Video stream image optimization method and device

Abstract

The embodiment of the application provides a video stream image optimization method and device, a dynamic blur kernel generation mechanism is innovatively designed, space-time decoupling and physical constraints are used, and data processing safety is ensured. A content-motion double-branch network is constructed, residual connection and space-time transformation are combined, and a reliable feature reconstruction system is established. A physical constraint modifier is introduced, adaptive fusion and rule correction are used, user privacy is protected, and high-quality video output is provided. The method effectively solves the deficiencies of traditional technologies in blur processing, feature decoupling and image optimization.

Description

Technical Field

[0001] This application relates to the field of image processing, specifically to a method and apparatus for optimizing video stream images. Background Technology

[0002] Existing video stream image optimization methods have significant shortcomings. Traditional systems perform poorly in dynamic blur processing, failing to effectively combine optical flow and motion features, thus affecting image quality.

[0003] Furthermore, existing technologies face bottlenecks in decoupling spatiotemporal features. Most systems lack robust content-motion separation mechanisms and physical constraint strategies, resulting in unsatisfactory motion region reconstruction.

[0004] Existing systems have technical shortcomings in eliminating motion blur and ghosting. They lack sufficient consideration of physical laws, making accurate image optimization through feature fusion difficult and impacting the visual experience. Solving these problems is crucial for improving video quality. Summary of the Invention

[0005] To address the problems in the prior art, this application provides a video stream image optimization method and apparatus, which can effectively solve the shortcomings of traditional techniques in blur processing, feature decoupling and image optimization.

[0006] To solve at least one of the above problems, this application provides the following technical solution:

[0007] In a first aspect, this application provides a video stream image optimization method, including:

[0008] A dynamic blur kernel is generated based on the motion speed of the video image and the target object, as well as optical flow parameters. Multiple convolutional units are used to perform optical flow modeling and occlusion detection on the video image. The video image and the blur kernel are then input into the physical perception module for data preprocessing to generate standardized features with motion trajectory constraints.

[0009] The standardized features are input into a spatiotemporal decoupling network, which includes a content branch and a motion branch. The content branch reconstructs the texture features of static regions, and the motion branch reconstructs the motion features of dynamic regions. The reconstruction results of the content branch and the motion branch are fused based on a cross-layer feature connection mechanism to generate spatiotemporal decoupling features.

[0010] The spatiotemporal decoupling features are input into a physical constraint corrector. The physical constraint corrector generates a dynamic weight matrix based on the motion trajectory constraints, performs adaptive fusion and physical law correction on the spatiotemporal decoupling features, and optimizes the corrected features through residual connections and a spatiotemporal transformation network to generate a video output that eliminates motion blur and ghosting.

[0011] Furthermore, it also includes: performing multi-scale feature matching on continuous video frames based on the pyramid optical flow algorithm to construct the motion trajectory matrix of the target object, inputting the motion trajectory matrix into a multi-layer convolutional neural network for feature extraction, wherein the multi-layer convolutional neural network includes a spatial attention module and a channel attention module, and calculating dynamic blur kernel parameters based on the spatiotemporal correlation of the motion trajectory matrix;

[0012] An optical flow model is constructed using multiple dilated convolutional units, each containing a residual connection layer and a normalization layer. Frame-by-frame optical flow estimation is performed on the video image sequence. An occlusion detection map is generated based on the consistency analysis of the optical flow vector field and the motion trajectory matrix. The occlusion detection map is then weighted and combined with the dynamic blur kernel parameters to obtain a dynamic blur kernel that considers the occlusion relationship.

[0013] Furthermore, it also includes: inputting the video image and the dynamic blur kernel into a physical perception module, the physical perception module comprising multiple deconvolution layers and upsampling layers, performing reverse deblurring processing on the video image based on the dynamic blur kernel, the deconvolution layers iteratively reconstructing the image features, and the upsampling layers improving the resolution of the reconstructed features to generate an initial feature map;

[0014] A motion trajectory constraint model is constructed using a bidirectional long short-term memory network. The motion trajectory constraint model includes a forward propagation layer and a backward propagation layer. The initial feature map is input into the motion trajectory constraint model. The forward propagation layer extracts forward temporal features, and the backward propagation layer extracts backward temporal features. Based on the correspondence between the forward temporal features and the backward temporal features, trajectory constraint conditions are constructed. The initial feature map is batch normalized to generate standardized features with motion trajectory constraints.

[0015] Furthermore, it also includes: performing static region segmentation on the standardized features, constructing a non-local attention module, wherein the non-local attention module includes a feature transformation sub-network and a similarity calculation sub-network, wherein the feature transformation sub-network maps the standardized features to a high-dimensional feature space, and the similarity calculation sub-network calculates the similarity matrix of each pixel in the feature map based on cosine distance to generate a static region feature map;

[0016] The static region feature map is input into a texture reconstruction network, which includes multiple dense connection modules and skip connection modules. The dense connection modules use multi-scale convolutional kernels to extract multi-level texture information, and the skip connection modules fuse texture information at different levels. Based on a self-attention mechanism, the fused features are enhanced in detail to generate reconstructed static region texture features.

[0017] Furthermore, it also includes: constructing a multi-scale motion encoder, wherein the multi-scale motion encoder includes a pyramid pooling module and a dynamic convolution module, wherein the pyramid pooling module performs multi-scale decomposition on the standardized features, and the dynamic convolution module adaptively adjusts the convolution kernel parameters based on the motion amplitude to reconstruct features of the dynamic region and generate dynamic region motion features;

[0018] The static region texture features and the dynamic region motion features are input into a feature fusion network. The feature fusion network includes a channel recalibration module and a spatial reweighting module. The channel recalibration module adaptively selects feature channels, and the spatial reweighting module adjusts feature weights based on a position sensitivity map. The recalibrated features are cascaded and combined through cross-layer residual connections to generate spatiotemporally decoupled features.

[0019] Furthermore, it also includes: constructing a motion physical constraint model, the motion physical constraint model including an acceleration analysis module and a velocity prediction module, the acceleration analysis module calculating the acceleration distribution map of the target object based on the motion trajectory constraints, and the velocity prediction module predicting the motion state at the next moment based on the acceleration distribution map, and generating physical constraint parameters;

[0020] The spatiotemporal decoupling features and the physical constraint parameters are input into a weight generation network. The weight generation network includes a feature association module and a weight calculation module. The feature association module performs temporal consistency analysis on the spatiotemporal decoupling features. The weight calculation module evaluates and scores the feature consistency based on the physical constraint parameters. The evaluation scores are converted into weight coefficients through normalization processing to generate a dynamic weight matrix.

[0021] Furthermore, it also includes: inputting the spatiotemporal decoupling features and the dynamic weight matrix into a feature correction network, wherein the feature correction network includes an adaptive fusion module and a physical constraint module, wherein the adaptive fusion module performs weighted combination of the spatiotemporal decoupling features based on the dynamic weight matrix, and the physical constraint module performs regular correction on the fused features according to motion trajectory constraints to generate a corrected feature map;

[0022] An optimization and enhancement network is constructed, which includes a residual learning module and a spatiotemporal transformation module. The residual learning module performs detail compensation on the correction feature map, and the spatiotemporal transformation module performs spatial alignment and temporal smoothing on the compensated features based on deformable convolution. Optimized video features are generated through multi-layer receptive field fusion and feature recalibration. The optimized video features are then deconvolved to reconstruct the video output with motion blur and ghosting eliminated.

[0023] Secondly, this application provides a video stream image optimization apparatus, comprising:

[0024] The preprocessing module is used to generate a dynamic blur kernel based on the motion speed of the video image and the target object, as well as optical flow parameters. Multiple convolutional units are used to perform optical flow modeling and occlusion detection on the video image. The video image and the blur kernel are then input into the physical perception module for data preprocessing to generate standardized features with motion trajectory constraints.

[0025] The spatiotemporal decoupling module is used to input the standardized features into the spatiotemporal decoupling network. The spatiotemporal decoupling network includes a content branch and a motion branch. The content branch reconstructs the texture features of static regions, and the motion branch reconstructs the motion features of dynamic regions. The reconstruction results of the content branch and the motion branch are fused based on the cross-layer feature connection mechanism to generate spatiotemporal decoupling features.

[0026] The image optimization module is used to input the spatiotemporal decoupling features into the physical constraint corrector. The physical constraint corrector generates a dynamic weight matrix based on the motion trajectory constraints, performs adaptive fusion and physical law correction on the spatiotemporal decoupling features, and optimizes the corrected features through residual connections and spatiotemporal transformation networks to generate video output with eliminated motion blur and ghosting.

[0027] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the video stream image optimization method.

[0028] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the video stream image optimization method described above.

[0029] Fifthly, this application provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the video stream image optimization method.

[0030] As can be seen from the above technical solution, this application provides a video stream image optimization method and apparatus. Through an innovative design of a dynamic blur kernel generation mechanism, and by using spatiotemporal decoupling and physical constraints, it ensures data processing security. A content-motion dual-branch network is constructed, combining residual connections and spatiotemporal transformation to establish a reliable feature reconstruction system. A physical constraint corrector is introduced, and through adaptive fusion and regular correction, it provides high-quality video output while protecting user privacy. This method effectively solves the shortcomings of traditional techniques in blur processing, feature decoupling, and image optimization. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a flowchart illustrating the video stream image optimization method in an embodiment of this application.

[0033] Figure 2 This is a structural diagram of the video stream image optimization device in the embodiments of this application;

[0034] Figure 3 This is a schematic diagram of the structure of the electronic device in the embodiments of this application.

