Video noise reduction method and device, medium and equipment

Abstract

The embodiment of the invention provides a video noise reduction method and device, a medium and equipment, and the method comprises the steps: determining the inter-frame motion intensity of an original video, constructing a key frame group of the original video, and constructing the key frame group of the original video according to the inter-frame motion intensity. According to the key frame group, obtaining dynamic channel characteristics for motion noise reduction through frequency domain conversion and channel attention weighting; static global features of space details in the original frame are determined and reserved; extracting semantic features of the original frame; and according to the semantic feature of each original frame, the static global feature of each original frame and the dynamic channel feature, obtaining a noise-reduced video through feature fusion. According to the method, through inter-frame motion intensity, key frame group construction, dynamic and static dual-feature collaborative noise reduction and semantic feature fusion, the original video after noise reduction has no motion blur and ensures space details.

Description

Technical Field

[0001] This application belongs to the field of computer technology, and in particular relates to a video noise reduction method, apparatus, medium and device. Background Technology

[0002] With the development of video technology, video noise has become a key factor affecting video quality. For example, historical media videos suffer from problems such as graininess, scratches, and analog interference due to factors such as analog equipment, film-to-digital conversion, and magnetic tape storage; new information sources such as drones, mobile phone shooting, and live streaming suffer from digital noise such as compression block effect and transmission errors.

[0003] Currently, temporal filtering methods are generally used for video denoising. These methods utilize the randomness of noise through inter-frame averaging filtering. However, this temporal filtering method assumes that the scene in the video is static or that there is global motion. When there are local moving targets in the video, it will misjudge the motion edges as noise, thus blurring the motion area and resulting in poor adaptability to moving scenes. Summary of the Invention

[0004] This application provides a video noise reduction method, apparatus, medium, and device that can be adapted to different scenarios and ensure the noise reduction accuracy of the video.

[0005] In a first aspect, embodiments of this application provide a video noise reduction method, the method comprising: Determine the inter-frame motion intensity of the original video; Based on the inter-frame motion intensity, construct the keyframe group of the original video; Based on the keyframe group, dynamic channel features for motion noise reduction are obtained through frequency domain transformation and channel attention weighting. For each original frame in the original video, determine the static global features that preserve the spatial details of that original frame; Extract the semantic features of each of the original frames; Based on the semantic features of each original frame, the static global features of each original frame, and the dynamic channel features, a noise-reduced video is obtained through feature fusion.

[0006] In one feasible implementation, determining the inter-frame motion intensity of the original video specifically includes: The forward and backward optical flow fields of the original video are determined using a lightweight optical flow network. For each pixel position in the original video, the inter-frame displacement of that pixel position is determined based on the forward optical flow field and the backward optical flow field; The inter-frame motion intensity of the original video is determined based on the inter-frame displacement of each pixel position.

[0007] In one feasible implementation, the keyframe group of the original video is constructed based on the inter-frame motion intensity, specifically including: Multiple preset frame-skipping intervals and multiple mutual exclusion threshold ranges are set, wherein the frame-skipping interval and the mutual exclusion threshold range have a one-to-one correlation, and the frame-skipping interval and the mutual exclusion threshold range have a negative correlation. The mutual exclusion threshold range into which the inter-frame motion intensity falls is taken as the matching threshold range; Based on the aforementioned correlation, the frame extraction interval corresponding to the matching threshold range is taken as the target inter-frame interval; Based on the target frame extraction interval, construct the keyframe group of the original video.

[0008] In one feasible implementation, based on the keyframe group, dynamic channel features for inter-frame noise reduction are obtained through frequency domain transformation and channel attention weighting, specifically including: The video frequency domain tensor of the keyframe group is obtained through frequency domain transformation; Noise separation is performed on the video frequency domain tensor to obtain a sparse signal-to-noise mask; Determine the channel attention weights of the sparse signal-to-noise mask; Based on the channel attention weights, the sparse signal-to-noise mask is subjected to channel attention weighting to obtain the target signal-to-noise mask; Based on the target signal-to-noise mask and the video frequency domain tensor, the dynamic channel features are obtained through time-domain transformation.

[0009] In one feasible implementation, for each original frame in the original video, static global features that preserve spatial details in that original frame are determined, specifically including: By performing frequency domain transformation, the single-frame frequency domain tensor of each original frame is obtained; For each original frame, the spectral coefficients of the frequency domain tensor of the single frame are sorted and divided according to the magnitude of the frequency domain amplitude to obtain multiple mutually exclusive subsets. Then, by performing time-domain transformation on the multiple mutually exclusive subsets, a multi-branch convolution kernel set is obtained. For each of the original frames, frequency domain modulation features are generated based on the original frame and the multi-branch convolution kernel set, and the global attention weights of the frequency domain modulation features are determined. For each of the original frames, the frequency domain modulation features are globally attention-weighted according to the global attention weights to obtain the target modulation features; For each original frame, the target modulation feature is transformed in the time domain and then fused with the original frame through residual connection to obtain the static global feature.

[0010] In one feasible implementation, a denoised video is obtained by feature fusion based on the semantic features of each original frame, the static global features of each original frame, and the dynamic channel features, specifically including: For each of the original frames, the original frame is downsampled to obtain semantic features; For each original frame, the static global features and dynamic channel features of the original frame are concatenated along the channel dimension to generate a fused feature; For each of the original frames, a fusion coefficient is obtained based on the semantic features, the inter-frame motion intensity, and the fusion features; For each original frame, the semantic features and the fusion features are weighted and upsampled according to the fusion coefficient to obtain a denoised frame; Based on each of the denoised frames, a denoised video is obtained.

[0011] In one feasible implementation, for each original frame, a fusion coefficient is obtained based on the original frame, the inter-frame motion intensity, and the fusion feature, specifically including: The inter-frame motion intensity in the scalar dimension is extended to the spatial dimension of the semantic features to obtain the motion intensity tensor; The fusion coefficients are obtained by concatenating the motion intensity tensor, the semantic features, and the fusion features along the channel dimension and then mapping them through the channels.

[0012] In one feasible implementation, the method further includes: Visual features, audio features, and text features of the original video are extracted, and the index feature vector of the original video is obtained through feature fusion and feature vectorization. Obtain the video parameter metadata of the original video; The noise-reduced video is stored and managed based on the video parameter metadata and the index feature vector.

