Data noise reduction method, device, computer equipment and storage medium

Abstract

The present application discloses a data denoising method, apparatus, computer equipment and storage medium. The method is implemented by: performing baseline filtering on motion state sample data to obtain motion state sample data with baseline drift removed, which includes static data and motion data; inputting the static data and motion data into a preset domain adaptation model, performing domain alignment processing and small motion estimation processing respectively to obtain domain alignment results and small motion estimation results, the domain alignment results including static global features and motion global features; determining alignment loss based on static global features and motion global features; determining estimation loss based on small motion estimation results; iteratively updating the preset domain adaptation model based on alignment loss and estimation loss. The present application adopts a multi-stage decoupling strategy to achieve data denoising, and significantly improves the denoising effect and model performance by processing different types of noise in a staged and targeted manner.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a data noise reduction method, apparatus, computer equipment, and storage medium. Background Art

[0002] When measuring an object's motion, inertial sensors (IMUs) often generate noise due to external interference and internal errors. Denoising is necessary to improve measurement accuracy. Traditional denoising methods, such as complementary filters, Kalman filters, and blind source separation, are highly dependent on task characteristics and specific parameter selection, resulting in poor generalization and difficulty adapting to complex and changing application scenarios.

[0003] Moreover, the above-mentioned traditional denoising methods mostly focus on processing the inherent mechanical noise of IMU, and are ineffective in processing other complex noise sources. In IMU-based micro-motion perception tasks, such as using an IMU placed on the human body to capture micro-motion, many challenges are faced. On the one hand, the motion signal itself can become a noise source; on the other hand, Figure 1 As shown in the figure, the spectra of macroscopic body movements and micro facial movements in the IMU signal overlap significantly, and the intensity of the human body motion interference signal far exceeds the micro motion characteristic signal, which makes the micro motion characteristic signal easily masked, greatly hindering the accurate perception and analysis of micro movements.

[0004] Although, with the development of deep learning, denoising autoencoders, generative adversarial networks, and domain adaptation methods have been applied to tasks such as IMU denoising, existing time series domain adaptation methods are mostly only applicable to classification tasks. In regression tasks, their aligned sample features lack information about temporal dynamic changes, making them difficult to effectively handle regression tasks such as micro-motion estimation, which have strict requirements on the dynamic characteristics of time series. They are also unable to meet the needs of practical applications such as IMU-based micro-motion perception. Summary of the Invention

[0005] Based on this, it is necessary to provide a data denoising method, apparatus, computer equipment and storage medium to address the above technical issues, so as to solve at least one problem existing in the above-mentioned prior art.

[0006] In a first aspect, a data noise reduction method is provided, comprising:

[0007] Performing baseline filtering on the motion state sample data to obtain motion state sample data with baseline drift removed, wherein the motion state sample data with baseline drift removed includes static data and motion data;

[0008] Inputting the static data and the motion data into a preset domain adaptation model, performing domain alignment processing and small motion estimation processing respectively, to obtain a domain alignment result and a small motion estimation result, wherein the domain alignment result includes a static global feature and a motion global feature;

[0009] determining an alignment loss based on the static global feature and the motion global feature;

[0010] determining an estimated loss based on the micro-motion estimation result;

[0011] The preset domain adaptation model is iteratively updated based on the alignment loss and the estimation loss.

[0012] In a possible implementation, inputting the static data and the motion data into a preset domain adaptation model to perform domain alignment processing and slight motion estimation processing respectively includes:

[0013] Extracting features from the static data and the motion data to obtain static features and motion features;

[0014] Normalizing the static features and the motion features in a time dimension to obtain standard static features and standard motion features;

[0015] Based on the category token, domain aligning the standard static features and the standard motion features to obtain the static global features and the motion global features;

[0016] Based on the standard static features and the standard motion features, a slight motion estimation result is obtained.

[0017] In a possible implementation, extracting features from the static data and the motion data to obtain static features and motion features respectively includes:

[0018] dividing the static data and the motion data into a plurality of static time series segments and a plurality of motion time series segments respectively;

[0019] extracting at least one target time series segment from the plurality of stationary time series segments and the plurality of moving time series segments according to a preset timestamp label;

[0020] According to preset extraction rules, feature extraction is performed on the target time series segments respectively by a preset feature extractor to obtain the static features and the motion features.

[0021] In a possible implementation, the normalizing the static features and the motion features in the time dimension to obtain standard static features and standard motion features includes:

[0022] Determining the mean and standard deviation of the static features and the motion features;

[0023] Based on the mean value and the standard deviation, the static feature and the motion feature are standardized to obtain a standard static feature and a standard motion feature.

[0024] In a possible implementation, the domain alignment of the standard static features and the standard motion features based on the category token to obtain the static global features and the motion global features includes:

[0025] Adding category tokens to the standard static feature and the standard motion feature respectively to obtain a static category token feature and a motion category token feature, wherein the category tokens are used to identify domain features that change dynamically over time;

[0026] The stationary category token features and the motion category token features are input into a preset encoder for processing to obtain a stationary global feature and a motion global feature.

[0027] In a possible implementation, obtaining a slight motion estimation result based on the standard static feature and the standard motion feature includes:

[0028] Expanding the prefix labels in the feature time series corresponding to the standard static features and the standard motion features to obtain an expanded label sequence, wherein the prefix labels refer to labels corresponding to prefix motion state data, and the prefix motion state data are motion state sample data corresponding to the first preset features in the feature time series;

[0029] Inputting the expanded label sequence into the encoder layer for self-attention calculation;

[0030] The label sequence calculated after self-attention is input into the decoder layer for feature fusion and decoding processing to obtain the micro-motion estimation result.

[0031] In a possible implementation manner, after iteratively updating the preset domain adaptation model, the method further includes:

[0032] Acquiring motion state data to be estimated;

[0033] Performing baseline filtering on the motion state data to be estimated to obtain the motion state data to be estimated with baseline drift removed;

[0034] performing feature extraction on the motion state data to be estimated with baseline drift removed to obtain features to be estimated;

[0035] Standardizing the features to be estimated in a preset time dimension to obtain standard features to be estimated;

[0036] Based on the standard feature to be estimated, a micro-motion estimation result corresponding to the motion state data to be estimated is obtained.