[0035] Figure label:

[0036] Electronic device 9600, central processing unit 9100, memory 9140, communication module 9110, input unit 9120, audio processor 9130, display 9160, power supply 9170, buffer memory 9141, application / function storage unit 9142, data storage unit 9143, driver storage unit 9144, antenna 9111, speaker 9131, microphone 9132. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0038] The acquisition, storage, use, and processing of data in this application all comply with the relevant provisions of national laws and regulations.

[0039] To address the problems existing in current technologies, this application provides a video stream image optimization method and apparatus. It innovatively designs a dynamic blur kernel generation mechanism, ensuring data processing security through spatiotemporal decoupling and physical constraints. A content-motion dual-branch network is constructed, combining residual connections and spatiotemporal transformation to establish a reliable feature reconstruction system. A physical constraint corrector is introduced, providing high-quality video output while protecting user privacy through adaptive fusion and regular correction. This method effectively solves the shortcomings of traditional techniques in blur processing, feature decoupling, and image optimization.

[0040] To effectively address the shortcomings of traditional techniques in areas such as blurring, feature decoupling, and image optimization, this application provides an embodiment of a video stream image optimization method, see [link to embodiment]. Figure 1 The video stream image optimization method specifically includes the following:

[0041] Step S101: Generate a dynamic blur kernel based on the motion speed of the video image and the target object and the optical flow parameters. Use multiple convolutional units to perform optical flow modeling and occlusion detection on the video image. Input the video image and the blur kernel into the physical perception module for data preprocessing to generate standardized features with motion trajectory constraints.

[0042] Optionally, this embodiment innovatively designs a feature processing scheme based on physical perception to address problems such as blurred moving targets, complex occlusion relationships, and inaccurate feature representation in video image processing. For feature processing evaluation, this embodiment designs a blur kernel scoring formula: Blur_Score=α×(Motion_Speed)+β×(Flow_Consistency)+γ×(Occlusion_Weight), where α, β, and γ are weighting coefficients, representing the influence of motion speed, optical flow consistency, and occlusion degree, respectively. Feature quality scoring uses: Feature_Score=(Physical_Constraint×Temporal_Consistency) / (Spatial_Error+Noise_Level), used to evaluate the expressive power of standardized features.

[0043] This embodiment first achieves accurate modeling of motion blur through a deeply optimized dynamic blur kernel generation mechanism. When processing video images, the system analyzes the motion characteristics of the target object, including velocity magnitude, direction changes, and acceleration distribution. For fast-moving targets, the blur kernel generation process considers the nonlinear characteristics of the motion trajectory, dynamically adjusting the shape and size of the blur kernel by establishing a velocity-blur mapping relationship. When processing complex scenes, the system captures the motion details of the target by analyzing optical flow parameters, including local deformation and overall displacement. The blur kernel design employs an adaptive strategy, dynamically adjusting the intensity of the blur effect according to the target's motion state. For multi-target scenes, the system establishes a target motion correlation graph to describe the motion dependencies between different targets, ensuring the consistency of the generated blur kernel. This optimized blur kernel generation mechanism significantly improves the accuracy of motion blur representation.

[0044] This embodiment innovatively implements optical flow modeling and occlusion detection mechanisms. The design of multiple convolutional units adopts a hierarchical structure, with different levels of convolutional units responsible for feature extraction at different scales. During optical flow modeling, the system constructs a pixel-level motion vector field through spatial gradient analysis and temporal correlation calculation. The convolutional unit design employs a multi-scale strategy, with shallow convolutional units responsible for extracting local motion features and deep convolutional units focusing on global motion patterns. For occlusion detection, the system identifies potential occlusion boundaries by analyzing the discontinuities and motion consistency of the optical flow field. Temporal constraints are introduced during occlusion detection; by analyzing the feature correspondence between consecutive frames, the accuracy of occlusion detection is improved. This optimized feature extraction mechanism provides reliable input for subsequent data preprocessing.

[0045] This embodiment achieves feature standardization through an optimized physical perception mechanism. The physical perception module employs a multi-level feature extraction strategy when processing input data. For video images, the system establishes a physical model to describe the motion patterns of the target, including inertia, drag, and external force influences. The application of the fuzzy kernel considers physical constraints, ensuring that the deblurring results conform to actual motion characteristics. Motion trajectory constraints are introduced during feature extraction; a trajectory prediction model is established to constrain the temporal consistency of features. This optimized physical perception mechanism significantly improves the physical plausibility of the features.

[0046] This embodiment establishes a complete feature processing framework. The generation of standardized features employs a multi-stage optimization strategy. First, the motion characteristics of the target are analyzed using a physical model, and then a feature extraction network is constructed based on these motion characteristics. The system establishes a feature quality evaluation mechanism to monitor the representation effect of features in real time and supports dynamic adjustment of the feature extraction strategy. This optimized feature processing mechanism provides high-quality feature support for subsequent video processing tasks. This embodiment provides an innovative solution for feature processing. Through optimized blur kernel generation and physical perception, the system achieves end-to-end optimization from raw video to standardized features. This reliable processing mechanism significantly improves feature quality, especially when processing fast-moving and complex occlusion scenes, where the generated features exhibit stronger physical consistency and temporal relevance.

[0047] The innovation of this embodiment lies primarily in the adaptive generation of the dynamic blur kernel and the physical constraint representation of features. Through multi-level feature extraction and physical perception mechanisms, the system achieves accurate modeling of motion blur and standardized representation of features, providing reliable feature support for video image processing tasks. Especially when processing high-speed motion scenes, this physical constraint-based feature processing method significantly improves the system's robustness and accuracy.

[0048] Step S102: Input the standardized features into the spatiotemporal decoupling network. The spatiotemporal decoupling network includes a content branch and a motion branch. The content branch reconstructs the texture features of static regions, and the motion branch reconstructs the motion features of dynamic regions. Based on the cross-layer feature connection mechanism, the reconstruction results of the content branch and the motion branch are fused to generate spatiotemporal decoupling features.

[0049] Optionally, this embodiment addresses the problems of feature confusion between static and dynamic regions, complex spatiotemporal feature correlation, and insufficient feature fusion in video processing by innovatively designing a spatiotemporal decoupling network based on a two-branch architecture. For feature decoupling evaluation, this embodiment designs a decoupling quality scoring formula: Decoupling_Score=α×(Static_Fidelity)+β×(Motion_Accuracy)+γ×(Feature_Independence), where α, β, and γ are weight coefficients representing static fidelity, motion accuracy, and feature independence, respectively. The feature fusion score uses: Fusion_Score=(Cross_Layer_Correlation×Feature_Complementarity) / (Information_Loss+Interference_Level) to evaluate the effectiveness of feature fusion.

[0050] This embodiment first achieves accurate reconstruction of static regions through a deeply optimized content branch. When processing standardized features, the content branch employs a multi-level feature extraction architecture, with different levels responsible for capturing texture information at different scales. For static region identification, the system constructs a static region confidence map by analyzing the temporal consistency and spatial correlation of features. An adaptive receptive field mechanism is introduced during feature extraction, dynamically adjusting the convolutional kernel size to achieve accurate capture of textures at different scales. When processing complex textures, the system establishes a multi-scale feature pyramid to achieve hierarchical representation of texture details. The reconstruction process employs an iterative optimization strategy, dynamically adjusting reconstruction parameters in each iteration based on the difference between the current reconstruction result and the original features. To maintain texture continuity, the system introduces local consistency constraints to ensure smooth transitions in local regions. This optimized static feature reconstruction mechanism significantly improves the accuracy of texture representation.

[0051] This embodiment innovatively implements a feature reconstruction mechanism for motion branches. When processing dynamic regions, the motion branch employs a time-aware feature extraction strategy. For motion feature extraction, the system constructs a motion feature field by analyzing feature changes between consecutive frames. The feature extraction network adopts a multi-resolution design, with low-resolution layers responsible for capturing the overall motion trend and high-resolution layers focusing on local motion details. When handling complex motion, the system decomposes complex motion into basic motion components by establishing a motion decomposition model. Motion continuity constraints are introduced during reconstruction, ensuring the temporal smoothness of the reconstruction results by analyzing the motion consistency between adjacent frames. For fast-moving regions, the system increases the temporal resolution of feature extraction through an adaptive sampling strategy, improving the accuracy of motion capture. This optimized dynamic feature reconstruction mechanism provides a reliable motion representation for subsequent feature fusion.

[0052] This embodiment achieves effective feature fusion through an optimized cross-layer feature connection mechanism. The feature fusion process employs a multi-level connection strategy, combining features from different levels using adaptive weights. The fusion network design considers feature complementarity, guiding selective feature fusion by establishing a feature correlation graph. To reduce information redundancy, a feature filtering mechanism is introduced, retaining the most discriminative feature combinations. A residual learning strategy is used during the fusion process, preserving original feature information through skip connections. This optimized fusion mechanism significantly improves the expressive power of spatiotemporally decoupled features.

[0053] This embodiment establishes a complete feature decoupling framework. The generation of spatiotemporally decoupled features employs a multi-stage optimization strategy. First, static and dynamic features are extracted separately through a dual-branch network. Then, effective feature fusion is achieved through a cross-layer connection mechanism. The system establishes a feature quality evaluation mechanism to monitor the decoupling effect in real time and supports dynamic adjustment of network parameters. This optimized decoupling mechanism provides high-quality feature support for subsequent video processing tasks. This embodiment offers an innovative solution for feature decoupling. Through an optimized dual-branch network and feature fusion, the system achieves end-to-end optimization from standardized features to spatiotemporally decoupled features. This reliable decoupling mechanism significantly improves the expressive power of features, especially when processing complex scenes, where decoupled features exhibit stronger discriminative and adaptive capabilities.