[0013] In one feasible implementation, extracting the visual features of the original video specifically includes: Based on the keyframe group, the content type features of the original video are determined through feature extraction and category recognition; Based on the lighting conditions of the original video, determine the characteristics of the lighting environment; Based on the inter-frame displacement of each original frame in the original video, determine the motion type features; The visual features are obtained by concatenating the content type features, the lighting environment features, and the motion type features.

[0014] Secondly, embodiments of this application provide a video noise reduction device, the device comprising: The motion intensity module is used to determine the inter-frame motion intensity of the original video; The keyframe module is used to construct a keyframe group of the original video based on the inter-frame motion intensity. The dynamic noise reduction module is used to obtain dynamic channel features for motion noise reduction based on the key frame group through frequency domain transformation and channel attention weighting. The static noise reduction module is used to determine, for each original frame in the original video, static global features that preserve the spatial details of that original frame; The semantic feature module is used to extract semantic features from each of the original frames; The noise reduction video module is used to obtain a noise reduction video by feature fusion based on the semantic features of each original frame, the static global features of each original frame, and the dynamic channel features.

[0015] Thirdly, embodiments of this application provide a video noise reduction device, the device including: a processor, and a memory storing computer program instructions; the processor reads and executes the computer program instructions to implement any of the above-described video noise reduction methods.

[0016] Fourthly, embodiments of this application provide a computer storage medium storing computer program instructions, which, when executed by a processor, implement any of the above-described video noise reduction methods.

[0017] Fifthly, the embodiments of this application provide a computer program product, wherein when the instructions in the computer program product are executed by the processor of an electronic device, the electronic device performs any of the above-mentioned video noise reduction methods.

[0018] This application discloses a video denoising method, apparatus, medium, and device that can construct a keyframe group of the original video by determining the inter-frame motion intensity, and then construct the keyframe group of the original video based on the inter-frame motion intensity. Based on the keyframe group, dynamic channel features for motion denoising are obtained through frequency domain transformation and channel attention weighting; static global features that preserve spatial details in the original frames are determined; semantic features of the original frames are extracted; and denoised video is obtained through feature fusion based on the semantic features of each original frame, the static global features of each original frame, and the dynamic channel features. In other words, this application constructs a keyframe group through inter-frame motion intensity, and the synergistic denoising of dynamic and static features, along with the fusion of semantic features, ensures that the denoised original video has no motion blur and preserves spatial details.

[0019] Furthermore, the inter-frame motion intensity and frame extraction interval are bound together by a mutually exclusive threshold range. The higher the motion intensity, the higher the threshold range, and the smaller the frame extraction interval; the lower the motion intensity, the lower the threshold range, and the larger the frame extraction interval. This allows keyframe groups to be extracted densely in high-motion scenes, preserving more motion details to support motion noise reduction; and to be extracted sparsely in static or low-motion scenes, reducing redundant calculations and improving processing efficiency. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic flowchart of a video noise reduction method provided in an embodiment of this application; Figure 2 This is a schematic diagram illustrating the principle of obtaining a denoised frame through feature fusion, as provided in an embodiment of this application. Figure 3 This is a schematic diagram of the architecture of a multilayer perceptron provided in an embodiment of this application; Figure 4 This is a schematic diagram illustrating the principle of video type feature extraction provided in the embodiments of this application; Figure 5 This is a schematic diagram of the media asset video management process provided in the embodiments of this application; Figure 6 This is a schematic diagram of the structure of a video noise reduction device provided in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of a video noise reduction device provided in an embodiment of this application. Detailed Implementation

[0022] The features and exemplary embodiments of various aspects of this application will be described in detail below. To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain this application and not to limit it. For those skilled in the art, this application can be implemented without some of these specific details. The following description of the embodiments is merely to provide a better understanding of this application by illustrating examples.

[0023] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0024] With the development of video technology, video noise has become a key factor affecting video quality. For example, historical media videos suffer from problems such as graininess, scratches, and analog interference due to factors such as analog equipment, film-to-digital conversion, and magnetic tape storage; new information sources such as drones, mobile phone shooting, and live streaming suffer from digital noise such as compression block effect and transmission errors.

[0025] Currently, temporal filtering methods are generally used for video denoising. These methods utilize the randomness of noise through inter-frame averaging filtering. However, this temporal filtering method assumes that the scene in the video is static or that there is global motion. When there are local moving targets in the video, it will misjudge the motion edges as noise, thus blurring the motion area and resulting in poor adaptability to moving scenes.

[0026] To address the problems in the prior art, embodiments of this application provide a video noise reduction method, apparatus, device, and computer storage medium.

[0027] In practical applications, the execution subject of the video noise reduction method in this application embodiment can be a terminal device, such as a desktop computer or laptop computer, or a remote device like a server. Of course, this application embodiment can also adopt an execution subject in the form of software, such as a client or software program installed on a terminal device. The specific type of execution subject corresponding to the technical solution provided in this application embodiment is not strictly limited here, and can be flexibly selected according to the actual application scenario and actual needs.

[0028] The following describes specific embodiments of a video noise reduction method, apparatus, medium, and device provided in this application. First, a video noise reduction method is introduced.

[0029] Figure 1 A schematic flowchart of a video noise reduction method according to an embodiment of this application is shown. Figure 1 As shown, the method includes steps S100 to S105: S100: Determine the inter-frame motion intensity of the original video.

[0030] In one or more embodiments of this application, in order to implement a dynamic adaptive frame-skipping strategy based on motion intensity in subsequent steps—to suppress motion blur by shortening the frame-skipping interval in high-motion scenes and to save computational power by extending the interval in low-motion scenes—this application needs to determine the inter-frame motion intensity of the original video by quantifying the severity of the inter-frame motion.

[0031] Specifically, this application can perform grayscale processing on adjacent original frames of the original video, and then calculate the absolute grayscale difference matrix between adjacent original frames. A difference threshold is set, and pixels in the absolute grayscale difference matrix that are greater than the threshold are marked as moving pixels. The inter-frame motion intensity is determined based on the proportion of moving pixels to the total number of pixels; or the mean / variance of the absolute grayscale difference matrix is ​​determined as the inter-frame motion intensity.

