YUV format and output it in three channels. The Y channel outputs the luminance component image, the U channel outputs the blue chrominance component image, and the V channel outputs the red chrominance component image. The Y-channel processing unit includes a pre-denoising module, an edge weight calculation module, a local variance calculation module, a spatial adaptive Q-map construction module, a multi-frame weighted memory prediction model, an improved Kalman filter, and a high-frequency compensation module. The input of the pre-denoising module is connected to the output of the Y channel in the color space conversion module, and is used to perform pre-denoising processing on the luminance component image to obtain the pre-denoised Y channel image. The inputs of the edge weight calculation module and the local variance calculation module are respectively connected to the output of the pre-denoising module. The edge weight calculation module is used to detect the edge regions of the pre-denoised Y channel image and obtain the edge weights. The local variance calculation module is used to estimate the local noise level of each pixel in the pre-denoised Y channel image and obtain the local variance. The two inputs of the spatial adaptive Q_map construction module are respectively connected to the outputs of the edge weight calculation module and the local variance calculation module, and are used to combine the edge weights and local variances to construct the pixel-level spatial adaptive process noise covariance Q_map. The spatial adaptive Q_map construction module constructs the pixel-level spatial adaptive process noise covariance Q_map using the following formula: ； Where Q_base is the baseline of the process noise, For frame-level adaptive coefficients, For local variance, Edge weights; The multi-frame weighted memory prediction model includes a frame buffer for storing historical frame images of the luminance component. The input of the model is connected to the output of the Y channel in the color space conversion module. This allows the model to add the luminance component image of the current frame output from the Y channel of the color space conversion module to the frame buffer, and then perform prediction based on a weighted average of the multiple frames in the frame buffer to obtain the predicted value. The multi-frame weighted memory prediction model calculates the predicted value using the following formula. : ； in, The exponential decay weight of the i-th frame in the frame buffer. Let M be the brightness value of the i-th frame preceding time t in the frame buffer, M be the current frame number in the frame buffer, and i be the frame sequence number. The three inputs of the improved Kalman filter are respectively connected to the outputs of the spatial adaptive Q_map construction module, the multi-frame weighted memory prediction model, and the pre-denoising module, and are used to calculate the noise covariance Q_map of the spatial adaptive process and the predicted value. Kalman gain is calculated pixel by pixel to fuse the predicted values. The Kalman-filtered Y-channel image is obtained by comparing the pre-denoised Y-channel image with the Kalman-filtered image; the improved Kalman filter calculates the Kalman-filtered Y-channel image using the following formula: ； in, To update the denoised values ​​pixel by pixel, that is, the pixel values ​​of the Y channel image after Kalman filtering; This is the predicted value for the current frame. This refers to the observation value of the current frame, i.e., the Y-channel image of the current frame after pre-denoising. For Kalman gain, , ∈[0,1], To predict covariance, To measure the noise covariance; The two input terminals of the high-frequency compensation module are respectively connected to the output terminals of the improved Kalman filter and the local variance calculation module. It is used to receive the Y-channel image and local variance after Kalman filtering, and to repair the high-frequency details that were smoothed during the Kalman filtering process to obtain the Y-channel image after high-frequency enhancement. The two input terminals of the UV channel processing module are respectively connected to the output terminals of the U channel and V channel of the color space conversion module. It is used to perform lightweight denoising on the chromaticity component images output by the U channel and V channel using a bilateral filtering algorithm to obtain the bilateral filtered U channel image and V channel image. The three input terminals of the channel fusion module are respectively connected to the output terminal of the high-frequency compensation module and the two output terminals of the UV channel processing module. It is used to merge the high-frequency enhanced Y channel image, the bilaterally filtered U channel image and V channel image back into a complete image, and convert it back to RGB format to obtain a denoised RGB format image frame. One input terminal of the output module is connected to the output terminal of the channel fusion module, and the other input terminal is used to receive the original video frame rate so as to stitch the denoised RGB format image frames into a complete denoised video according to the original video frame rate.

2. The video noise reduction system for simulated test of a liquid rocket engine according to claim 1, characterized in that: The high-frequency compensation module calculates the enhanced Y-channel image using the following formula: Y_enh=clip(Y_blur+λ×ΔY,0,255); Where Y_enh is the Y channel image after high-frequency enhancement, Y_blur is the Gaussian blurred image, ΔY is the high-frequency component, λ is the high-frequency enhancement gain, and clip(·,0,255) is the clipping function used to ensure that the pixel value is within the range of [0,255].