[0054] The innovation of this embodiment lies primarily in the decoupled representation of spatiotemporal features and the adaptive fusion of features. By constructing a dual-branch network and a cross-layer connection mechanism, the system achieves accurate modeling of both static and dynamic regions, providing reliable feature support for video processing tasks. Especially when processing scenes with complex motion patterns, this decoupled feature processing method significantly improves the system's expressive power and processing accuracy.

[0055] Step S103: Input the spatiotemporal decoupling features into the physical constraint corrector. The physical constraint corrector generates a dynamic weight matrix based on the motion trajectory constraints, performs adaptive fusion and physical law correction on the spatiotemporal decoupling features, and optimizes the corrected features through residual connections and spatiotemporal transformation networks to generate video output with eliminated motion blur and ghosting.

[0056] Optionally, this embodiment innovatively designs a feature correction scheme based on physical constraints to address the issues of motion blur and ghosting in video processing. For feature correction evaluation, this embodiment designs a correction effect scoring formula: Correction_Score=α×(Physics_Constraint)+β×(Feature_Consistency)+γ×(Motion_Smoothness), where α, β, and γ are weighting coefficients, representing the degree of physical constraint satisfaction, feature consistency, and motion smoothness, respectively. The output quality score is calculated as: Quality_Score=(Detail_Preservation×Artifact_Removal) / (Motion_Blur+Ghost_Effect), used to evaluate the quality of the video output.

[0057] This embodiment first achieves accurate feature correction through a deeply optimized physical constraint correction mechanism. When processing spatiotemporally decoupled features, the physical constraint corrector employs a multi-level constraint modeling strategy. Based on the motion trajectory constraints obtained in the preceding steps, the corrector first establishes a physical motion model to describe the target's motion patterns. For high-speed moving targets, the system predicts the target's motion state by analyzing acceleration distribution and velocity changes. The construction of physical constraints considers multiple aspects, including motion continuity, physical feasibility, and environmental interactivity. Based on these constraints, the system generates a dynamic weight matrix for adaptive feature fusion. The weight matrix generation process employs an attention mechanism, dynamically adjusting the weight allocation in different regions by calculating the correlation between features and physical constraints. This optimized constraint modeling mechanism significantly improves the accuracy of feature correction.

[0058] This embodiment innovatively implements a feature fusion and correction mechanism. In the adaptive fusion stage, the system uses a dynamic weight matrix to weight and combine spatiotemporally decoupled features. The fusion process considers the spatial correlation and temporal consistency of features, describing the dependencies between different features by establishing a feature correlation graph. For regions that may produce motion blur, the system dynamically adjusts the fusion weights by analyzing motion trajectories and velocity fields to suppress the motion blur effect. The physical law correction process employs an iterative optimization strategy, adjusting features based on physical constraints in each iteration. Temporal smoothing constraints are introduced during the correction process to ensure the temporal continuity of the correction results. This optimized fusion mechanism provides a reliable foundation for subsequent feature optimization.

[0059] This embodiment achieves final feature optimization through an optimized network structure. The residual connection and spatiotemporal transformation network are designed with a multi-branch architecture. The residual learning branch is responsible for compensating and enhancing detailed information, preserving high-frequency details of the image by establishing residual mapping relationships. The spatiotemporal transformation branch achieves spatial alignment and temporal smoothing of features through deformable convolution. During the transformation process, the system employs a multi-scale receptive field fusion strategy to capture image information at different scales. The feature recalibration process considers physical constraints to ensure that the optimization results conform to actual motion laws. This optimized network structure significantly improves the quality of the video output.

[0060] This embodiment establishes a complete video optimization framework. The generation of video output employs a multi-stage optimization strategy. First, features are corrected using a physical constraint corrector, and then details are enhanced using an optimization network. The system establishes a video quality evaluation mechanism to monitor the output effect in real time and supports dynamic adjustments to the optimization strategy. This optimization mechanism significantly improves video quality, especially when handling complex motion scenes, resulting in better visual effects and physical plausibility. This embodiment provides an innovative solution for video optimization. Through optimized physical constraints and feature correction, the system achieves end-to-end optimization from feature processing to video output. This reliable optimization mechanism significantly improves video quality, especially when handling fast motion and complex scenes, resulting in output videos with clearer details and fewer artifacts.

[0061] The innovation of this embodiment lies primarily in the dynamic application of physical constraints and the adaptability of feature optimization. Through a physical constraint corrector and an optimization network, the system achieves feature correction and video quality enhancement based on physical laws, providing reliable optimization support for video processing tasks. Particularly when processing high-speed motion scenes, this physical constraint-based optimization method significantly improves system performance and effectively eliminates visual artifacts such as motion blur and ghosting. For example, when processing sports videos, the system can accurately capture the rapid movements of athletes while maintaining image clarity and continuity.

[0062] As described above, the video stream image optimization method provided in this application can ensure data processing security through an innovative design of a dynamic blur kernel generation mechanism, spatiotemporal decoupling, and physical constraints. A content-motion dual-branch network is constructed, combining residual connections and spatiotemporal transformation to establish a reliable feature reconstruction system. A physical constraint corrector is introduced, and through adaptive fusion and regular correction, high-quality video output is provided while protecting user privacy. This method effectively addresses the shortcomings of traditional techniques in blur processing, feature decoupling, and image optimization.

[0063] In one embodiment of the video stream image optimization method of this application, it may further include the following:

[0064] Step S201: Perform multi-scale feature matching on continuous video frames based on the pyramid optical flow algorithm to construct the motion trajectory matrix of the target object. Input the motion trajectory matrix into a multi-layer convolutional neural network for feature extraction. The multi-layer convolutional neural network includes a spatial attention module and a channel attention module. Calculate the dynamic blur kernel parameters based on the spatiotemporal correlation of the motion trajectory matrix.

[0065] Step S202: An optical flow model is constructed using multiple dilated convolutional units, each containing a residual connection layer and a normalization layer. Frame-by-frame optical flow estimation is performed on the video image sequence. An occlusion detection map is generated based on the consistency analysis of the optical flow vector field and the motion trajectory matrix. The occlusion detection map is then weighted and combined with the dynamic blur kernel parameters to obtain a dynamic blur kernel that considers the occlusion relationship.

[0066] Optionally, this embodiment addresses the problems of inaccurate moving target tracking, imperfect occlusion detection, and unreasonable blur kernel generation in traditional video processing by innovatively designing a feature processing scheme based on multi-scale optical flow. For optical flow estimation and evaluation, this embodiment designs a matching quality scoring formula: Flow_Score=α×(Feature_Similarity)+β×(Motion_Consistency)+γ×(Scale_Coverage), where α, β, and γ are weighting coefficients, representing feature similarity, motion consistency, and scale coverage, respectively. The occlusion detection score uses: Occlusion_Score=(Flow_Discontinuity×Trajectory_Consistency) / (Noise_Level+Estimation_Error) to evaluate the accuracy of occlusion detection.

[0067] This embodiment first achieves multi-scale feature matching through a depth-optimized pyramid optical flow algorithm. When processing consecutive video frames, the system constructs a multi-level image pyramid, with each level corresponding to a different spatial resolution. For tracking moving targets, the system adopts a top-down matching strategy, first performing coarse matching at low-resolution levels to obtain initial motion estimates, and then gradually performing refined matching at high-resolution levels. Local deformation constraints are introduced during feature matching, describing the non-rigid deformation of the target by establishing a local affine transformation model. For fast-moving targets, the system dynamically adjusts the search range of feature matching through an adaptive search window mechanism. The construction of the motion trajectory matrix adopts temporal consistency constraints, ensuring the continuity of the trajectory by analyzing the feature correspondence between consecutive frames. This optimized feature matching mechanism significantly improves the accuracy of motion estimation.

[0068] This embodiment innovatively implements an attention-enhanced feature extraction mechanism. When processing motion trajectory matrices, the multi-layer convolutional neural network employs an architecture combining spatial attention and channel attention. The spatial attention module adaptively highlights the feature representation of important regions by calculating the spatial response map of the feature map. For the localization of moving targets, the attention mechanism focuses on the spatial distribution and local structural features of the target. The channel attention module enhances discriminative motion features by learning the dependencies between feature channels. When dealing with complex backgrounds, the attention mechanism effectively suppresses background interference and improves the extraction quality of target features. The calculation of dynamic blur kernel parameters is based on spatiotemporal correlation analysis, and the motion trend of the target is predicted by establishing a motion state transition model. This optimized feature extraction mechanism provides reliable feature support for subsequent optical flow estimation.

[0069] This embodiment achieves accurate optical flow estimation through an optimized dilated convolutional network. The optical flow model is constructed using multiple dilated convolutional units, with kernels of varying dilation rates capturing motion information at different scales. The design of the residual connection layer ensures efficient transfer of feature information, and the skip connection mechanism reduces the gradient vanishing problem. The normalization layer employs a batch normalization strategy, improving network training stability by standardizing feature distribution. During frame-by-frame optical flow estimation, the system establishes temporal consistency constraints to ensure the temporal continuity of the optical flow field. Occlusion detection is implemented based on consistency analysis of the optical flow vector field and motion trajectory matrix. By identifying discontinuous regions in the optical flow field and interruptions in the trajectory, an occlusion detection map is generated. Finally, the dynamic blur kernel incorporates occlusion relationships into its generation process through a weighted combination mechanism. This optimized optical flow estimation mechanism significantly improves the accuracy of occlusion detection.