[0032] It should be noted that this application does not restrict the source of the original video and can be set according to actual needs. For example, the original video can be internet media asset video from an internet platform; it can also connect with the license holder and the content distribution network of each province through the HTTP protocol and the back-to-origin strategy, supporting multiple format protocols such as TS / MP4 / HLS to pull media asset metadata and bitstream, and the video management system can uniformly obtain internet media assets, network-wide media assets, and provincial media assets to form the original video input source. This application does not restrict the method of determining the inter-frame motion intensity and can be set according to actual needs, such as motion vector statistics based on fast matching, that is, dividing the original frame into image blocks of fixed size, finding matching image blocks in adjacent original frames through fast matching, and then determining the inter-frame motion intensity by calculating the block displacement vector between the matching image blocks; or determining the inter-frame motion intensity based on feature point matching; or motion intensity prediction based on deep learning. In one or more embodiments of this application, forward and backward optical flow consistency can be used to effectively distinguish real motion from interference signals such as illumination and noise, thereby improving the accuracy of inter-frame motion intensity. This includes: firstly, determining the forward and backward optical flow fields of the original video using a lightweight optical flow network; secondly, determining the inter-frame displacement of each pixel position in the original video based on the forward and backward optical flow fields; and finally, determining the inter-frame motion intensity of the original video based on the inter-frame displacements of each pixel position. Determining the inter-frame motion intensity based on this embodiment includes: In equations (1), (2), and (3), the original video is , The optical flow vector at a single pixel location. This represents the pixel coordinates of a specific pixel location within the original video frame. This represents the horizontal displacement component of the pixel position. This represents the vertical displacement component of the pixel position. It represents a two-dimensional vector consisting of displacement components in the horizontal direction and displacement components in the vertical direction. Represents the L1 norm per pixel. Inter-frame motion intensity is used to characterize the degree of motion in the original video. The higher the value, the more intense the exercise. Let be the backward optical flow field, representing the optical flow field from frame t to frame t-1; Let be the forward optical flow field, representing the optical flow field from frame t to frame t+1. It is a three-dimensional tensor with the shape of , and These represent the height and width of the original video, respectively, for any number of pixels. The two-dimensional vector is used to calculate the inter-frame displacement between adjacent original frames. The global motion intensity is quantized by checking the consistency of the forward / backward optical flow (if there is illumination or noise interference, the bidirectional displacement is usually inconsistent and will be weakened after summation) to obtain the inter-frame motion intensity of the original video.

[0033] S101: Construct the keyframe group of the original video based on the inter-frame motion intensity.

[0034] In one or more embodiments of this application, in order to accurately extract the dynamic channel features of motion adaptation in subsequent steps, it is necessary to ensure that the number of keyframes is sufficient to support inter-frame temporal consistency analysis, while avoiding keyframe redundancy that would lead to wasted computational power and reduced efficiency in feature extraction. In this step, this application can dynamically adapt the frame extraction interval of keyframes according to the inter-frame motion intensity to construct a keyframe group of the original video.

[0035] Specifically, this application can determine the target frame extraction interval based on the inter-frame motion intensity and the preset frame extraction interval, and then select key frames according to the target frame extraction interval to construct the key frame group of the original video.

[0036] It should be noted that this application does not limit the specific method of determining the target frame extraction interval based on inter-frame motion intensity, and can be set according to actual needs. For example, a step-by-step adaptation can be achieved by pre-setting multiple sets of correspondences between frame extraction intervals and intensity thresholds; or a deep learning model can be used to fit the mapping relationship between motion intensity and frame extraction intervals. In one or more embodiments of this application, multiple preset frame extraction intervals and multiple mutually exclusive threshold ranges can be set, with a one-to-one correlation between the frame extraction interval and the mutually exclusive threshold range, and a negative correlation between the preset frame extraction interval and the mutually exclusive threshold range; the mutually exclusive threshold range into which the inter-frame motion intensity falls is used as the matching threshold range; according to the correlation, the frame extraction interval corresponding to the matching threshold range is used as the target inter-frame interval; and the keyframe group of the original video is constructed according to the target frame extraction interval. The coverage areas of each mutually exclusive threshold range do not overlap and are mutually exclusive.

[0037] The reference formula for determining the target frame extraction interval based on inter-frame motion intensity is as follows: In equation (4), and This is the boundary value of the mutually exclusive threshold range. Less than The mutual exclusion threshold range includes , as well as “1, 4, 2” represent the frame extraction intervals corresponding to each mutual exclusion threshold range. The target frame extraction interval is given. The values ​​in formula (4) are for reference only. In this application, the specific content of the frame extraction interval and the mutual exclusion threshold range is not limited and can be set according to actual needs. After determining the target frame extraction interval, the keyframe group of the original video can be constructed based on the target frame extraction interval, as follows: In equation (5), For keyframe groups; These are the original video frames.

[0038] S102: Based on the keyframe group, dynamic channel features for motion noise reduction are obtained through frequency domain transformation and channel attention weighting.

[0039] In one or more embodiments of this application, features designed to enhance temporal consistency for motion region noise reduction in subsequent steps are provided. In this step, the application analyzes the frequency domain correlation information between keyframes through frequency domain transformation based on keyframe groups to highlight effective motion signals and suppress noise interference. Then, channel attention weighting is applied to accurately match the noise reduction requirements of different motion intensity scenarios, thereby obtaining dynamic channel features for motion noise reduction.

[0040] Specifically, this application can perform frequency domain transformation (such as Fourier transform or wavelet transform) on each keyframe in the keyframe group, mapping the frame information in the spatial domain to the frequency domain. Then, through a channel attention mechanism, the feature importance of different channels is analyzed and weighted to enhance the representation of motion-related frequency domain components (such as high-frequency motion signals), weaken noise-dominated redundant components, and integrate to obtain dynamic channel features for focused motion noise reduction.