3. The video noise reduction system for simulated test of a liquid rocket engine according to claim 2, characterized in that: The UV channel processing module calculates the bilaterally filtered U channel image using the following formula: ； in, This is the U-channel image after bilateral filtering, where W is the normalization constant. For the current pixel, For neighboring pixels, This is the chroma component image of the current pixel output by the U channel. σ is the chromaticity component image of the neighboring pixels output by the U channel; σ is the spatial kernel. For the range kernel; The calculation method for the bilaterally filtered V channel image and the bilaterally filtered U channel image in the UV channel processing module is the same.

4. The video noise reduction system for simulated test of a liquid rocket engine according to claim 1, characterized in that: The color space conversion module converts the input original RGB format image frame into YUV format through linear matrix operations, and its expression is as follows: ； Where Y is the luminance component, U is the blue chromaticity component, V is the red chromaticity component, and R, G, and B are the red, green, and blue primary color components, respectively.

5. The video noise reduction system for simulated test of a liquid rocket engine according to claim 4, characterized in that: While performing pre-denoising processing on the luminance component image, the pre-denoising module utilizes the weighted average characteristic of similar patches to initially retain key details and obtain the pre-denoised Y channel image.

6. The video noise reduction system for simulated test of a liquid rocket engine according to claim 5, characterized in that: The edge weight calculation module calculates edge weights based on the Laplace operator and the exponential decay function; The local variance calculation module calculates the local variance based on the box filter of the local variance calculation window.

7. A method for noise reduction in simulated test videos of liquid rocket engines, based on the noise reduction system for simulated test videos of liquid rocket engines as described in any one of claims 1-6, characterized in that, Includes the following steps: Step 1: The input module reads the original frame sequence of the liquid rocket engine simulation test video and outputs the original RGB format image frames frame by frame, which are then transmitted to the color space conversion module. Step 2: The color space conversion module converts the input original RGB format image frame into YUV format, and outputs the luminance component image, blue chrominance component image and red chrominance component image through the Y channel, U channel and V channel respectively. Step 3: The UV channel processing module uses a bilateral filtering algorithm to perform lightweight denoising on the blue chroma component image output by the U channel and the red chroma component image output by the V channel, respectively, to obtain the bilaterally filtered U channel image and V channel image. Simultaneously, the brightness component image output from the Y channel undergoes adaptive Kalman filtering via the Y channel processing unit, specifically including: The first step involves the pre-denoising module performing pre-denoising on the luminance component of the Y channel output to obtain the pre-denoised Y channel image; the edge weight calculation module then detects the edge regions of the pre-denoised Y channel image and calculates the edge weights. Simultaneously, the local variance calculation module estimates the local noise level of each pixel in the pre-denoised Y-channel image and calculates the local variance σ²(x); the spatial adaptive Q-map construction module combines edge weights... And the local variance σ²(x) are used to construct the pixel-level spatial adaptive process noise covariance Q_map; The second step involves the multi-frame weighted memory prediction model adding the current frame's luminance component image output from the Y channel to the frame buffer, and then performing predictions based on a weighted average of multiple frames in the frame buffer to obtain pixel-level predictions without residual noise. ; The third step involves the improved Kalman filter adapting to the spatial adaptive process noise covariance Q_map and the predicted value. Kalman gain is calculated pixel by pixel to fuse the predicted values. The Kalman-filtered Y-channel image is obtained by comparing the pre-denoised Y-channel image with the Y-channel image. The fourth step involves the high-frequency compensation module receiving the Kalman-filtered Y-channel image and local variance σ²(x), repairing the high-frequency details smoothed during the Kalman filtering process, and avoiding high-frequency enhancement in edge regions and noise-dense areas, while outputting the high-frequency enhanced Y-channel image. Step 4: The channel fusion module merges the high-frequency enhanced Y channel image, the bilaterally filtered U channel image, and the V channel image back into a complete image, and converts it back to RGB format to obtain a denoised RGB format image frame. Step 5: The output module stitches the denoised RGB format image frames together into a complete denoised video at the original video frame rate.

Images