[0070] This embodiment establishes a complete blur kernel generation framework. Dynamic blur kernel generation considering occlusion relationships employs a multi-stage optimization strategy. First, motion information is obtained through multi-scale feature matching. Then, accurate modeling of the blur kernel is achieved based on optical flow estimation and occlusion detection. The system establishes a blur kernel quality evaluation mechanism to monitor the generation effect in real time and supports dynamic parameter adjustment. This optimized generation mechanism provides high-quality blur kernel representations for subsequent video processing tasks. This embodiment provides an innovative solution for blur kernel generation. Through optimized feature matching and optical flow estimation, the system achieves end-to-end optimization from video sequences to dynamic blur kernels. This reliable generation mechanism significantly improves the quality of blur kernels, especially when handling complex motion and occlusion scenes, where the generated blur kernels exhibit stronger adaptability and accuracy.

[0071] The innovation of this embodiment lies primarily in the adaptive processing of multi-scale feature matching and the accurate modeling of occlusion relationships. By constructing a multi-level feature processing network and an optimized optical flow estimation model, the system achieves accurate tracking of moving targets and accurate detection of occlusion relationships, providing reliable blur kernel support for video deblurring tasks. Especially when dealing with complex scenes, this multi-scale feature processing method significantly improves the robustness and accuracy of the system.

[0072] In one embodiment of the video stream image optimization method of this application, it may further include the following:

[0073] Step S301: Input the video image and the dynamic blur kernel into the physical perception module. The physical perception module includes multiple deconvolution layers and upsampling layers. The video image is deblurred based on the dynamic blur kernel. The deconvolution layer iteratively reconstructs the image features. The upsampling layer improves the resolution of the reconstructed features to generate an initial feature map.

[0074] Step S302: A motion trajectory constraint model is constructed using a bidirectional long short-term memory network. The motion trajectory constraint model includes a forward propagation layer and a backward propagation layer. The initial feature map is input into the motion trajectory constraint model. The forward propagation layer extracts forward temporal features, and the backward propagation layer extracts backward temporal features. Based on the correspondence between the forward temporal features and the backward temporal features, trajectory constraint conditions are constructed. Batch normalization is performed on the initial feature map to generate standardized features with motion trajectory constraints.

[0075] Optionally, this embodiment innovatively designs a feature processing scheme based on physical perception to address issues such as inaccurate feature reconstruction, insufficient temporal constraints, and unreasonable feature standardization in video deblurring. For feature processing evaluation, this embodiment designs a deblurring effect scoring formula: Deblur_Score=α×(Feature_Reconstruction)+β×(Resolution_Enhancement)+γ×(Physical_Consistency), where α, β, and γ are weight coefficients, representing feature reconstruction quality, resolution improvement effect, and physical consistency, respectively. The feature standardization score uses: Normalization_Score=(Temporal_Consistency×Trajectory_Constraint) / (Feature_Distortion+Noise_Level), used to evaluate the quality of standardized features.

[0076] This embodiment first implements video deblurring through a deeply optimized physical perception module. When processing input data, the physical perception module employs a multi-level feature reconstruction strategy. The deconvolution layer is designed based on a deep learning architecture, achieving inverse recovery of blurred features through multi-level feature mapping. Each deconvolution layer is equipped with an attention mechanism, dynamically adjusting reconstruction weights by calculating the feature response map. The iterative reconstruction process adopts a progressive strategy, focusing on recovering the main structural features in the initial stage, while subsequent iterations gradually optimize detailed information. For complex blurred scenes, the system adaptively adjusts the deconvolution parameters by analyzing the characteristics of the dynamic blur kernel. Physical constraints are introduced during the reconstruction process to ensure that the reconstruction results conform to the actual imaging process. The upsampling layer uses a learnable upsampling kernel, improving resolution through multi-scale feature fusion. This optimized deblurring mechanism significantly improves the quality of feature reconstruction.

[0077] This embodiment innovatively implements a motion trajectory constraint mechanism. The bidirectional long short-term memory network is designed with a symmetrical architecture, with the forward propagation layer and backpropagation layer handling the forward and backward dependencies of temporal information, respectively. The forward propagation layer selectively retains and updates temporal features through a gating mechanism, focusing on the continuity and predictability of motion. The backpropagation layer supplements the causal relationships of motion through inverse analysis, enhancing the completeness of trajectory constraints. Temporal consistency loss is introduced during network training to optimize feature extraction by comparing the correspondence between forward and backward features. For complex motion scenarios, the system establishes a multi-scale feature pyramid to achieve feature capture at different time scales. This optimized constraint mechanism provides reliable temporal support for feature standardization.

[0078] This embodiment achieves standardized feature representation through an optimized feature standardization mechanism. Batch normalization considers the statistical properties and physical constraints of the features during the standardization process. The calculation of standardization parameters adopts an adaptive strategy, dynamically adjusting the normalization coefficients according to the feature distribution. To maintain the discriminative power of the features, the system introduces local response normalization, preserving local feature differences while maintaining global consistency. Trajectory constraints during the standardization process are implemented by establishing a feature correspondence graph, ensuring that the standardization results meet the requirements of motion continuity. This optimized standardization mechanism significantly improves the expressive power of the features.

[0079] This embodiment establishes a complete feature processing framework. The generation of standardized features employs a multi-stage optimization strategy: first, feature reconstruction is performed through a physical perception module, and then feature normalization is achieved through motion trajectory constraints. The system establishes a feature quality evaluation mechanism to monitor the processing effect in real time and supports dynamic adjustment of the processing strategy. This optimized processing mechanism provides high-quality feature support for subsequent video processing tasks. This embodiment offers an innovative solution for feature processing; through optimized physical perception and trajectory constraints, the system achieves end-to-end optimization from blurred images to standardized features. This reliable processing mechanism significantly improves feature quality, especially when processing complex motion scenes, where the generated features exhibit stronger physical plausibility and temporal consistency.

[0080] The innovation of this embodiment lies primarily in the feature reconstruction based on physical perception and the constraint representation of motion trajectories. Through multi-level feature processing and bidirectional temporal modeling, the system achieves accurate reconstruction of fuzzy features and effective constraints on motion features, providing reliable feature support for video processing tasks. Especially when processing high-speed motion scenes, this physically constrained feature processing method significantly improves the system's reconstruction quality and temporal consistency.

[0081] In one embodiment of the video stream image optimization method of this application, it may further include the following:

[0082] Step S401: Perform static region segmentation on the standardized features and construct a non-local attention module. The non-local attention module includes a feature transformation sub-network and a similarity calculation sub-network. The feature transformation sub-network maps the standardized features to a high-dimensional feature space. The similarity calculation sub-network calculates the similarity matrix of each pixel in the feature map based on cosine distance and generates a static region feature map.

[0083] Step S402: Input the static region feature map into the texture reconstruction network. The texture reconstruction network includes multiple dense connection modules and skip connection modules. The dense connection modules use multi-scale convolutional kernels to extract multi-level texture information. The skip connection modules fuse texture information at different levels and perform detail enhancement on the fused features based on a self-attention mechanism to generate reconstructed static region texture features.

[0084] Optionally, this embodiment innovatively designs a static feature reconstruction scheme based on non-local attention to address issues such as inaccurate static region segmentation, incomplete texture reconstruction, and insufficient feature representation in video processing. For feature reconstruction evaluation, this embodiment designs a reconstruction quality scoring formula: Reconstruction_Score=α×(Region_Accuracy)+β×(Texture_Fidelity)+γ×(Detail_Preservation), where α, β, and γ are weight coefficients, representing region accuracy, texture fidelity, and detail preservation, respectively. Feature enhancement scoring uses: Enhancement_Score=(Local_Detail×Global_Coherence) / (Information_Loss+Artifact_Level) to evaluate the effect of feature enhancement.

[0085] This embodiment first achieves accurate segmentation of static regions through a deeply optimized nonlocal attention mechanism. When processing standardized features, the feature transformation subnetwork employs a multi-level mapping architecture, mapping features to a high-dimensional space through nonlinear transformation. An adaptive weighting mechanism is introduced during the mapping process, dynamically adjusting transformation parameters based on feature importance. For complex texture regions, the system establishes a multi-scale feature pyramid to achieve hierarchical representation of features. In the high-dimensional feature space, the system constructs a feature similarity map by analyzing the local structure and global distribution of features. The similarity calculation subnetwork uses an improved cosine distance metric, improving the accuracy of similarity calculation by considering the spatial relationship of pixels and feature distribution characteristics. For processing region boundaries, the system introduces a boundary awareness module, improving the accuracy of boundary localization by analyzing feature gradient changes. This optimized attention mechanism significantly improves the accuracy of static region recognition.

[0086] This embodiment innovatively implements a densely connected texture reconstruction mechanism. When processing static region feature maps, the texture reconstruction network employs multiple densely connected modules for feature extraction. Each densely connected module contains multiple parallel convolutional branches, with different branches using convolutional kernels of different scales to capture multi-scale texture features. The convolutional kernel design considers the directionality and periodicity of the texture, enhancing feature extraction in the main texture directions through an adaptive weighting mechanism. The feature fusion process adopts a progressive strategy, first combining features in local regions and then gradually expanding the fusion range to ensure the continuity of texture reconstruction. The skip connection module is designed with a multi-level connection strategy, maintaining the transmission of detailed information by establishing direct pathways between features at different levels. This optimized reconstruction mechanism provides a reliable texture foundation for subsequent feature enhancement.