[0037] In a second aspect, a data noise reduction device is provided, comprising:

[0038] a first noise reduction unit, configured to perform baseline filtering on the motion state sample data to obtain motion state sample data with baseline drift removed, wherein the motion state sample data with baseline drift removed includes static data and motion data;

[0039] a second denoising unit, configured to input the static data and the motion data into a preset domain adaptation model, and perform domain alignment processing and small motion estimation processing respectively, to obtain a domain alignment result and a small motion estimation result, wherein the domain alignment result includes a static global feature and a motion global feature;

[0040] an alignment loss determining unit, configured to determine an alignment loss based on the static global feature and the motion global feature;

[0041] an estimated loss determining unit, configured to determine an estimated loss based on the micro-motion estimation result;

[0042] A model iterative updating unit is used to iteratively update the preset domain adaptation model based on the alignment loss and the estimation loss.

[0043] In a third aspect, a computer device is provided, comprising a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor, wherein the processor implements the steps of the data denoising method as described above when executing the computer-readable instructions.

[0044] In a fourth aspect, a readable storage medium is provided, wherein the readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by a processor, the steps of the data denoising method as described above are implemented.

[0045] The above-mentioned data denoising method, device, computer equipment and storage medium, and its method implementation include: performing baseline filtering on the motion state sample data to obtain motion state sample data with baseline drift removed, the motion state sample data with baseline drift removed includes static data and motion data; inputting the static data and motion data into a preset domain adaptation model, performing domain alignment processing and small motion estimation processing respectively to obtain domain alignment results and small motion estimation results, the domain alignment results include static global features and motion global features; determining the alignment loss based on the static global features and the motion global features; determining the estimation loss based on the small motion estimation results; iteratively updating the preset domain adaptation model based on the alignment loss and the estimation loss. In an embodiment of the present application, an innovative multi-stage decoupling strategy is adopted to achieve data denoising, and the noise reduction effect and model performance are significantly improved by processing different types of noise in a staged and targeted manner. First, in the first stage, a baseline filtering mechanism is designed to deal with the low-frequency interference noise caused by external interference. This mechanism can dynamically adjust the baseline according to signal changes, effectively suppress the drift characteristics of low-frequency noise, accurately filter out such noise interference, and lay a pure data foundation for subsequent processing. Secondly, the second stage focuses on the interference noise introduced by human movement. This stage introduces time domain information to deeply explore the changing characteristics of the data in the time dimension. By capturing the dynamic change patterns of the data over a long time span, it can not only comprehensively handle complex and changeable interference noise, but also greatly enhance the model's adaptability to time series data. In complex motion scenarios, whether it is fast and intense body movements or subtle and continuous posture changes, this method can stably and efficiently handle noise, showing excellent robustness, and providing reliable guarantees for tasks such as small motion perception based on inertial sensors (IMUs). BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments of the present application. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0047] Figure 1 is a curve diagram illustrating a frequency domain aliasing phenomenon between a human motion signal and a facial signal in an embodiment of the prior art;

[0048] Figure 2 This is a schematic diagram of an overall framework of a data noise reduction method in an embodiment of the present application;

[0049] Figure 3This is a schematic diagram of a network architecture of a domain adaptation model TADA that combines time domain information in one embodiment of the present application;

[0050] Figure 4 This is a flow chart of a data noise reduction method in one embodiment of the present application;

[0051] Figure 5 This is a schematic diagram of a network architecture of a DACT encoder in one embodiment of the present application;

[0052] Figure 6 This is a schematic diagram of a network architecture of a micro-motion estimation module in one embodiment of the present application;

[0053] Figure 7 This is a structural diagram of a data noise reduction device in one embodiment of the present application;

[0054] Figure 8 Schematic diagram of a computer device in one embodiment of the present application. DETAILED DESCRIPTION

[0055] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0056] The data denoising method provided in this embodiment can be applied to Figure 2 In this application environment, we first use an adaptive baseline correction algorithm to process interference signals with drift characteristics. This algorithm dynamically adjusts the signal baseline using baseline filtering, effectively removing low-frequency interference signals. This allows for flexible baseline adjustment based on actual conditions, significantly improving accuracy. In the second stage, we constructed a domain adaptation model (TADA) that incorporates time domain information. This model can handle interference signals over longer time spans and demonstrates greater robustness in complex motion scenarios. Finally, we perform micro-motion estimation on the denoised signal.

[0057] like Figure 3As shown, the temporal domain adaptation model (TADA) for temporal information may include a feature extractor, a TPN module (temporal segment normalization module), a DACT encoder, and a micro-motion estimation module. The feature extractor extracts features from both static and moving data, obtaining basic feature representations of motion state samples. The TPN module receives the features output by the feature extractor and further processes them to capture the data's positional information and temporal variation patterns, providing time-aware features for subsequent processing. The DACT encoder utilizes a self-attention mechanism to deeply encode the features processed by the TPN module, mining correlations between different features within the data. It also uses category tokens to achieve domain alignment between the static and moving domains, enhancing feature representation capabilities. The micro-motion estimation module estimates the features processed by the TPN module and outputs task-relevant micro-motion estimation results, such as micro-motion estimation values.

[0058] The DACT encoder can include multiple stacked Encoder Blocks, such as three Feed-Forward neural networks and Attention Blocks. The Encoder Block is the fundamental building block of the entire encoder structure, and stacking multiple such modules enables deep encoding of input data. Within the Feed-Forward Encoder Block, the feed-forward neural network performs nonlinear transformations on the input data, enhancing the data's feature representation capabilities through the form of a multi-layer perceptron, allowing the model to learn more complex feature patterns. The Attention Block is a key component of the Encoder Block. It determines the importance of each position in feature calculation by calculating the attention weights between different positions in the input sequence, effectively capturing long-distance dependencies in the sequence and enabling the model to focus on information more important to the current task.