[0041] It should be noted that this application does not limit the specific method for determining dynamic channel features, and can be set according to actual needs. For example, short-time Fourier transform can be used to achieve frequency domain transformation and combined with an attention module to calculate weights; or the frequency domain tensor can be decomposed by wavelet transform and then channel weighting can be performed using an attention mechanism; or a Transformer-based self-attention mechanism can be directly used to replace channel attention to capture the frequency domain correlation information of key frame groups. In one or more embodiments of this application, in order to improve the accuracy of frequency domain noise separation and enhance the motion adaptability of features, this application can obtain dynamic channel features through a frequency domain and spatial domain linkage processing method, including: first, by frequency domain transformation, the video frequency domain tensor of key frame groups is obtained, and then noise separation is performed on the video frequency domain tensor to obtain a sparse signal-to-noise mask; second, the channel attention weights of the sparse signal-to-noise mask are determined, and the sparse signal-to-noise mask is weighted by channel attention according to the channel attention weights to obtain a target signal-to-noise mask; finally, the dynamic channel features are obtained by time domain transformation based on the target signal-to-noise mask and the video frequency domain tensor.

[0042] Continuing with the previous example, the process of determining the sparse signal-to-noise mask through frequency domain sparse representation learning includes: converting the spatial and temporal pixel information into a video frequency domain tensor by performing a three-dimensional discrete Fourier transform on the keyframe group along the temporal dimension. 3 represents the number of keyframes; a lightweight 3D U-Net is constructed, containing only two 3×3×3 convolutions, with the input... Output sparse signal-to-noise mask .

[0043] The channel attention acquisition process includes: in the sparse SNR mask, a mask position of 1 indicates that the corresponding frequency domain coefficients are preserved, corresponding to the real signal frequency band; a mask position of 0 indicates that the corresponding frequency domain coefficients are suppressed, corresponding to the noise frequency band. Global average pooling is performed on the sparse SNR mask of each channel in the spatial dimension to obtain a channel description vector of length 3. Then, the channel dimension is reduced through two fully connected layers, and after a nonlinear transformation by an activation function, it is increased back to 3 channels. Finally, softmax activation is applied to generate three channel attention weights that sum to 1. The larger the weight of this channel attention, the higher the reliability of the sparse SNR mask of the corresponding frame (frames with high motion consistency have higher weights).

[0044] The process of generating dynamic channel features through temporal consistency reconstruction includes: based on the channel attention weights, performing channel attention weighting on the sparse signal-to-noise mask to obtain the target signal-to-noise mask; multiplying the target signal-to-noise mask element-wise with the video frequency domain tensor to filter noise frequency bands and retain effective signals; and then obtaining the dynamic channel features through temporal transformation (three-dimensional inverse Fourier transform). The formula for determining this example is as follows: In equation (6), This is a dynamic channel feature; This is element-wise multiplication; For video frequency domain tensors; This is a sparse signal-to-noise mask. Specifically, the sparse signal-to-noise mask is optimized by channel attention weights to obtain the target signal-to-noise mask. Then, based on this target signal-to-noise mask, the video frequency domain tensor can be filtered to obtain dynamic channel features focused on temporal motion consistency.

[0045] This application captures inter-frame motion correlation information by using motion intensity-adaptive keyframe groups combined with 3D Fourier transform. Then, sparse signal-to-noise masking accurately separates real motion signals from noise signals, avoiding motion blur caused by inter-frame signal misalignment. This application eliminates the need for manually designed frequency domain basis functions; it learns frequency domain sparse representations end-to-end using a lightweight 3D U-Net, adapting to different video motion patterns and noise distributions to generate filtering strategies more closely suited to real-world scenarios, demonstrating strong adaptability. The lightweight network structure also reduces feature extraction time, motion-adaptive frame extraction reduces redundant keyframes, and channel attention weighting focuses on highly reliable features, significantly improving processing efficiency while maintaining noise reduction accuracy, thus meeting real-time video noise reduction requirements.

[0046] S103: For each original frame in the original video, determine the static global features that preserve the spatial details in that original frame.

[0047] In one or more embodiments of this application, in order to provide spatial detail for noise reduction of stationary or weakly moving regions in subsequent steps, this application may determine static global features that preserve spatial details in each original frame of the original video.

[0048] Specifically, this application first performs frequency domain transformation on each original frame to extract the frequency component distribution of a single original frame; then, it strengthens the representation of spatial details (such as texture and edge) through a global attention mechanism, suppresses noise interference, and generates static global features that preserve spatial integrity, ensuring that the texture and details of static areas are not lost after noise reduction.

[0049] It should be noted that this application does not limit the method of determining static global features, and can be set according to actual needs. For example, Gaussian pyramid decomposition can be used to extract multi-scale static features and combined with global pooling to obtain global information; or a non-local attention mechanism can be used to capture long-distance spatial correlations within a single frame. In one or more embodiments of this application, in order to accurately separate the signal and noise in the frequency domain of a single original frame and enhance spatial details, this application can determine static global features by combining frequency domain modulation and global attention, including: First, by frequency domain transformation, the single-frame frequency domain tensor of each original frame is obtained. Second, according to the magnitude of the frequency domain amplitude, the spectral coefficients of the single-frame frequency domain tensor are sorted and divided to obtain multiple mutually exclusive subsets, and by performing time domain transformation on the multiple mutually exclusive subsets, a multi-branch convolution kernel set is obtained. Then, for each original frame, based on the original frame and the multi-branch convolution kernel set, frequency domain modulation features are generated, and the global attention weights of the frequency domain modulation features are determined. Finally, for each original frame, the frequency domain modulation features are globally attention-weighted according to the global attention weights to obtain the target modulation features. After the target modulation features are transformed in the time domain, they are fused with the original frame through residual connections to obtain static global features.

[0050] For example, the process of determining the frequency domain tensor of a single frame includes: the original frames in the original video. (resolution is) The original frame (with 3 RGB channels) is mapped to a high-dimensional feature space using a 1*1 convolution, outputting a high-dimensional feature tensor with dimensions of 1. ( The preset number of channels ranges from 32 to 128 (in this application, C can be set to 64). This high-dimensional feature tensor is rearranged along the channel dimension into a two-dimensional complex tensor, resulting in a single-frame frequency domain tensor. (HC=H*C, WC=W*C), each element in this single-frame frequency domain tensor Corresponding frequency domain coordinates The spectral coefficients are used to complete the high-dimensional feature mapping from the spatial domain to the frequency domain.