[0087] This embodiment achieves feature detail enhancement through an optimized self-attention mechanism. When processing the fused features, the system employs a multi-head attention mechanism, with different attention heads responsible for capturing different types of texture features. The calculation of attention weights considers both local correlation and global consistency of features, and a feature dependency graph is established to describe the relationships between different regions. For detail enhancement, the system adaptively adjusts the enhancement intensity by analyzing the frequency distribution of features. Edge preservation constraints are introduced during the enhancement process to ensure that the enhanced features retain edge sharpness. This optimized enhancement mechanism significantly improves the representation quality of texture features.

[0088] This embodiment establishes a complete static feature reconstruction framework. Texture feature generation employs a multi-stage optimization strategy: first, region segmentation is achieved through a non-local attention module; then, texture reconstruction is performed based on a densely connected network; and finally, detail enhancement is achieved through a self-attention mechanism. The system establishes a feature quality evaluation mechanism to monitor the reconstruction effect in real time and supports dynamic parameter adjustment. This optimized reconstruction mechanism provides high-quality static feature representation for subsequent video processing tasks. This embodiment offers an innovative solution for static feature reconstruction. Through optimized attention mechanisms and feature reconstruction, the system achieves end-to-end optimization from standardized features to static texture features. This reliable reconstruction mechanism significantly improves feature expressiveness, especially when processing complex texture scenes, where the reconstructed features exhibit stronger detail representation and spatial consistency.

[0089] The innovation of this embodiment lies primarily in the adaptive processing of non-local attention and the feature reconstruction of dense connections. By constructing a multi-level feature processing network and an optimized attention mechanism, the system achieves accurate segmentation of static regions and high-quality texture reconstruction, providing reliable static feature support for video processing tasks. Especially when processing scenes with rich texture details, this attention-based feature reconstruction method significantly improves the system's expressive power and reconstruction quality.

[0090] In one embodiment of the video stream image optimization method of this application, it may further include the following:

[0091] Step S501: Construct a multi-scale motion encoder, which includes a pyramid pooling module and a dynamic convolution module. The pyramid pooling module performs multi-scale decomposition on the standardized features, and the dynamic convolution module adaptively adjusts the convolution kernel parameters based on the motion amplitude to reconstruct features of the dynamic region and generate motion features of the dynamic region.

[0092] Step S502: Input the static region texture features and the dynamic region motion features into the feature fusion network. The feature fusion network includes a channel recalibration module and a spatial reweighting module. The channel recalibration module adaptively selects feature channels, and the spatial reweighting module adjusts feature weights based on the position sensitivity map. The recalibrated features are cascaded and combined through cross-layer residual connections to generate spatiotemporally decoupled features.

[0093] Optionally, this embodiment innovatively designs a feature processing scheme based on multi-scale coding to address issues such as inaccurate dynamic feature extraction, insufficient feature fusion, and spatiotemporal feature coupling in video processing. For feature processing evaluation, this embodiment designs a coding quality scoring formula: Encoding_Score=α×(Scale_Coverage)+β×(Motion_Adaptability)+γ×(Feature_Discriminability), where α, β, and γ are weight coefficients, representing scale coverage, motion adaptability, and feature discriminability, respectively. The fusion effect score is calculated using: Fusion_Score=(Channel_Selection×Spatial_Weighting) / (Information_Loss+Feature_Interference), used to evaluate the quality of feature fusion.

[0094] This embodiment first achieves accurate extraction of dynamic features through a deep-optimized multi-scale motion encoder. The pyramid pooling module employs a multi-level feature decomposition strategy when processing standardized features. For motion features at different scales, the system aggregates features through an adaptive pooling window to ensure multi-scale representation. An attention mechanism is introduced during pooling, dynamically adjusting pooling weights by calculating the feature response map. For fast-moving targets, the system increases the receptive field of features by adding pooling levels. The feature decomposition process considers spatial correlation, describing the relationships between features at different scales by establishing a feature dependency map. The dynamic convolution module adopts a deformable convolution design, with convolution kernel parameters dynamically adjusted according to the motion amplitude. For complex motion patterns, the system adaptively adjusts the shape and size of the convolution kernel by analyzing the motion gradient field. This optimized encoding mechanism significantly improves the quality of dynamic feature extraction.

[0095] This embodiment innovatively implements a feature fusion mechanism. When processing static and dynamic features, the feature fusion network employs a strategy combining channel recalibration and spatial reweighting. The channel recalibration module learns the dependencies between feature channels to achieve adaptive selection of feature channels. For important feature channels, the system increases their contribution to the fusion process by increasing channel weights. A channel attention mechanism is introduced during calibration, dynamically adjusting channel weights by calculating channel response maps. The spatial reweighting module constructs a spatial weight map of features based on location sensitivity analysis. For target regions, the system enhances feature representation by increasing spatial weights. The reweighting process considers the spatial continuity of features and ensures the rationality of weight distribution through smoothing constraints. This optimized fusion mechanism provides reliable support for feature decoupling.

[0096] This embodiment achieves spatiotemporal decoupling through an optimized feature combination mechanism. The cross-layer residual connection design employs a multi-level feature transfer strategy. For features at different levels, the system maintains the independence and integrity of features through residual learning. A feature selection mechanism is introduced during the connection process, dynamically adjusting the connection strength by calculating feature correlations. The cascade combination adopts a progressive strategy, maintaining independent feature representation in the initial stage and gradually achieving effective feature fusion in subsequent stages. Spatiotemporal consistency is considered during the combination process, ensuring the continuity of decoupled features by establishing feature correspondences. This optimized combination mechanism significantly improves the feature decoupling effect.

[0097] This embodiment establishes a complete feature processing framework. The generation of spatiotemporally decoupled features employs a multi-stage optimization strategy. First, dynamic features are extracted through a multi-scale encoder, and then an effective combination of static and dynamic features is achieved through a feature fusion network. The system establishes a feature quality evaluation mechanism to monitor the processing effect in real time and supports dynamic adjustment of the processing strategy. This optimized processing mechanism provides high-quality feature support for subsequent video processing tasks. This embodiment provides an innovative solution for feature processing. Through optimized feature encoding and fusion mechanisms, the system achieves end-to-end optimization from feature extraction to spatiotemporal decoupling. This reliable processing mechanism significantly improves feature quality, especially when processing complex motion scenes, where decoupled features exhibit stronger discriminative and adaptive characteristics.

[0098] The innovation of this embodiment lies primarily in the dynamic encoding of multi-scale features and the adaptive fusion of features. By constructing a multi-level feature processing network and an optimized fusion mechanism, the system achieves accurate extraction of dynamic features and effective combination of static features, providing reliable feature support for video processing tasks. Especially when processing complex scenes, this decoupled feature processing method significantly improves the system's expressive power and processing accuracy. For example, when processing sports competition videos, the system can accurately separate the dynamic features of athletes from the static features of the field, providing a clearer visual effect.

[0099] In one embodiment of the video stream image optimization method of this application, it may further include the following:

[0100] Step S601: Construct a motion physical constraint model, which includes an acceleration analysis module and a velocity prediction module. The acceleration analysis module calculates the acceleration distribution map of the target object based on the motion trajectory constraints, and the velocity prediction module predicts the motion state at the next moment based on the acceleration distribution map and generates physical constraint parameters.

[0101] Step S602: Input the spatiotemporal decoupling features and the physical constraint parameters into the weight generation network. The weight generation network includes a feature association module and a weight calculation module. The feature association module performs temporal consistency analysis on the spatiotemporal decoupling features. The weight calculation module evaluates and scores the feature consistency based on the physical constraint parameters. The evaluation scores are converted into weight coefficients through normalization processing to generate a dynamic weight matrix.

[0102] Optionally, this embodiment innovatively designs a dynamic weight generation scheme based on physical constraints to address issues such as inaccurate motion prediction, unreasonable weight allocation, and insufficient feature fusion in video processing. For motion prediction evaluation, this embodiment designs a prediction accuracy scoring formula: Prediction_Score=α×(Acceleration_Accuracy)+β×(Velocity_Consistency)+γ×(State_Coherence), where α, β, and γ are weight coefficients, representing acceleration accuracy, velocity consistency, and state coherence, respectively. The weight generation score uses: Weight_Score=(Feature_Correlation×Physical_Constraint) / (Temporal_Error+Spatial_Drift), to evaluate the rationality of the weight allocation.

[0103] This embodiment first achieves accurate prediction of motion state through a deeply optimized physical constraint model. When handling motion trajectory constraints, the acceleration analysis module employs a multi-level feature extraction strategy. For the motion analysis of the target object, the system establishes a kinematic model to describe the object's acceleration variation. The generation of the acceleration distribution map considers multiple physical factors, including inertia, external force influences, and environmental constraints. When handling complex motion, the system decomposes the motion components and analyzes the acceleration characteristics in different directions. For handling abrupt motion, the system introduces a motion state transition model, predicting possible state changes by analyzing historical motion data. The velocity prediction module, based on the acceleration distribution map, uses a dynamic model to predict the target's motion trend. Temporal smoothing constraints are introduced during the prediction process to ensure the continuity of the prediction results. This optimized prediction mechanism significantly improves the accuracy of motion state estimation.

[0104] This embodiment innovatively implements a weight generation mechanism based on physical constraints. The weight generation network employs a two-stream architecture when processing input data. The feature association module constructs a feature dependency graph by analyzing the temporal changes of spatiotemporally decoupled features. Temporal consistency analysis uses a sliding window strategy, evaluating feature stability by comparing feature changes within consecutive time windows. For modeling long-term dependencies, the system establishes a memory mechanism to maintain the transmission of important feature information. The weight calculation module uses a multi-level evaluation strategy based on physical constraint parameters. The evaluation process considers the matching degree between features and physical constraints, quantifying the physical rationality of features through a scoring model. This optimized weight generation mechanism provides reliable weight guidance for feature fusion.