[0059] The micro-motion estimation module includes an input embedding layer, a Transformer Encoder, a Transformer Decoder, and Linear Projection. The input embedding layer converts input features into a low-dimensional dense vector representation suitable for model processing, giving the data its initial feature representation. The Transformer Encoder processes the features processed by the input embedding layer. It incorporates a self-attention mechanism to capture correlations between data features in different subspaces. The feedforward neural network further performs nonlinear transformations on the features. Residual connections and layer normalization (Add & Norm) stabilize network training by normalizing the inputs of each layer to prevent vanishing or exploding gradients, thereby extracting deep-level features from the data. The Transformer Decoder receives the output of the Transformer Encoder and processes it in conjunction with observed IMU features. Similarly, using the cross-attention mechanism and feedforward neural network, it establishes connections between the encoder output and the observed features, decoding the relevant information. Linear Projection linearly transforms the feature vector output by the Transformer Decoder to the required output dimension, resulting in the final micro-motion estimation result or other task-related output.

[0060] In one embodiment, if Figure 4 As shown, a data noise reduction method is provided, comprising the following steps:

[0061] In step S110, baseline filtering is performed on the motion state sample data to obtain motion state sample data with baseline drift removed, wherein the motion state sample data with baseline drift removed includes static data and motion data;

[0062] It's important to note that motion state sample data refers to IMU data, which is time-series data collected by an inertial measurement unit (IMU). This data describes an object's motion state in three-dimensional space, including angular velocity and acceleration. An IMU is an inertial sensor capable of measuring an object's three-axis acceleration and angular velocity. An IMU placed on the human body not only captures minute movements, but these motion signals themselves constitute a source of noise that needs to be eliminated.

[0063] It is understandable that the motion state sample data includes static data and motion data, which are presented in the form of time series. Static data refers to the motion state data collected by the inertial measurement unit when the measured target, such as an object or a human body, is in a relatively static state. Motion data refers to data related to motion. It is usually generated by a moving object or human body, such as the acceleration, velocity, displacement and other data recorded by the sensor during the object's movement, or various dynamic information collected by a mobile device during movement. It should be noted that since the inertial measurement unit usually only collects sensitive data on small movements, such as hand tremors (which can monitor diseases) and facial muscle twitching, in a motion scene, the signal of a person walking will also be collected by the IMU. This data will interfere with the above-mentioned small motion data, so the collected data needs to be denoised.

[0064] The motion state sample data is the IMU signal originally collected by the IMU sensor, which often contains noise interference (such as sensor errors and noise generated by external environmental interference). For example, low-frequency interference noise with drift characteristics and interference noise introduced for human motion cannot directly meet the needs of high-precision applications. Since the movement frequency of this type of noise with drift characteristics is relatively low, it is much smaller than the frequency of small movements, such as the falling off of the device where the IMU is located. Therefore, in response to this type of noise, this application constructs an adaptive baseline correction algorithm that can extract the low-frequency baseline component in the IMU signal.

[0065] Specifically, a preset frequency threshold, such as 0.1Hz, can be set. Signal components below 0.1Hz are considered baseline signals. These baseline signals are primarily generated by low-frequency interference. Therefore, by extracting and removing these low-frequency baseline signals, the original collected IMU data can be effectively corrected.

[0066] The specific processing process of the adaptive baseline correction algorithm is as follows:

[0067] Perform high-pass filtering on the motion state sample data to remove low-frequency components below the preset frequency threshold. Assuming that the motion state sample data is x(t), the signal after high-pass filtering is . Then its expression is as follows:

[0068] ;

[0069] in, represents the baseline signal extracted by the low-pass filter. The preset frequency threshold of the low-pass filter can be set to 0.1Hz to ensure that only the low-frequency components of the motion are extracted. The high-pass filter can be implemented by the following formula:

[0070] ;

[0071] Where H(s) is the filter transfer function, s is the complex frequency variable, c is the cutoff frequency (e.g., a preset frequency threshold of 0.1 Hz). This allows for adaptive correction of the IMU signal to remove low-frequency interference, thereby improving the accuracy of subtle motion perception. This adaptive baseline correction algorithm is not only simple and efficient, but also significantly improves the robustness of subtle motion perception.

[0072] In step S120, the static data and the motion data are input into a preset domain adaptation model, and domain alignment processing and small motion estimation processing are performed respectively to obtain a domain alignment result and a small motion estimation result, wherein the domain alignment result includes a static global feature and a motion global feature;

[0073] Specifically, if Figure 2 As shown, the signal that has been denoised by the adaptive baseline correction algorithm in the first stage can enter the second stage and undergo denoising again through the domain adaptation model TADA that combines time domain information. TADA may include Feature Extractor (feature extractor), TPN Module (time segment normalization module), DACT encoder and micro-motion estimation module. It should be noted that the preset domain adaptation model may include two branches, one branch may be provided with a DACT encoder for aligning the static and motion domains, and the other branch may be provided with a micro-motion estimation module for performing micro-motion estimation.

[0074] It should be noted that the specific processing process of TADA can be as follows: the static data and motion data are input into a shared feature extractor, which processes the static data and motion data separately to extract preliminary static and motion feature representations. These static and motion feature representations are input into the TPN module. The TPN module then normalizes the data information along the temporal dimension to ensure the stability of subsequent network inputs and help subsequent models better account for dependencies in time series data. Class tokens are introduced into the standard static and motion features obtained after processing by the TPN module and input into the DACT encoder to obtain the corresponding static global features [CLS]s and motion global features [CLS]m, achieving domain alignment between the static and motion domains and improving the model's generalization ability.

[0075] At the same time, the standard static features and the standard motion features can be input into the small motion estimation module, and the small motion estimation module can perform supervised learning based on these features and the true labels.

[0076] In step S130, an alignment loss is determined based on the static global feature and the motion global feature;

[0077] Specifically, the DACT encoder can obtain static global features [CLS]s and motion global features [CLS]m based on the standard static features and standard motion features after normalization by the TPN module, and then calculate Lalign (alignment loss) based on the static global features [CLS]s and motion global features [CLS]m. This alignment loss is used to measure and narrow the difference between the static domain and the motion domain, prompting the model to learn more general and aligned features.