[0051] The process of determining the multi-branch convolution kernel set includes: processing the frequency domain tensor of a single frame. Calculate the frequency domain amplitude of all spectral coefficients, sort them in ascending order according to their frequency domain amplitude to obtain an ordered index list. and divide it into Mutually exclusive subsets In which each mutually exclusive subset satisfies (i, j are positive integers not greater than n); perform an inverse two-dimensional discrete Fourier transform on each mutually exclusive subset to obtain the corresponding spatial kernel. ,in, Indicates the kernel size (e.g., 3*3). All By concatenating along the output channel dimension, a multi-branch convolutional kernel set is obtained.

[0052] The process of determining frequency domain modulation features through frequency domain and spatial modulation includes: taking the original frame as input, processing it through convolutional layers and activation functions to generate a pixel-level frequency domain modulation matrix. ,in, Indicates position The noise reduction strength coefficient; Expand along the channel dimension The dimension is obtained by performing element-wise Hadamard multiplication on the result of convolving the original frame with each branch of a multi-branch convolution kernel set to obtain the corresponding high-frequency regions (edges, textures). When the value approaches 0, weak noise reduction is performed to avoid blurring of details; low-frequency regions (flat background) correspond to When the value approaches 1, strong denoising is performed to suppress particle noise; the frequency domain modulation features are obtained by summing the results of all modulated branch convolutions along the channel dimension. .

[0053] The global attention extraction process includes: inputting the frequency domain modulation features into the global attention module to perform global average pooling on the frequency domain modulation features in the spatial dimension, obtaining a one-dimensional channel statistical vector of length C, and then generating channel attention weights through two fully connected layers (dimensionality reduction → ReLU activation → dimensionality increase). After that and By multiplying each channel sequentially, adaptive weighting of the channel dimension is achieved to strengthen effective feature channels and weaken redundant noise channels.

[0054] The process of determining static global features includes: performing a two-dimensional inverse Fourier transform on the frequency domain modulation features after global attention weighting to obtain the spatial domain noise reduction residual. By using residual connections, the spatial domain denoising residual is added pixel-by-pixel to the original frame to output static global features. This static global feature can avoid color shift and brightness drift during the noise reduction process, and preserve the spatial details of the original frame.

[0055] This application fully leverages the computational advantages of frequency domain convolution by migrating the extraction process of static global features to the frequency domain: compared to spatial domain convolution, which requires point-by-point multiplication and addition operations between each pixel and the convolution kernel (computational complexity of O(N²K²), where N is the image size and K is the convolution kernel size), frequency domain processing only requires one Fourier transform, element-by-element multiplication, and inverse Fourier transform, reducing the complexity to O(N²logN) and improving computational efficiency. Meanwhile, frequency domain convolution can quickly capture the global frequency features of an image. By using multi-branch convolution kernel sets to process different frequency domain subsets, it can accurately separate high-frequency details (edges, textures) from low-frequency noise. Combined with pixel-level modulation matrices, it achieves differentiated processing of "weak noise reduction for high frequencies and strong noise reduction for low frequencies". Then, it strengthens the effective channel features through global attention. Finally, the static global features output through residual connections not only completely preserve the spatial details of the original frame, but also avoid color shift and brightness drift during the noise reduction process. It complements the dynamic channel features in terms of "static detail fidelity and dynamic motion continuity", and together improves the noise reduction effect of videos in all scenes.

[0056] S104: Extract the semantic features of each of the original frames.

[0057] In one or more embodiments of this application, in order to provide an adaptive weighting basis for the fusion of dynamic channel features and static global features in subsequent steps and avoid information loss caused by excessive noise reduction of key semantic regions, this application may extract semantic features of each original frame in this step.

[0058] Specifically, this application can downsample each original frame to obtain a low-resolution frame, and then process it through convolutional layers and activation functions to output semantic features that match the size of the low-resolution frame. This application does not limit the specific method of extracting semantic features and can set it according to actual needs. For example, a pre-trained object detection model can be used to extract semantic features.

[0059] S105: Based on the semantic features of each original frame, the static global features of each original frame, and the dynamic channel features, a noise-reduced video is obtained through feature fusion.

[0060] In one or more embodiments of this application, in order to output a denoised video that balances motion continuity, spatial detail, and semantic fidelity, this application can obtain denoised frames with no motion blur in moving areas, preserved spatial detail in static areas, and unambiguous semantics through multi-dimensional feature collaboration and adaptive weighted fusion. In this step, this application needs to combine the semantic features of each original frame, the static global features of each original frame, and the dynamic channel features to obtain the denoised video.

[0061] Specifically, this application performs convolution and average pooling on the original frame to extract semantic features. Then, based on these semantic features, dynamic channel features (temporal consistency), and static global features (frequency domain noise reduction), the features are concatenated along the channel dimension. After normalization and convolution, the GELU activation function is used to maintain feature sparsity and filter redundant noise. The concatenated features and semantic features are then input into a multilayer perceptron to obtain denoised frames. A denoised video is obtained from each denoised frame.

[0062] It should be noted that this application does not limit the specific method of obtaining the denoised video. The method can be set according to actual needs, such as using a Transformer structure to model long-distance dependencies of features before fusion, or achieving cross-level fusion through a multi-scale feature pyramid structure. In one or more embodiments of this application, to enhance the accuracy of semantic features and motion information in the fusion process and avoid fusion conflicts between static and dynamic features in complex scenes, this application introduces pixel-level dynamically adjusted fusion coefficients for adaptive noise reduction, including: First, for each original frame, downsampling the original frame to obtain semantic features; then, concatenating the static global features and dynamic channel features of the original frame along the channel dimension to generate fused features. Second, for each original frame, fusion coefficients are obtained based on semantic features, inter-frame motion intensity, and fusion features; and based on the fusion coefficients, the semantic features and fused features are weighted and upsampled to obtain denoised frames. Finally, a denoised video is obtained based on each of the denoised frames.

[0063] In one or more embodiments of this application, the inter-frame motion intensity in the scalar dimension can be extended to the spatial dimension of semantic features to obtain a motion intensity tensor; based on the motion intensity tensor, semantic features, and fusion features, the fusion coefficients are obtained by concatenating along the channel dimension and then mapping through the channel.

[0064] Continuing with the previous example, we downsample the original frame to obtain low-resolution semantic features. ,in, For the preset number of channels, the inter-frame motion intensity in the scalar dimension is... Expand to isomorphic one-dimensional tensors, and The concatenation is performed along the channel dimension, and then the concatenated tensor is mapped to a single-channel feature map through a pointwise convolutional layer Conv1×1. Finally, the fusion coefficients are output through a sigmoid activation function. .