[0105] This embodiment achieves a standardized representation of weights through optimized normalization. The transformation process of weight coefficients employs an adaptive normalization strategy, dynamically adjusting the normalization parameters based on the feature distribution. To maintain the smoothness of weight allocation, the system introduces local consistency constraints to avoid drastic changes in weight distribution. When generating the dynamic weight matrix, the system establishes a spatial correlation graph to describe the weight dependencies between different regions. This optimized normalization mechanism significantly improves the rationality of weight allocation.

[0106] This embodiment establishes a complete weight generation framework. The generation of the dynamic weight matrix employs a multi-stage optimization strategy. First, motion states are predicted using a physical constraint model, and then the weights are accurately allocated based on the weight generation network. The system establishes a weight quality evaluation mechanism to monitor the generation effect in real time and supports dynamic parameter adjustment. This optimized generation mechanism provides high-quality weight support for subsequent feature fusion. This embodiment offers an innovative solution for weight generation. Through optimized physical constraints and weight calculation, the system achieves end-to-end optimization from motion prediction to weight allocation. This reliable generation mechanism significantly improves the quality of the weights, especially when dealing with complex motion scenes, where the generated weights exhibit stronger physical rationality and temporal consistency.

[0107] The innovation of this embodiment lies primarily in the dynamic modeling of physical constraints and the adaptive generation of weights. By constructing a multi-level constraint model and an optimized weight generation network, the system achieves accurate prediction of motion states and reasonable weight allocation, providing reliable fusion support for video processing tasks. Especially when processing highly dynamic scenes, this physical constraint-based weight generation method significantly improves the system's adaptability and accuracy. For example, when processing sports videos, the system can accurately predict changes in athlete acceleration and generate appropriate feature fusion weights accordingly, improving the accuracy of motion trajectory reconstruction.

[0108] In one embodiment of the video stream image optimization method of this application, it may further include the following:

[0109] Step S701: Input the spatiotemporal decoupling features and the dynamic weight matrix into the feature correction network. The feature correction network includes an adaptive fusion module and a physical constraint module. The adaptive fusion module performs weighted combination of the spatiotemporal decoupling features based on the dynamic weight matrix. The physical constraint module performs regular correction on the fused features according to the motion trajectory constraint to generate a corrected feature map.

[0110] Step S702: Construct an optimization and enhancement network, which includes a residual learning module and a spatiotemporal transformation module. The residual learning module performs detail compensation on the correction feature map, and the spatiotemporal transformation module performs spatial alignment and temporal smoothing on the compensated features based on deformable convolution. Optimized video features are generated through multi-layer receptive field fusion and feature recalibration. The optimized video features are then deconvolved to reconstruct the video output with motion blur and ghosting eliminated.

[0111] Optionally, this embodiment innovatively designs a feature optimization scheme based on physical constraints to address problems such as inaccurate feature correction, insufficient detail compensation, and incomplete elimination of visual artifacts in video processing. For feature optimization evaluation, this embodiment designs a correction quality scoring formula: Correction_Score=α×(Fusion_Accuracy)+β×(Physical_Consistency)+γ×(Detail_Enhancement), where α, β, and γ are weighting coefficients, representing fusion accuracy, physical consistency, and the degree of detail enhancement, respectively. The optimization effect score uses: Enhancement_Score=(Residual_Learning×Spatial_Alignment) / (Artifact_Level+Temporal_Discontinuity), to evaluate the quality of feature optimization.

[0112] This embodiment first achieves accurate feature correction through a deeply optimized feature correction network. The adaptive fusion module uses a dynamic weight matrix for feature combination when processing spatiotemporally decoupled features. The weight allocation process considers the spatiotemporal correlation of features, describing the relationships between different regions by establishing a feature dependency graph. For regions with intense motion, the system increases weight values ​​to highlight the expression of motion features. An adaptive threshold mechanism is introduced during the fusion process, dynamically adjusting fusion parameters according to the importance of features. The physical constraint module performs feature correction based on motion trajectory constraints, ensuring that the correction results conform to actual motion characteristics by analyzing the target's motion patterns. The correction process employs an iterative optimization strategy, adjusting features based on physical laws in each iteration. For complex motion scenarios, the system establishes a multi-level constraint model to achieve hierarchical feature correction. This optimized correction mechanism significantly improves the physical rationality of the features.

[0113] This embodiment innovatively implements a feature enhancement mechanism. The optimized enhancement network employs a strategy combining residual learning and spatiotemporal transformation when processing correction features. The residual learning module establishes residual mapping relationships to effectively compensate for detail information. For image edges and textured regions, the system enhances the weights of the residual learning to improve the quality of detail reconstruction. Edge-preserving constraints are introduced during the compensation process to ensure that detail enhancement does not introduce new artifacts. The spatiotemporal transformation module uses a deformable convolution design, dynamically adjusting the shape and position of the convolution kernel to achieve precise feature alignment. For moving targets, the system adaptively adjusts the transformation parameters by analyzing the motion state. This optimized enhancement mechanism provides reliable feature support for video output.

[0114] This embodiment achieves high-quality video output through an optimized feature reconstruction mechanism. Multi-layer receptive field fusion employs a hierarchical feature combination strategy, with receptive fields of different scales responsible for capturing image information at different levels. A spatial attention mechanism is introduced during the fusion process, dynamically adjusting the fusion weights by calculating the feature response map. The feature recalibration process considers spatiotemporal consistency, ensuring the continuity of the reconstruction results by establishing feature correspondences. Deconvolutional reconstruction adopts a progressive strategy, focusing on restoring the main structure in the initial stage and gradually optimizing details in subsequent stages. Motion blur and ghosting removal during the reconstruction process are achieved through physical constraints and temporal smoothing, ensuring the visual quality of the output video. This optimized reconstruction mechanism significantly improves the quality of the video output.

[0115] This embodiment establishes a complete video optimization framework. The generation of video output employs a multi-stage optimization strategy. First, feature correction is performed through a feature correction network, followed by feature enhancement and reconstruction through an optimization and enhancement network. The system establishes a video quality evaluation mechanism to monitor the processing effect in real time and supports dynamic adjustment of the optimization strategy. This optimized processing mechanism provides reliable technical support for video quality improvement. This embodiment offers an innovative solution for video optimization. Through optimized feature correction and enhancement mechanisms, the system achieves end-to-end optimization from feature processing to video output. This reliable processing mechanism significantly improves video quality, especially when processing complex motion scenes, resulting in output videos with clearer details and fewer visual artifacts.

[0116] The innovation of this embodiment lies primarily in the feature correction based on physical constraints and the multi-level feature enhancement. By constructing an optimized feature processing network and a robust reconstruction mechanism, the system achieves accurate optimization and high-quality reconstruction of video features, providing reliable quality assurance for video processing tasks. Particularly when processing high-speed motion scenes, this physically constrained optimization method significantly improves the system's processing capabilities and output quality. For example, when processing moving target tracking videos, the system can effectively eliminate motion-induced blurring and ghosting, providing a clearer display of the target trajectory.

[0117] To effectively address the shortcomings of traditional technologies in areas such as blur processing, feature decoupling, and image optimization, this application provides an embodiment of a video stream image optimization apparatus for implementing all or part of the aforementioned video stream image optimization method. See [link to relevant documentation]. Figure 2 The video stream image optimization device specifically includes the following components:

[0118] The preprocessing module 10 is used to generate a dynamic blur kernel based on the motion speed of the video image and the target object and the optical flow parameters. It uses multiple convolutional units to perform optical flow modeling and occlusion detection on the video image, and inputs the video image and the blur kernel into the physical perception module for data preprocessing to generate standardized features with motion trajectory constraints.

[0119] The spatiotemporal decoupling module 20 is used to input the standardized features into the spatiotemporal decoupling network. The spatiotemporal decoupling network includes a content branch and a motion branch. The content branch reconstructs the texture features of static regions, and the motion branch reconstructs the motion features of dynamic regions. The reconstruction results of the content branch and the motion branch are fused based on the cross-layer feature connection mechanism to generate spatiotemporal decoupling features.

[0120] The image optimization module 30 is used to input the spatiotemporal decoupling features into the physical constraint corrector. The physical constraint corrector generates a dynamic weight matrix based on the motion trajectory constraints, performs adaptive fusion and physical law correction on the spatiotemporal decoupling features, and optimizes the corrected features through residual connections and spatiotemporal transformation networks to generate a video output that eliminates motion blur and ghosting.

[0121] As described above, the video stream image optimization device provided in this application embodiment can ensure data processing security through an innovative design of a dynamic blur kernel generation mechanism, spatiotemporal decoupling, and physical constraints. A content-motion dual-branch network is constructed, combined with residual connections and spatiotemporal transformation, to establish a reliable feature reconstruction system. A physical constraint corrector is introduced, and through adaptive fusion and regular correction, high-quality video output is provided while protecting user privacy. This method effectively solves the shortcomings of traditional technologies in blur processing, feature decoupling, and image optimization.

[0122] From a hardware perspective, in order to effectively address the shortcomings of traditional technologies in areas such as blur processing, feature decoupling, and image optimization, this application provides an embodiment of an electronic device for implementing all or part of the video stream image optimization method. The electronic device specifically includes the following components:

[0123] The system comprises a processor, memory, a communications interface, and a bus; wherein the processor, memory, and communications interface communicate with each other via the bus; the communications interface is used to realize information transmission between the video stream image optimization device and core business systems, user terminals, and related databases and other related devices; the logic controller can be a desktop computer, tablet computer, or mobile terminal, etc., and this embodiment is not limited to these. In this embodiment, the logic controller can be implemented with reference to the embodiments of the video stream image optimization method and the embodiments of the video stream image optimization device in the embodiments, the contents of which are incorporated herein, and repeated details will not be described again.