[0078] The alignment loss can be calculated using the following formula:

[0079] ;

[0080] In step S140, based on the micro-motion estimation result, an estimated loss is determined;

[0081] Specifically, the standard static features and standard motion features normalized by the TPN module are input into the subtle motion estimation module. The subtle motion estimation module first expands and embeds these standard static and motion features. Expansion allows the features to contain richer information, while embedding converts the features into vector representations suitable for model processing. The processed features are then input into the encoder layer, which employs a self-attention mechanism to perform attention calculations to capture correlations between features. Finally, the decoder layer decodes the features using a specific decoding algorithm, outputting the subtle motion estimation results.

[0082] The micro-motion estimation results may include displacement information, speed information, movement direction, finger joint angles, or facial muscle movement intensity.

[0083] Then, the estimated loss can be calculated based on the preset loss function and the true label.

[0084] The estimated loss can be calculated using the following formula:

[0085] ;

[0086] in, represents the estimation, L is the total length of the time series, h represents the number of prefix sequences, represents the sequence of decoder layer outputs, expressed as , since the prefix h is a known condition, the estimated loss is calculated from m onwards.

[0087] In step S150 , the preset domain adaptation model is iteratively updated based on the alignment loss and the estimation loss.

[0088] Specifically, the estimation loss is used to measure the difference between small motion indicators (such as vibration amplitude, motion intensity, etc.) and the actual value, while the alignment loss is used to ensure that the feature distributions of the static domain and the motion domain are as similar as possible, thereby improving the generalization ability of the model in different fields. In order to comprehensively consider the optimization of these two tasks, the alignment loss can be used to and estimated losses , obtaining the overall loss. Based on this overall loss, the preset domain adaptation model is iterated, adjusting the model weight coefficients until the preset convergence conditions are met. For example, when the overall loss is less than the preset loss threshold or the number of iterations reaches the preset number, training is completed, and the trained domain adaptation model is obtained. Optimizing the model using alignment loss and estimation loss balances the importance of estimation and domain feature alignment to optimize model performance.

[0089] The overall loss can be calculated using the following formula:

[0090]

[0091] Among them, λ is a hyperparameter used to control the alignment loss the relative importance of .

[0092] In an embodiment of the present application, a data denoising method is provided, comprising: performing baseline filtering on motion state sample data to obtain motion state sample data with baseline drift removed, wherein the motion state sample data with baseline drift removed includes stationary data and motion data; inputting the stationary data and motion data into a preset domain adaptation model, performing domain alignment processing and small motion estimation processing respectively to obtain domain alignment results and small motion estimation results, wherein the domain alignment results include stationary global features and motion global features; determining alignment loss based on the stationary global features and the motion global features; determining estimation loss based on the small motion estimation results; and iteratively updating the preset domain adaptation model based on the alignment loss and the estimation loss. In an embodiment of the present application, a multi-stage decoupling strategy is innovatively adopted to achieve data denoising, and by processing different types of noise in a staged and targeted manner, the denoising effect and model performance are significantly improved. First, in the first stage, a baseline filtering mechanism is designed to deal with low-frequency interference noise caused by external interference. This mechanism can dynamically adjust the baseline according to signal changes, effectively suppress the drift characteristics of low-frequency noise, accurately filter out such noise interference, and lay a pure data foundation for subsequent processing. Secondly, the second phase focuses on the interference noise introduced by human motion. This phase introduces time-domain information to deeply explore the changing characteristics of the data over time. By capturing the dynamic changes in the data over a long period of time, this method not only comprehensively handles complex and variable interference noise but also greatly enhances the model's adaptability to time series data. In complex motion scenarios, whether rapid and intense body movements or subtle and sustained posture changes, this method can stably and efficiently handle noise, demonstrating excellent robustness, providing reliable support for tasks such as subtle motion sensing based on inertial sensors (IMUs).

[0093] In one embodiment of the present application, the static data and the motion data are input into a preset domain adaptation model to perform domain alignment processing and small motion estimation processing respectively, including:

[0094] Extracting features from the static data and the motion data to obtain static features and motion features;

[0095] Normalizing the static features and the motion features in a time dimension to obtain standard static features and standard motion features;

[0096] Based on the category token, domain aligning the standard static features and the standard motion features to obtain the static global features and the motion global features;

[0097] Based on the standard static features and the standard motion features, a slight motion estimation result is obtained.

[0098] Specifically, the first stage uses an adaptive baseline correction algorithm to remove low-frequency interference noise with drifting characteristics. Then, in the second stage, a domain adaptation model (TADA) that incorporates temporal information removes interference noise introduced by the motion of objects or people being measured. This phased and targeted approach to addressing different types of noise significantly improves noise reduction and model performance.

[0099] It should be noted that the domain adaptation model TADA for time domain information in the training phase may include a Feature Extractor (feature extractor), a TPN module (time segment normalization module), a DACT encoder, and a micro-motion estimation module.

[0100] Among them, the specific processing process of TADA is: the static data and motion data are input into the shared feature extractor respectively, the feature extractor processes the original data, and extracts the preliminary static feature representation and motion feature representation. The static feature representation and motion feature representation are input into the TPN module. The TPN module standardizes the data information from the time dimension to ensure the stability of subsequent network inputs, which helps subsequent models to better focus on the dependencies in time series data. Features with time-aware characteristics enter the DACT encoder. The DACT encoder uses a self-attention mechanism to establish associations between different feature dimensions and different time steps, deeply encodes the features, and obtains more representative and discriminative feature representations. At the same time, the static data and motion data are processed by their respective DACT encoders to obtain corresponding and , which is the global feature representation. Based on the obtained and , calculate the alignment loss This alignment loss is used to measure and narrow the gap between the static and motion domains, encouraging the model to learn more general, aligned features and improving model generalization. Normalizing the time series along the temporal dimension before domain alignment helps improve the smoothness of the domain alignment input and enables subsequent models to better capture dependencies in time series data.