[0065] In equations (7) and (8), This is an example of the sigmoid activation function in this application, used to map any real number input to... The interval is where the output of the pointwise convolutional layer is normalized. `downsample(I)` t) ) represents semantic features; concat(*) is the feature concatenation function. For Gaussian error linear units, The weighted sum of the i-th neuron in this layer. The cumulative distribution function of the standard normal distribution. is the scaling factor of the Gaussian CDF.

[0066] Figure 2 This is a schematic diagram illustrating the principle of obtaining a denoised frame through feature fusion, as provided in an embodiment of this application. Figure 2 As shown, this application fuses fusion features and semantic features based on fusion coefficients to obtain denoised frames after motion denoising and static denoising, as shown in the following formula: In equation (9), Represents the characteristics of integration; This indicates a high-resolution residual branch, which can be set according to actual needs and can be omitted. Upsampling is used to restore the full-resolution output, generating pixel-level fusion weights in one go, reducing computation and noise interference; To achieve downsampling, the current frame is aligned with the low-resolution noise reduction features. Simultaneously, convolution and sigmoid activation are applied to the dynamic channel features to generate weights for enhancing temporal consistency. These weights are then multiplied with the dynamic channel features to obtain temporally consistent features. Max pooling and average pooling are performed on the static global features to obtain two different feature representations. These two feature representations are then input into a multilayer perceptron to generate the first feature vector. Second eigenvector The two feature vectors are then concatenated via channels and activated by sigmoid to generate frequency domain denoising weights. These weights are then multiplied by the static global features to obtain the frequency domain denoising features. The temporal consistency feature and frequency domain denoising feature processing methods in this application ensure that β approaches 1 in large motion regions, prioritizing temporal information and reducing motion blur; β approaches 0 in static regions, prioritizing static denoising results and preserving texture; β is adaptively updated frame by frame, eliminating the need for manual adjustment.

[0067] Figure 3 This is a schematic diagram of the architecture of the multilayer perceptron provided in this application. Figure 3As shown, through a multilayer perceptron, the feature vectors of the extracted temporal consistency features and frequency domain noise reduction features can be concatenated along the channel dimension. Feature mapping is performed in the first fully connected layer, mapping the 2C-dimensional features to H-dimensional features, where H < 2C, to compress redundant information. A nonlinear activation layer is then added for nonlinearization. In the second fully connected layer, the weight matrix W ∈ RH × H and the bias b ∈ RH are shared by the first fully connected layer and the downstream task. A second nonlinear activation layer is then added to maintain sparse activation. Finally, a fully connected layer is added to concatenate the features according to the channel dimension.

[0068] In the above method, dynamic noise reduction and static noise reduction are performed in parallel. Furthermore, the weights of temporal consistency features or frequency domain noise reduction features are adaptively increased based on motion type characteristics, eliminating the need for manual parameter tuning. Simultaneously, this application shortens the sampling interval in high-motion scenes by adaptively adjusting the keyframe extraction interval, thereby avoiding motion blur. In static noise reduction, the frequency domain replaces the spatial domain, and noise separation is achieved through Fourier domain parameter decoupling and dynamic frequency domain convolution, unaffected by prior noise distribution, thus adapting to different video types.

[0069] Furthermore, this application can store and manage denoised videos using a feature vectorization approach. Specifically, by extracting multi-dimensional features from the original video and generating indexed feature vectors, orderly management of the denoised videos can be achieved, as detailed below: In one or more embodiments of this application, the application can extract visual features, audio features, and text features from the original video, and obtain an index feature vector of the original video through feature fusion and feature vectorization. Video parameter metadata of the original video is obtained. Based on the video parameter metadata and the index feature vector, the denoised video is stored and managed.

[0070] In one or more embodiments of this application, the visual features include lighting environment features, motion type features, and content type features. This application can extract these three types of features to obtain the visual features of the original video, including: determining the content type features of the original video based on keyframe groups through feature extraction and category recognition; determining the lighting environment features based on the lighting conditions of the original video; determining the motion type features based on the inter-frame displacement of each original frame in the original video; and concatenating the content type features, lighting environment features, and motion type features to obtain the visual features.

[0071] It should be noted that this application does not limit the specific method for determining content type features, and can be set according to actual needs. For example, a feature extractor and a type decision maker can be used to classify content types (people, cartoons, nature, and animals, etc.), as shown in the following formula: In equations (11), (12), (13), and (14), xi Representing the i-th feature in a keyframe group, ResNet is a common residual neural network in deep learning. It is calculated by performing ResNet computation on each frame. f i 2D Transformer is a transformer based on attention mechanism and preorder network. Softmax is the formula for probability normalization. MLP is a multilayer perceptron. The output C in the type decision maker P is the preset number of content categories. It is the fused index feature vector. , , These are visual, audio, and text feature vectors, respectively. This indicates a weighted fusion operation. Visual features. It includes content type features, lighting environment features, and motion type features; For audio features, The vector is the text after BERT transformation, i.e., the text feature.

[0072] By using feature vectorization, the collected basic information, such as clarity, frame rate, resolution, director, lead actors, and release date, is stored as video parameter metadata, index feature vectors, and noise-reduced video data, making it easier to retrieve and analyze.

[0073] Figure 4 This is a schematic diagram illustrating the principle of video type feature extraction provided in an embodiment of this application. Figure 4 As shown, this application needs to obtain basic information about the original video, including its content type, lighting environment, and motion type, as auxiliary information for storing and managing the noise-reduced video.

[0074] Figure 5 This is a flowchart illustrating the media asset video management process provided in an embodiment of this application. Figure 5 As shown, this application needs to achieve the storage management of media asset videos through feature information extraction 501, feature vectorization 502, vector storage 503, index construction 504, similarity association 505, and metadata management 506.

[0075] Based on the aforementioned video noise reduction method, this application also provides a specific embodiment of a video noise reduction apparatus.

[0076] like Figure 6 As shown, Figure 6 This is a schematic diagram of a video noise reduction device provided in an embodiment of this application. The device includes a motion intensity module 601, a keyframe module 602, a dynamic noise reduction module 603, a static noise reduction module 604, a semantic feature module 605, and a noise-reduced video module 606.