[0124] It is understood that the user terminal may include smartphones, tablet computers, network set-top boxes, portable computers, desktop computers, personal digital assistants (PDAs), in-vehicle devices, smart wearable devices, etc. Among these, the smart wearable devices may include smart glasses, smartwatches, smart bracelets, etc.

[0125] In practical applications, some parts of the video stream image optimization method can be executed on the electronic device side as described above, or all operations can be completed in the client device. The choice can be made based on the processing power of the client device and the limitations of the user's usage scenario. This application does not impose any limitations on this. If all operations are completed in the client device, the client device may further include a processor.

[0126] The aforementioned client device may have a communication module (i.e., a communication unit) that can communicate with a remote server to achieve data transmission. The server may include a server on the task scheduling center side; in other implementation scenarios, it may also include a server on an intermediate platform, such as a server on a third-party server platform that has a communication link with the task scheduling center server. The server may include a single computer device, a server cluster consisting of multiple servers, or a distributed server structure.

[0127] Figure 3 This is a schematic block diagram illustrating the system configuration of the electronic device 9600 according to an embodiment of this application. Figure 3 As shown, the electronic device 9600 may include a central processing unit 9100 and a memory 9140; the memory 9140 is coupled to the central processing unit 9100. It is worth noting that... Figure 3 This is an example; other types of structures can also be used to supplement or replace this structure to achieve telecommunications functions or other functions.

[0128] In one embodiment, the video stream image optimization method functionality can be integrated into the central processing unit 9100. The central processing unit 9100 can be configured to perform the following control:

[0129] Step S101: Generate a dynamic blur kernel based on the motion speed of the video image and the target object and the optical flow parameters. Use multiple convolutional units to perform optical flow modeling and occlusion detection on the video image. Input the video image and the blur kernel into the physical perception module for data preprocessing to generate standardized features with motion trajectory constraints.

[0130] Step S102: Input the standardized features into the spatiotemporal decoupling network. The spatiotemporal decoupling network includes a content branch and a motion branch. The content branch reconstructs the texture features of static regions, and the motion branch reconstructs the motion features of dynamic regions. Based on the cross-layer feature connection mechanism, the reconstruction results of the content branch and the motion branch are fused to generate spatiotemporal decoupling features.

[0131] Step S103: Input the spatiotemporal decoupling features into the physical constraint corrector. The physical constraint corrector generates a dynamic weight matrix based on the motion trajectory constraints, performs adaptive fusion and physical law correction on the spatiotemporal decoupling features, and optimizes the corrected features through residual connections and spatiotemporal transformation networks to generate video output with eliminated motion blur and ghosting.

[0132] As described above, the electronic device provided in this application embodiment, through an innovative design of a dynamic blur kernel generation mechanism, ensures data processing security through spatiotemporal decoupling and physical constraints. A content-motion dual-branch network is constructed, combining residual connections and spatiotemporal transformation to establish a reliable feature reconstruction system. A physical constraint corrector is introduced, providing high-quality video output while protecting user privacy through adaptive fusion and regular correction. This method effectively addresses the shortcomings of traditional techniques in blur processing, feature decoupling, and image optimization.

[0133] In another embodiment, the video stream image optimization device can be configured separately from the central processing unit 9100. For example, the video stream image optimization device can be configured as a chip connected to the central processing unit 9100, and the video stream image optimization method function can be implemented through the control of the central processing unit.

[0134] like Figure 3 As shown, the electronic device 9600 may further include: a communication module 9110, an input unit 9120, an audio processor 9130, a display 9160, and a power supply 9170. It is worth noting that the electronic device 9600 does not necessarily need to include these components. Figure 3 All components shown; in addition, the electronic device 9600 may also include Figure 3For components not shown, please refer to existing technologies.

[0135] like Figure 3 As shown, the central processing unit 9100, sometimes also referred to as a controller or operating control, may include a microprocessor or other processor device and / or logic device, which receives inputs and controls the operation of various components of the electronic device 9600.

[0136] The memory 9140 may be, for example, one or more of a cache, flash memory, hard drive, removable media, volatile memory, non-volatile memory, or other suitable devices. It may store the aforementioned failure-related information, and also store a program for executing that information. The central processing unit 9100 may execute the program stored in the memory 9140 to perform information storage or processing, etc.

[0137] Input unit 9120 provides input to central processing unit 9100. Input unit 9120 may be, for example, a keypad or touch input device. Power supply 9170 provides power to electronic device 9600. Display 9160 displays images and text. Display may be, for example, an LCD display, but is not limited thereto.

[0138] The memory 9140 can be a solid-state memory, such as a read-only memory (ROM), random access memory (RAM), a SIM card, etc. It can also be a memory that retains information even when power is off, can be selectively erased, and contains more data; examples of this type of memory are sometimes referred to as EPROMs. The memory 9140 can also be some other type of device. The memory 9140 includes a buffer memory 9141 (sometimes referred to as a buffer). The memory 9140 may include an application / function storage unit 9142 for storing application programs and function programs or processes for executing the operation of the electronic device 9600 via the central processing unit 9100.

[0139] The memory 9140 may also include a data storage unit 9143 for storing data, such as contacts, digital data, pictures, sounds, and / or any other data used by the electronic device. The driver storage unit 9144 of the memory 9140 may include various drivers for the electronic device's communication functions and / or for performing other functions of the electronic device (such as messaging applications, address book applications, etc.).

[0140] The communication module 9110 is a transmitter / receiver that sends and receives signals via the antenna 9111. The communication module 9110 (transmitter / receiver) is coupled to the central processing unit 9100 to provide input signals and receive output signals, which is the same as in a conventional mobile communication terminal.

[0141] Based on different communication technologies, multiple communication modules 9110 can be configured in the same electronic device, such as cellular network modules, Bluetooth modules, and / or wireless LAN modules. The communication module 9110 (transmitter / receiver) is also coupled to a speaker 9131 and a microphone 9132 via an audio processor 9130 to provide audio output via the speaker 9131 and receive audio input from the microphone 9132, thereby realizing typical telecommunications functions. The audio processor 9130 may include any suitable buffer, decoder, amplifier, etc. Additionally, the audio processor 9130 is coupled to a central processing unit 9100, enabling on-device recording via the microphone 9132 and on-device playback of stored audio via the speaker 9131.

[0142] Embodiments of this application also provide a computer-readable storage medium capable of implementing all steps of the video stream image optimization method with a server or client as the execution subject in the above embodiments. The computer-readable storage medium stores a computer program that, when executed by a processor, implements all steps of the video stream image optimization method with a server or client as the execution subject in the above embodiments. For example, when the processor executes the computer program, it implements the following steps:

[0143] Step S101: Generate a dynamic blur kernel based on the motion speed of the video image and the target object and the optical flow parameters. Use multiple convolutional units to perform optical flow modeling and occlusion detection on the video image. Input the video image and the blur kernel into the physical perception module for data preprocessing to generate standardized features with motion trajectory constraints.

[0144] Step S102: Input the standardized features into the spatiotemporal decoupling network. The spatiotemporal decoupling network includes a content branch and a motion branch. The content branch reconstructs the texture features of static regions, and the motion branch reconstructs the motion features of dynamic regions. Based on the cross-layer feature connection mechanism, the reconstruction results of the content branch and the motion branch are fused to generate spatiotemporal decoupling features.

[0145] Step S103: Input the spatiotemporal decoupling features into the physical constraint corrector. The physical constraint corrector generates a dynamic weight matrix based on the motion trajectory constraints, performs adaptive fusion and physical law correction on the spatiotemporal decoupling features, and optimizes the corrected features through residual connections and spatiotemporal transformation networks to generate video output with eliminated motion blur and ghosting.

[0146] As described above, the computer-readable storage medium provided in this application embodiment, through an innovative design of a dynamic blur kernel generation mechanism, ensures data processing security through spatiotemporal decoupling and physical constraints. A content-motion dual-branch network is constructed, combining residual connections and spatiotemporal transformation to establish a reliable feature reconstruction system. A physical constraint corrector is introduced, providing high-quality video output while protecting user privacy through adaptive fusion and regular correction. This method effectively addresses the shortcomings of traditional techniques in blur processing, feature decoupling, and image optimization.

[0147] Embodiments of this application also provide a computer program product capable of implementing all steps of the video stream image optimization method with the execution subject being a server or client in the above embodiments. When executed by a processor, this computer program / instruction implements the steps of the video stream image optimization method. For example, the computer program / instruction implements the following steps:

[0148] Step S101: Generate a dynamic blur kernel based on the motion speed of the video image and the target object and the optical flow parameters. Use multiple convolutional units to perform optical flow modeling and occlusion detection on the video image. Input the video image and the blur kernel into the physical perception module for data preprocessing to generate standardized features with motion trajectory constraints.

[0149] Step S102: Input the standardized features into the spatiotemporal decoupling network. The spatiotemporal decoupling network includes a content branch and a motion branch. The content branch reconstructs the texture features of static regions, and the motion branch reconstructs the motion features of dynamic regions. Based on the cross-layer feature connection mechanism, the reconstruction results of the content branch and the motion branch are fused to generate spatiotemporal decoupling features.