[0101] At the same time, the features processed by the TPN module can also enter the micro-motion estimation module, which performs micro-motion estimation based on these features and outputs the micro-motion estimation results. Then, the estimated loss can be calculated based on the micro-motion estimation results and the true label. , the estimated loss can measure the gap between the estimated result and the true value, and is used to guide the parameter update of the model in the estimation task.

[0102] Finally, the calculated and , the parameters in the model are adjusted and optimized through the back-propagation algorithm, and the training is continuously iterated, so that the performance of the TADA model in domain feature alignment and estimation tasks is gradually improved.

[0103] In one embodiment of the present application, extracting features from the static data and the motion data to obtain static features and motion features respectively includes:

[0104] dividing the static data and the motion data into a plurality of static time series segments and a plurality of motion time series segments respectively;

[0105] extracting at least one target time series segment from the plurality of stationary time series segments and the plurality of moving time series segments according to a preset timestamp label;

[0106] According to preset extraction rules, feature extraction is performed on the target time series segments respectively by a preset feature extractor to obtain the static features and the motion features.

[0107] Specifically, static data and motion data can be input into a shared feature extractor. After convolution, pooling, and activation, static and motion features are obtained, respectively. Specifically, the motion state sample data is continuously collected by the sensor, so the static and motion data are time series signals. The time series signal is segmented into multiple time series segments. The corresponding sequence segments are then selected and input into the feature extractor for feature extraction. This feature extractor consists of a multi-layer convolutional neural network (CNN), for example, three layers. The time series segments enter the convolution layer, which may include multiple convolution kernels. The convolution kernels move point by point along the time dimension and perform convolution operations with the corresponding data. A single convolution operation produces a feature map, while multiple convolution kernels produce multiple feature maps. Each feature map represents a local feature representation of the data. To reduce data volume and complexity, the data processed by the convolution layer can be pooled using a pooling layer to reduce data dimensionality. Finally, activation functions (such as ReLU and Sigmoid) are used to extract static and motion features.

[0108] For example, for a time series signal of length T, x = [ , , … , ], extract the sequence segment centered on the tag timestamp with a window size of ω. Given the tag timestamp T={ , , … , }, then the corresponding sequence segment is ,in Then the Input into the feature extractor for feature extraction.

[0109] It should be noted that in order to reduce the number of parameters and improve the generalization ability of the model, the same feature extractor can be applied to different sequence segments, that is, shared parameters. The shared parameters are expressed as:

[0110] ;

[0111] in, is the sequence segment obtained by the feature extractor The feature representation of , θ represents the shared parameters of the feature extractor.

[0112] In one embodiment of the present application, the static features and the motion features are respectively normalized in the time dimension to obtain standard static features and standard motion features, including:

[0113] Determining the mean and standard deviation of the static features and the motion features;

[0114] Based on the mean value and the standard deviation, the static feature and the motion feature are standardized to obtain a standard static feature and a standard motion feature.

[0115] Specifically, although the shared parameter mechanism of the feature extractor can ensure that the meaning of each sequence segment channel remains consistent, the feature value itself is not constrained. Therefore, in order to ensure the stability of subsequent network input, the static features and motion features extracted by the feature extractor can be normalized through the TPN module (time segment normalization module).

[0116] The features extracted by the feature extractor are expressed as , where B is the batch size, C is the number of channels, and L is the length of the time series. The TPN module can be standardized along the L dimension. To perform TPN on H, the mean μ and standard deviation σ of the features along the time dimension can be calculated first. The mean μ and standard deviation σ can be calculated using the following formulas:

[0117] ;

[0118] ;

[0119] in, is a small constant used to prevent division by zero.

[0120] Then, the feature H is normalized and a learnable parameter is introduced and By zooming and panning, we can get the following calculation formula:

[0121] ;

[0122] Among them, the standardized feature is expressed as , which will be used as input to the subsequent network.

[0123] In one embodiment of the present application, the domain alignment of the standard static features and the standard motion features based on the category token to obtain the static global features and the motion global features includes:

[0124] Adding category tokens to the standard static feature and the standard motion feature respectively to obtain a static category token feature and a motion category token feature, wherein the category tokens are used to identify domain features that change dynamically over time;

[0125] The stationary category token features and the motion category token features are input into a preset encoder for processing to obtain a stationary global feature and a motion global feature.

[0126] It's important to note that temporal dynamics can effectively reflect the temporal dependencies of data, whereas samples in traditional domain adaptation lack this temporal dynamic pattern. Therefore, we introduce category tokens (CLS tokens) into standard static and motion features. These tokens can identify domain features that dynamically change over time, thereby aligning features along this dimension. By capturing the temporal variations of data, we improve the model's robustness in complex motion scenarios.

[0127] Then, the input time series data (normalized features) with [CLS] tags are fed into the DACT encoder, as Figure 5As shown, the DACT encoder may include multiple stacked Encoder Blocks, for example, three, a Feed-Forward Neural Network (FNN), and an Attention Block. The Attention Block within each Encoder Block performs attention calculations on the input sequence. Specifically, attention scores are calculated between different positions in the sequence to generate an attention weight matrix. This weight matrix represents the importance of each position in the sequence to other positions. The input sequence is then weighted and summed based on this weight matrix, highlighting information from important positions and suppressing unimportant information, thereby capturing long-range dependencies and key features in the sequence. The feature representations processed by the attention blocks are input to the Feed-Forward Neural Network (FNN). The FNN performs nonlinear transformations on these features. Through computations across multiple layers of neurons, it further extracts and transforms features, increasing their expressive power and learning more complex feature patterns. The output of one Encoder Block serves as the input to the next Encoder Block, repeating the aforementioned attention calculation and FNN processing steps. By stacking multiple Encoder Blocks, we can perform in-depth feature extraction and encoding on the input data, continuously enhancing the richness and accuracy of the feature representation. After processing by all Encoder Blocks, we ultimately obtain a feature representation that contains the global information of the entire input sequence. The feature representation at the [CLS] position aggregates the global semantics of the sequence and can be used in subsequent tasks, such as classification tasks, as a basis for classification, or in other tasks that require global features.