[0077] The motion intensity module 601 is used to determine the inter-frame motion intensity of the original video. The keyframe module 602 is used to construct a keyframe group of the original video based on the inter-frame motion intensity. The dynamic noise reduction module 603 is used to obtain dynamic channel features for motion noise reduction based on the key frame group through frequency domain transformation and channel attention weighting. The static noise reduction module 604 is used to determine, for each original frame in the original video, static global features that preserve the spatial details in that original frame; The semantic feature module 605 is used to extract semantic features for each of the original frames; The noise reduction video module 606 is used to obtain a noise reduction video by feature fusion based on the semantic features of each original frame, the static global features of each original frame, and the dynamic channel features.

[0078] In one feasible implementation, the motion intensity module is specifically used to determine the forward and backward optical flow fields of the original video using a lightweight optical flow network; for each pixel position of the original video, determine the inter-frame displacement of that pixel position based on the forward and backward optical flow fields; and determine the inter-frame motion intensity of the original video based on the inter-frame displacements of each pixel position.

[0079] In one feasible implementation, the aforementioned keyframe module is specifically used to set multiple preset frame extraction intervals and multiple mutually exclusive threshold ranges, wherein the frame extraction interval and the mutually exclusive threshold range have a one-to-one correlation, and the frame extraction interval and the mutually exclusive threshold range have a negative correlation; the mutually exclusive threshold range into which the inter-frame motion intensity falls is used as a matching threshold range; according to the correlation, the frame extraction interval corresponding to the matching threshold range is used as the target inter-frame interval; and the keyframe group of the original video is constructed according to the target frame extraction interval.

[0080] In one feasible implementation, the aforementioned dynamic noise reduction module is specifically used to obtain the video frequency domain tensor of the keyframe group through frequency domain transformation; perform noise separation on the video frequency domain tensor to obtain a sparse signal-to-noise mask; determine the channel attention weights of the sparse signal-to-noise mask; perform channel attention weighting on the sparse signal-to-noise mask according to the channel attention weights to obtain a target signal-to-noise mask; and obtain the dynamic channel features through time domain transformation based on the target signal-to-noise mask and the video frequency domain tensor.

[0081] In one feasible implementation, the aforementioned static noise reduction module is specifically used to obtain a single-frame frequency domain tensor for each original frame through frequency domain transformation; for each original frame, sort and divide the spectral coefficients of the single-frame frequency domain tensor according to the magnitude of the frequency domain amplitude to obtain multiple mutually exclusive subsets, and obtain a multi-branch convolution kernel set by performing time domain transformation on the multiple mutually exclusive subsets; for each original frame, generate frequency domain modulation features based on the original frame and the multi-branch convolution kernel set, and determine the global attention weight of the frequency domain modulation features; for each original frame, perform global attention weighting on the frequency domain modulation features according to the global attention weight to obtain target modulation features; for each original frame, after time domain transformation of the target modulation features, fuse them with the original frame through residual connection to obtain the static global features.

[0082] In one feasible implementation, the aforementioned noise-reduced video module is specifically used to: downsample each original frame to obtain semantic features; concatenate the static global features and dynamic channel features of each original frame along the channel dimension to generate fusion features; obtain fusion coefficients for each original frame based on the semantic features, the inter-frame motion intensity, and the fusion features; perform weighted upsampling on the semantic features and the fusion features for each original frame based on the fusion coefficients to obtain a noise-reduced frame; and obtain a noise-reduced video based on each noise-reduced frame.

[0083] In one feasible implementation, the above-mentioned noise reduction video module can also be used to extend the inter-frame motion intensity in the scalar dimension to the spatial dimension of the semantic features to obtain a motion intensity tensor; and according to the motion intensity tensor, the semantic features, and the fusion features, the fusion coefficients are obtained by concatenating along the channel dimension and then mapping through the channels.

[0084] In one feasible implementation, the above-mentioned device further includes a storage management module, specifically used to extract visual features, audio features, and text features of the original video, and obtain an index feature vector of the original video through feature fusion and feature vectorization; obtain video parameter metadata of the original video; and store and manage the noise-reduced video according to the video parameter metadata and the index feature vector.

[0085] In one feasible implementation, the storage management module described above can also be used to determine the content type features of the original video based on the keyframe group through feature extraction and category recognition; determine the light environment features based on the lighting conditions of the original video; determine the motion type features based on the inter-frame displacement of each original frame in the original video; and concatenate the content type features, the light environment features, and the motion type features to obtain the visual features.

[0086] Figure 7 A schematic diagram of the hardware structure of a video noise reduction device provided in an embodiment of this application is shown.

[0087] A video noise reduction device may include a processor 701 and a memory 702 storing computer program instructions.

[0088] Specifically, the processor 701 may include a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0089] Memory 702 may include mass storage for data or instructions. For example, and not limitingly, memory 702 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. In one embodiment, memory 702 may include removable or non-removable (or fixed) media, or memory 702 may be non-volatile solid-state memory. Memory 702 may be internal or external to the integrated gateway disaster recovery device.

[0090] In one instance, memory 702 may be read-only memory (ROM). In one instance, the ROM may be a mask-programmed ROM, a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), an electrically rewritable ROM (EAROM), or flash memory, or a combination of two or more of these.

[0091] Memory 702 may include read-only memory (ROM), random access memory (RAM), disk storage media device, optical storage media device, flash memory device, electrical, optical, or other physical / tangible memory storage device. Therefore, generally, memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described with reference to the video noise reduction method according to one aspect of this application.

[0092] The processor 701 reads and executes computer program instructions stored in the memory 702 to achieve... Figure 1 The video noise reduction method in the illustrated embodiment.

[0093] In one example, a device for determining the maximum transmission unit of a network may further include a communication interface 703 and a bus 704. Wherein, as Figure 7 As shown, the processor 701, memory 702, and communication interface 703 are connected through bus 704 and complete communication with each other.

[0094] The communication interface 703 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of this application.

[0095] Bus 704 includes hardware, software, or both, that couples components of an online data traffic metering device together. For example, and not limitingly, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Extended Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), a Hyper Transport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or combinations of two or more of these. Where appropriate, bus 504 may include one or more buses. Although specific buses are described and illustrated in embodiments of this application, this application contemplates any suitable bus or interconnect.