[0128] Specifically, you can add CLS Token in front of the extracted sequence features to get the sequence representation By adding CLS Token to the standard static features and standard motion features respectively, the marked static features can be obtained. And the marked motion characteristics After the introduction of the category token CLS Token, in order to further process the data marked with [CLS], it is input into the DACT encoder. The DACT encoder performs attention calculation on the input features through the self-attention mechanism to capture the long-distance dependencies and key information of the sequence, and then performs nonlinear transformation to enhance the feature expression ability. Finally, normalization and residual connection can be performed. After processing by multiple layers of encoders, the added classification tag [CLS] integrates the entire static feature sequence or motion feature sequence information. At this time, the feature vector corresponding to [CLS] represents the global feature of the sequence, which is represented as and ,in, and , D represents the feature dimension. For example, the stationary feature sequence The vector contains the global information of the static state and the motion feature sequence The vector contains global information about the motion state and can be used for subsequent tasks such as classification and estimation.

[0129] In one embodiment of the present application, a slight motion estimation result is obtained based on the standard static features and the standard motion features, including:

[0130] Expanding the prefix labels in the feature time series corresponding to the standard static features and the standard motion features to obtain an expanded label sequence, wherein the prefix labels refer to labels corresponding to prefix motion state data, and the prefix motion state data are motion state sample data corresponding to the first preset features in the feature time series;

[0131] Inputting the expanded label sequence into the encoder layer for self-attention calculation;

[0132] The label sequence calculated after self-attention is input into the decoder layer for feature fusion and decoding processing to obtain the micro-motion estimation result;

[0133] The micro-motion estimation results may include displacement information, speed information, movement direction, finger joint angles, or facial muscle movement intensity.

[0134] Specifically, standard static features and standard motion features are input into the micro-motion estimation module for estimation processing. This module includes Input Embedding, a TransformerEncoder, and a Transformer Decoder. The input embedding transforms the input features into a low-dimensional dense vector representation suitable for model processing using a specific mapping function, giving the data an initial feature representation. The encoder processes the embedded features using a self-attention mechanism, a feedforward neural network, residual connections, and layer normalization to extract deep features. The decoder similarly processes the encoder's output deep features using a cross-attention mechanism and a feedforward neural network to decode task-specific information. The micro-motion estimation module regresses the IMU micro-motion features into task-specific output metrics (such as finger joint angles and facial muscle movement intensity). Through the micro-motion estimation module, the model learns the mapping changes between labels and the IMU from prefix information to better estimate micro-motion information.

[0135] like Figure 6As shown, it is assumed that the feature representation of the TPN module output is , where the first m features come from the prefix motion state data, L is the total length of the time series, and Lm is the length of the observed features. To avoid the serious exposure bias caused by the autoregressive training strategy, the use of the lower triangle mask for regression training is eliminated. It should be noted that the motion state sample data is continuously collected by the sensor. Therefore, the features of the IMU data observed at the current time can be used as the observed features, while the IMU data collected before the current time is used as the prefix motion state data, and the corresponding features are the prefix IMU features.

[0136] The prefix label can be extended to the total length by zero padding. The extended label is represented as . Make the label length consistent with the total length of the feature time series.

[0137] These expanded labels can then be used as input to the encoder layer, so that the labels and features match in dimension and length. The encoder layer includes a self-attention mechanism that performs self-attention calculations on the input features, allowing the model to learn the association between the prefix IMU features and the prefix labels. The cross-attention mechanism allows the decoder to focus on the features output by the encoder. In this case, it allows the model to learn how the prefix IMU features correspond to the labels and explore the mapping relationship between them. Finally, the decoder layer can output a time series of estimated indicators ,in The part represents the estimation result. This means that after learning and processing, the model outputs a time series of estimated values. This estimated value corresponds to the output indicator of a specific task (such as the micro-motion estimation value) and can be used for subsequent analysis and application.

[0138] It should be noted that the IMU length can be set to 200 samples. The output metric's label time series length L can be set to a corresponding number of samples based on actual task requirements, for example, 60 samples. The prefix time series length m can be set to 15 samples. It is important to note that although L is 60 (corresponding to 2 seconds), this method can still maintain a sustained high frame rate output through the overlapping sliding window method. For model training, the Adam optimizer was used with a learning rate of 1e-4 and a batch size of 32. The loss weight parameter λ was set to 10. In addition, the OneCycleLR scheduling strategy was applied to adjust the learning rate. In the micro-motion estimation module, the number of Transformer layers can be set to 6, the number of heads to 8, and the hidden dimension size to 256.

[0139] In an embodiment of the present application, after iteratively updating the preset domain adaptation model, the method further includes:

[0140] Acquiring motion state data to be estimated;

[0141] Performing baseline filtering on the motion state data to be estimated to obtain the motion state data to be estimated with baseline drift removed;

[0142] performing feature extraction on the motion state data to be estimated with baseline drift removed to obtain features to be estimated;

[0143] Standardizing the features to be estimated in a preset time dimension to obtain standard features to be estimated;

[0144] Based on the standard feature to be estimated, a micro-motion estimation result corresponding to the motion state data to be estimated is obtained.

[0145] It should be noted that when performing inference through the trained domain adaptation model TADA, domain alignment is not required. Therefore, the TADA model during inference may include a Feature Extractor, a TPN Module (time segment normalization module), and a micro-motion estimation module.

[0146] After baseline filtering of the motion state data to be estimated, baseline drift-free motion state data can be obtained, achieving low-frequency noise filtering. This data is then input into a feature extractor, which performs feature extraction on the baseline drift-free motion state data to extract preliminary features to be estimated. These features to be estimated are input into the TPN module. The TPN module normalizes data information in the time dimension to ensure the stability of subsequent network inputs and help subsequent models better focus on dependencies in time series data. The features with time-aware characteristics processed by the TPN module enter the micro-motion estimation module, which performs micro-motion estimation based on these features and outputs the micro-motion estimation results.

[0147] In the embodiments of this application, an innovative multi-stage decoupling strategy is employed to achieve data noise reduction. By addressing different types of noise in a targeted and phased manner, the noise reduction effect and model performance are significantly improved. First, in the first stage, a baseline filtering mechanism is designed to address low-frequency interference noise caused by external interference. This mechanism dynamically adjusts the baseline based on signal changes, effectively suppressing the drift characteristics of low-frequency noise and accurately filtering out such noise interference, laying a clean data foundation for subsequent processing. The second stage focuses on interference noise introduced by human motion. This stage introduces time domain information to deeply explore the temporal characteristics of data changes. By capturing the dynamic changes in data over long time spans, this method not only comprehensively handles complex and variable interference noise but also greatly enhances the model's adaptability to time series data. In complex motion scenarios, whether rapid and intense body movements or subtle and continuous posture changes, this method can stably and efficiently handle noise, demonstrating excellent robustness and providing reliable support for tasks such as micro-motion sensing based on inertial sensors (IMUs).

[0148] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0149] In one embodiment, a data noise reduction device is provided, which corresponds to the data noise reduction method in the above embodiment. Figure 7 As shown, the data denoising device includes a first denoising unit 10, a second denoising unit 20, an alignment loss determining unit 30, an estimation loss determining unit 40 and a model iterative updating unit 50. Each functional module is described in detail as follows:

[0150] A first noise reduction unit 10 is configured to perform baseline filtering on the motion state sample data to obtain motion state sample data with baseline drift removed, wherein the motion state sample data with baseline drift removed includes static data and motion data;

[0151] a second denoising unit 20, configured to input the static data and the motion data into a preset domain adaptation model, and perform domain alignment processing and small motion estimation processing, respectively, to obtain a domain alignment result and a small motion estimation result, wherein the domain alignment result includes a static global feature and a motion global feature;

[0152] an alignment loss determining unit 30, configured to determine an alignment loss based on the static global feature and the motion global feature;

[0153] an estimated loss determining unit 40, configured to determine an estimated loss based on the micro-motion estimation result;

[0154] The model iterative updating unit 50 is configured to iteratively update the preset domain adaptation model based on the alignment loss and the estimation loss.

[0155] In one embodiment of the present application, the second noise reduction unit 20 is further configured to:

[0156] Extracting features from the static data and the motion data to obtain static features and motion features;

[0157] Normalizing the static features and the motion features in a time dimension to obtain standard static features and standard motion features;

[0158] Based on the category token, domain aligning the standard static features and the standard motion features to obtain the static global features and the motion global features;

[0159] Based on the standard static features and the standard motion features, a slight motion estimation result is obtained.

[0160] In one embodiment of the present application, the second noise reduction unit 20 is further configured to:

[0161] dividing the static data and the motion data into a plurality of static time series segments and a plurality of motion time series segments respectively;

[0162] extracting at least one target time series segment from the plurality of stationary time series segments and the plurality of moving time series segments according to a preset timestamp label;

[0163] According to preset extraction rules, feature extraction is performed on the target time series segments respectively by a preset feature extractor to obtain the static features and the motion features.

[0164] In one embodiment of the present application, the second noise reduction unit 20 is further configured to:

[0165] Determining the mean and standard deviation of the static features and the motion features;

[0166] Based on the mean value and the standard deviation, the static feature and the motion feature are standardized to obtain a standard static feature and a standard motion feature.

[0167] In one embodiment of the present application, the second noise reduction unit 20 is further configured to:

[0168] Adding category tokens to the standard static feature and the standard motion feature respectively to obtain a static category token feature and a motion category token feature, wherein the category tokens are used to identify domain features that change dynamically over time;

[0169] The stationary category token features and the motion category token features are input into a preset encoder for processing to obtain a stationary global feature and a motion global feature.

[0170] In one embodiment of the present application, the second noise reduction unit 20 is further configured to:

[0171] Expanding the prefix labels in the feature time series corresponding to the standard static features and the standard motion features to obtain an expanded label sequence, wherein the prefix labels refer to labels corresponding to prefix motion state data, and the prefix motion state data are motion state sample data corresponding to the first preset features in the feature time series;

[0172] Inputting the expanded label sequence into the encoder layer for self-attention calculation;

[0173] The label sequence calculated after self-attention is input into the decoder layer for feature fusion and decoding processing to obtain the micro-motion estimation result.

[0174] In one embodiment of the present application, the apparatus further includes an actual prediction unit, configured to:

[0175] After iteratively updating the preset domain adaptation model, the method further includes:

[0176] Acquiring motion state data to be estimated;

[0177] Performing baseline filtering on the motion state data to be estimated to obtain the motion state data to be estimated with baseline drift removed;

[0178] performing feature extraction on the motion state data to be estimated with baseline drift removed to obtain features to be estimated;

[0179] Standardizing the features to be estimated in a preset time dimension to obtain standard features to be estimated;

[0180] Based on the standard feature to be estimated, a micro-motion estimation result corresponding to the motion state data to be estimated is obtained.

[0181] In the embodiments of this application, an innovative multi-stage decoupling strategy is employed to achieve data noise reduction. By addressing different types of noise in a targeted, phased manner, the noise reduction effect and model performance are significantly improved. In the first stage, a baseline filtering mechanism is designed to address low-frequency interference noise caused by external interference. This mechanism dynamically adjusts the baseline based on signal changes, effectively suppressing the drift characteristics of low-frequency noise and accurately filtering out such noise interference, laying a clean data foundation for subsequent processing. The second stage focuses on interference noise introduced by human motion. This stage introduces time domain information to deeply explore the temporal characteristics of data changes. By capturing the dynamic changes in data over long time spans, it not only comprehensively handles complex and variable interference noise but also greatly enhances the model's adaptability to time series data. In complex motion scenarios, whether rapid and intense body movements or subtle and continuous posture changes, this method can stably and efficiently handle noise, demonstrating excellent robustness and providing reliable support for tasks such as micro-motion sensing based on inertial sensors (IMUs).

[0182] The specific definition of the data denoising device can be found in the definition of the data denoising method above and will not be repeated here. Each module in the above-mentioned data denoising device can be implemented in whole or in part through software, hardware, or a combination thereof. Each of the above-mentioned modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory of the computer device in software form, so that the processor can call and execute the corresponding operations of each of the above modules.