[0096] In addition, in conjunction with the video noise reduction method in the above embodiments, this application embodiment can provide a computer storage medium for implementation. The computer storage medium stores computer program instructions; when these computer program instructions are executed by a processor, they implement any of the video noise reduction methods in the above embodiments.

[0097] This application also provides a computer program product, including a computer program, which, when executed, implements any of the video noise reduction methods described in the above embodiments.

[0098] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0099] The functional blocks shown in the above-described block diagram can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. Programs or code segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, read-only memory (ROM), flash memory, erasable read-only memory (EROM), floppy disks, compact disc read-only memory (CD-ROM), optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. Code segments can be downloaded via computer networks such as the Internet, intranets, etc.

[0100] It should also be noted that the exemplary embodiments mentioned in this application describe methods or systems based on a series of steps or apparatus. However, this application is not limited to the order of the above steps; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0101] The aspects of this application have been described above with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It should be understood that each block in the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that these instructions, executable via the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field-programmable logic circuit. It is also understood that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by dedicated hardware performing the specified functions or actions, or can be implemented by a combination of dedicated hardware and computer instructions.

[0102] The above description is merely a specific implementation of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the protection scope of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the protection scope of this application.

Claims

1. A video noise reduction method, characterized in that, The method includes: Determine the inter-frame motion intensity of the original video; Based on the inter-frame motion intensity, construct the keyframe group of the original video; Based on the keyframe group, dynamic channel features for motion noise reduction are obtained through frequency domain transformation and channel attention weighting. For each original frame in the original video, determine the static global features that preserve the spatial details of that original frame; Extract the semantic features of each of the original frames; Based on the semantic features of each original frame, the static global features of each original frame, and the dynamic channel features, a noise-reduced video is obtained through feature fusion.

2. The method according to claim 1, characterized in that, Determining the inter-frame motion intensity of the original video specifically includes: The forward and backward optical flow fields of the original video are determined using a lightweight optical flow network. For each pixel position in the original video, the inter-frame displacement of that pixel position is determined based on the forward optical flow field and the backward optical flow field; The inter-frame motion intensity of the original video is determined based on the inter-frame displacement of each pixel position.

3. The method according to claim 1, characterized in that, Based on the inter-frame motion intensity, the keyframe group of the original video is constructed, specifically including: Multiple preset frame-skipping intervals and multiple mutual exclusion threshold ranges are set, wherein the frame-skipping interval and the mutual exclusion threshold range have a one-to-one correlation, and the frame-skipping interval and the mutual exclusion threshold range have a negative correlation. The mutual exclusion threshold range into which the inter-frame motion intensity falls is taken as the matching threshold range; Based on the aforementioned correlation, the frame extraction interval corresponding to the matching threshold range is taken as the target inter-frame interval; Based on the target frame extraction interval, construct the keyframe group of the original video.

4. The method according to claim 1, characterized in that, Based on the keyframe group, dynamic channel features for inter-frame noise reduction are obtained through frequency domain transformation and channel attention weighting, specifically including: The video frequency domain tensor of the keyframe group is obtained through frequency domain transformation; Noise separation is performed on the video frequency domain tensor to obtain a sparse signal-to-noise mask; Determine the channel attention weights of the sparse signal-to-noise mask; Based on the channel attention weights, the sparse signal-to-noise mask is subjected to channel attention weighting to obtain the target signal-to-noise mask; Based on the target signal-to-noise mask and the video frequency domain tensor, the dynamic channel features are obtained through time-domain transformation.

5. The method according to claim 1, characterized in that, For each original frame in the original video, determine the static global features that preserve the spatial details of that original frame, specifically including: By performing frequency domain transformation, the single-frame frequency domain tensor of each original frame is obtained; For each original frame, the spectral coefficients of the frequency domain tensor of the single frame are sorted and divided according to the magnitude of the frequency domain amplitude to obtain multiple mutually exclusive subsets. Then, by performing time-domain transformation on the multiple mutually exclusive subsets, a multi-branch convolution kernel set is obtained. For each of the original frames, frequency domain modulation features are generated based on the original frame and the multi-branch convolution kernel set, and the global attention weights of the frequency domain modulation features are determined. For each of the original frames, the frequency domain modulation features are globally attention-weighted according to the global attention weights to obtain the target modulation features; For each original frame, the target modulation feature is transformed in the time domain and then fused with the original frame through residual connection to obtain the static global feature.

6. The method according to claim 1, characterized in that, Based on the semantic features, static global features, and dynamic channel features of each original frame, a noise-reduced video is obtained through feature fusion, specifically including: For each of the original frames, the original frame is downsampled to obtain semantic features; For each original frame, the static global features and dynamic channel features of the original frame are concatenated along the channel dimension to generate a fused feature; For each of the original frames, a fusion coefficient is obtained based on the semantic features, the inter-frame motion intensity, and the fusion features; For each original frame, the semantic features and the fusion features are weighted and upsampled according to the fusion coefficient to obtain a denoised frame; Based on each of the denoised frames, a denoised video is obtained.

7. The method according to claim 6, characterized in that, For each of the original frames, a fusion coefficient is obtained based on the original frame, the inter-frame motion intensity, and the fusion features, specifically including: The inter-frame motion intensity in the scalar dimension is extended to the spatial dimension of the semantic features to obtain the motion intensity tensor; The fusion coefficients are obtained by concatenating the motion intensity tensor, the semantic features, and the fusion features along the channel dimension and then mapping them through the channels.

8. The method according to claim 1, characterized in that, The method further includes: Visual features, audio features, and text features of the original video are extracted, and the index feature vector of the original video is obtained through feature fusion and feature vectorization. Obtain the video parameter metadata of the original video; The noise-reduced video is stored and managed based on the video parameter metadata and the index feature vector.

9. The method according to claim 8, characterized in that, Extracting the visual features of the original video specifically includes: Based on the keyframe group, the content type features of the original video are determined through feature extraction and category recognition; Based on the lighting conditions of the original video, determine the characteristics of the lighting environment; Based on the inter-frame displacement of each original frame in the original video, determine the motion type features; The visual features are obtained by concatenating the content type features, the lighting environment features, and the motion type features.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions that, when executed by a processor, implement the video noise reduction method as described in any one of claims 1-9.

Images