[0183] In one embodiment, a computer device is provided. The computer device may be a terminal device, and its internal structure diagram may be as follows: Figure 8 As shown. The computer device includes a processor, a memory, and a network interface connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a readable storage medium. The readable storage medium stores computer-readable instructions. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer-readable instructions are executed by the processor, a data noise reduction method is implemented. The readable storage medium provided in this embodiment includes a non-volatile readable storage medium and a volatile readable storage medium.

[0184] In an embodiment of the present application, a computer device is provided, including a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor. When the processor executes the computer-readable instructions, the steps of the above-mentioned data denoising method are implemented.

[0185] In an embodiment of the application, a readable storage medium is provided, which stores computer-readable instructions. When the computer-readable instructions are executed by a processor, the steps of the data denoising method described above are implemented.

[0186] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing related hardware through computer-readable instructions. The computer-readable instructions can be stored in a non-volatile readable storage medium or a volatile readable storage medium. When the computer-readable instructions are executed, they can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0187] Those skilled in the art will clearly understand that for the sake of convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0188] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present application, and should all be included in the scope of protection of the present application.

Claims

1. A data denoising method, characterized in that: The method comprises: Performing baseline filtering on the motion state sample data to obtain motion state sample data with baseline drift removed, wherein the motion state sample data with baseline drift removed includes static data and motion data; Inputting the static data and the motion data into a preset domain adaptation model, performing domain alignment processing and small motion estimation processing respectively, to obtain a domain alignment result and a small motion estimation result, wherein the domain alignment result includes a static global feature and a motion global feature; determining an alignment loss based on the static global feature and the motion global feature; determining an estimated loss based on the micro-motion estimation result; Iteratively updating the preset domain adaptation model based on the alignment loss and the estimation loss; The step of inputting the static data and the motion data into a preset domain adaptation model and performing domain alignment processing and small motion estimation processing respectively includes: Extracting features from the static data and the motion data to obtain static features and motion features; Normalizing the static features and the motion features in a time dimension to obtain standard static features and standard motion features; Based on the category token, domain aligning the standard static features and the standard motion features to obtain the static global features and the motion global features; The slight motion estimation result is predicted based on the standard static feature and the standard motion feature.

2. The data denoising method according to claim 1, wherein: The extracting features of the static data and the motion data respectively to obtain static features and motion features includes: dividing the static data and the motion data into a plurality of static time series segments and a plurality of motion time series segments respectively; extracting at least one target time series segment from the plurality of stationary time series segments and the plurality of moving time series segments according to a preset timestamp label; According to preset extraction rules, feature extraction is performed on the target time series segments respectively by a preset feature extractor to obtain the static features and the motion features.

3. The data denoising method according to claim 2, wherein: The standardization of the static features and the motion features in the time dimension to obtain standard static features and standard motion features includes: Determining the mean and standard deviation of the static features and the motion features; Based on the mean value and the standard deviation, the static feature and the motion feature are standardized to obtain a standard static feature and a standard motion feature.

4. The data denoising method according to claim 1, wherein: The performing domain alignment on the standard static features and the standard motion features based on the category token to obtain the static global features and the motion global features includes: Adding category tokens to the standard static feature and the standard motion feature respectively to obtain a static category token feature and a motion category token feature, wherein the category tokens are used to identify domain features that change dynamically over time; The stationary category token features and the motion category token features are input into a preset encoder for processing to obtain the stationary global features and the motion global features.

5. The data denoising method according to claim 1, wherein: The predicting and obtaining the slight motion estimation result based on the standard static feature and the standard motion feature includes: Expanding the prefix labels in the feature time series corresponding to the standard static features and the standard motion features to obtain an expanded label sequence, wherein the prefix labels refer to labels corresponding to prefix motion state data, and the prefix motion state data are motion state sample data corresponding to the first preset features in the feature time series; Inputting the expanded label sequence into the encoder layer for self-attention calculation; The label sequence calculated after self-attention is input into the decoder layer for feature fusion and decoding processing to obtain the micro-motion estimation result.

6. The data denoising method according to any one of claims 1 to 5, wherein: After iteratively updating the preset domain adaptation model, the method further includes: Acquiring motion state data to be estimated; Performing baseline filtering on the motion state data to be estimated to obtain the motion state data to be estimated with baseline drift removed; performing feature extraction on the motion state data to be estimated with baseline drift removed to obtain features to be estimated; Standardizing the features to be estimated in a preset time dimension to obtain standard features to be estimated; Based on the standard feature to be estimated, a micro-motion estimation result corresponding to the motion state data to be estimated is obtained.

7. A data noise reduction device, characterized in that: The device comprises: a first noise reduction unit, configured to perform baseline filtering on the motion state sample data to obtain motion state sample data with baseline drift removed, wherein the motion state sample data with baseline drift removed includes static data and motion data; a second denoising unit, configured to input the static data and the motion data into a preset domain adaptation model, and perform domain alignment processing and small motion estimation processing respectively, to obtain a domain alignment result and a small motion estimation result, wherein the domain alignment result includes a static global feature and a motion global feature; an alignment loss determining unit, configured to determine an alignment loss based on the static global feature and the motion global feature; an estimated loss determining unit, configured to determine an estimated loss based on the micro-motion estimation result; A model iterative updating unit, configured to iteratively update the preset domain adaptation model based on the alignment loss and the estimation loss; Wherein, the second noise reduction unit is further used for: Extracting features from the static data and the motion data to obtain static features and motion features; Normalizing the static features and the motion features in a time dimension to obtain standard static features and standard motion features; Based on the category token, domain aligning the standard static features and the standard motion features to obtain the static global features and the motion global features; The slight motion estimation result is predicted based on the standard static feature and the standard motion feature.

8. A computer device comprising a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor, wherein: When the processor executes the computer-readable instructions, the steps of the data denoising method according to any one of claims 1 to 6 are implemented.

9. A readable storage medium storing computer-readable instructions, characterized in that: When the computer-readable instructions are executed by a processor, the steps of the data denoising method according to any one of claims 1 to 6 are implemented.

Images