[0150] Step S103: Input the spatiotemporal decoupling features into the physical constraint corrector. The physical constraint corrector generates a dynamic weight matrix based on the motion trajectory constraints, performs adaptive fusion and physical law correction on the spatiotemporal decoupling features, and optimizes the corrected features through residual connections and spatiotemporal transformation networks to generate video output with eliminated motion blur and ghosting.

[0151] As described above, the computer program product provided in this application, through an innovative design of a dynamic fuzzy kernel generation mechanism, ensures data processing security through spatiotemporal decoupling and physical constraints. A content-motion dual-branch network is constructed, combining residual connections and spatiotemporal transformation to establish a reliable feature reconstruction system. A physical constraint corrector is introduced, providing high-quality video output while protecting user privacy through adaptive fusion and regular correction. This method effectively addresses the shortcomings of traditional techniques in fuzzing processing, feature decoupling, and image optimization.

[0152] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0153] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (devices), and computer program products according to embodiments of the invention. It will be understood that each block of 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, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0154] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0155] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0156] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A video stream image optimization method, characterized in that, The method includes: A dynamic blur kernel is generated based on the motion speed of the video image and the target object, as well as optical flow parameters. Multiple convolutional units are used to perform optical flow modeling and occlusion detection on the video image. This includes: performing multi-scale feature matching on consecutive video frames based on the pyramid optical flow algorithm to construct the motion trajectory matrix of the target object; inputting the motion trajectory matrix into a multi-layer convolutional neural network for feature extraction, wherein the multi-layer convolutional neural network includes a spatial attention module and a channel attention module; calculating dynamic blur kernel parameters based on the spatiotemporal correlation of the motion trajectory matrix; constructing an optical flow model using multiple dilated convolutional units, wherein the dilated convolutional units include residual connection layers and normalization layers; performing frame-by-frame optical flow estimation on the video image sequence; generating an occlusion detection map based on the consistency analysis of the optical flow vector field and the motion trajectory matrix; and weighting and combining the occlusion detection map with the dynamic blur kernel parameters to obtain a dynamic blur kernel that considers occlusion relationships. The video image and the blur kernel are input into a physical perception module for data preprocessing to generate standardized features with motion trajectory constraints. This includes: inputting the video image and the dynamic blur kernel into the physical perception module, which contains multiple deconvolution layers and upsampling layers; performing reverse deblurring on the video image based on the dynamic blur kernel; iteratively reconstructing image features using the deconvolution layers; and increasing the resolution of the reconstructed features using the upsampling layers to generate an initial feature map; constructing a motion trajectory constraint model using a bidirectional long short-term memory network, which contains forward propagation layers and backward propagation layers; inputting the initial feature map into the motion trajectory constraint model; extracting forward temporal features using the forward propagation layer; extracting backward temporal features using the backward propagation layer; constructing trajectory constraints based on the correspondence between the forward and backward temporal features; and batch normalizing the initial feature map to generate standardized features with motion trajectory constraints. The standardized features are input into a spatiotemporal decoupling network, which includes a content branch and a motion branch. The content branch reconstructs the texture features of static regions, and the motion branch reconstructs the motion features of dynamic regions. The reconstruction results of the content branch and the motion branch are fused based on a cross-layer feature connection mechanism to generate spatiotemporal decoupling features. The spatiotemporal decoupling features are input into a physical constraint corrector. The physical constraint corrector generates a dynamic weight matrix based on the motion trajectory constraints, performs adaptive fusion and physical law correction on the spatiotemporal decoupling features, and optimizes the corrected features through residual connections and a spatiotemporal transformation network to generate a video output that eliminates motion blur and ghosting.

2. The video stream image optimization method according to claim 1, characterized in that, The standardized features are input into a spatiotemporal decoupling network, which includes a content branch and a motion branch. The content branch reconstructs the texture features of static regions, including: Static region segmentation is performed on the standardized features to construct a non-local attention module. The non-local attention module includes a feature transformation sub-network and a similarity calculation sub-network. The feature transformation sub-network maps the standardized features to a high-dimensional feature space, and the similarity calculation sub-network calculates the similarity matrix of each pixel in the feature map based on cosine distance to generate a static region feature map. The static region feature map is input into a texture reconstruction network, which includes multiple dense connection modules and skip connection modules. The dense connection modules use multi-scale convolutional kernels to extract multi-level texture information, and the skip connection modules fuse texture information at different levels. Based on a self-attention mechanism, the fused features are enhanced in detail to generate reconstructed static region texture features.

3. The video stream image optimization method according to claim 1, characterized in that, The motion branch reconstructs the motion features of the dynamic region, and fuses the reconstruction results of the content branch and the motion branch based on a cross-layer feature connection mechanism to generate spatiotemporal decoupled features, including: A multi-scale motion encoder is constructed, which includes a pyramid pooling module and a dynamic convolution module. The pyramid pooling module performs multi-scale decomposition on the standardized features, and the dynamic convolution module adaptively adjusts the convolution kernel parameters based on the motion amplitude to reconstruct features of the dynamic region and generate dynamic region motion features. The static region texture features and the dynamic region motion features are input into a feature fusion network. The feature fusion network includes a channel recalibration module and a spatial reweighting module. The channel recalibration module adaptively selects feature channels, and the spatial reweighting module adjusts feature weights based on a position sensitivity map. The recalibrated features are cascaded and combined through cross-layer residual connections to generate spatiotemporally decoupled features.

4. The video stream image optimization method according to claim 1, characterized in that, The step of inputting the spatiotemporal decoupling features into a physical constraint corrector, wherein the physical constraint corrector generates a dynamic weight matrix based on the motion trajectory constraints, includes: A motion physical constraint model is constructed, which includes an acceleration analysis module and a velocity prediction module. The acceleration analysis module calculates the acceleration distribution map of the target object based on the motion trajectory constraints, and the velocity prediction module predicts the motion state at the next moment based on the acceleration distribution map, thereby generating physical constraint parameters. The spatiotemporal decoupling features and the physical constraint parameters are input into a weight generation network. The weight generation network includes a feature association module and a weight calculation module. The feature association module performs temporal consistency analysis on the spatiotemporal decoupling features. The weight calculation module evaluates and scores the feature consistency based on the physical constraint parameters. The evaluation scores are converted into weight coefficients through normalization processing to generate a dynamic weight matrix.

5. The video stream image optimization method according to claim 1, characterized in that, The adaptive fusion and physical law correction of the spatiotemporal decoupling features, followed by optimization of the corrected features through residual connections and a spatiotemporal transformation network, to generate video output with eliminated motion blur and ghosting, includes: The spatiotemporal decoupling features and the dynamic weight matrix are input into a feature correction network. The feature correction network includes an adaptive fusion module and a physical constraint module. The adaptive fusion module performs weighted combination of the spatiotemporal decoupling features based on the dynamic weight matrix. The physical constraint module performs regular correction on the fused features according to motion trajectory constraints to generate a corrected feature map. An optimization and enhancement network is constructed, which includes a residual learning module and a spatiotemporal transformation module. The residual learning module performs detail compensation on the correction feature map, and the spatiotemporal transformation module performs spatial alignment and temporal smoothing on the compensated features based on deformable convolution. Optimized video features are generated through multi-layer receptive field fusion and feature recalibration. The optimized video features are then deconvolved to reconstruct the video output with motion blur and ghosting eliminated.

6. A video stream image optimization device, characterized in that, The device includes: The preprocessing module generates a dynamic blur kernel based on the motion speed of the video image and the target object, as well as optical flow parameters. It employs multiple convolutional units to perform optical flow modeling and occlusion detection on the video image. This includes: performing multi-scale feature matching on consecutive video frames based on the pyramid optical flow algorithm to construct the motion trajectory matrix of the target object; inputting the motion trajectory matrix into a multi-layer convolutional neural network for feature extraction, where the multi-layer convolutional neural network includes a spatial attention module and a channel attention module; calculating dynamic blur kernel parameters based on the spatiotemporal correlation of the motion trajectory matrix; constructing an optical flow model using multiple dilated convolutional units, where each dilated convolutional unit includes a residual connection layer and a normalization layer; performing frame-by-frame optical flow estimation on the video image sequence; generating an occlusion detection map based on the consistency analysis of the optical flow vector field and the motion trajectory matrix; weighting and combining the occlusion detection map with the dynamic blur kernel parameters to obtain a dynamic blur kernel considering occlusion relationships; and then combining the video image and the target object's motion trajectory matrix into a dynamic blur kernel. The process involves inputting a blur kernel into a physical perception module for data preprocessing to generate standardized features with motion trajectory constraints. This includes: inputting the video image and the dynamic blur kernel into the physical perception module, which contains multiple deconvolutional layers and upsampling layers; performing reverse deblurring on the video image based on the dynamic blur kernel; iteratively reconstructing image features using the deconvolutional layers; and enhancing the resolution of the reconstructed features using the upsampling layers to generate an initial feature map. A bidirectional long short-term memory network is then used to construct a motion trajectory constraint model, which includes forward propagation layers and backward propagation layers. The initial feature map is input into the motion trajectory constraint model; the forward propagation layer extracts forward temporal features; the backward propagation layer extracts backward temporal features; trajectory constraints are constructed based on the correspondence between the forward and backward temporal features; and batch normalization is performed on the initial feature map to generate standardized features with motion trajectory constraints. The spatiotemporal decoupling module is used to input the standardized features into the spatiotemporal decoupling network. The spatiotemporal decoupling network includes a content branch and a motion branch. The content branch reconstructs the texture features of static regions, and the motion branch reconstructs the motion features of dynamic regions. The reconstruction results of the content branch and the motion branch are fused based on the cross-layer feature connection mechanism to generate spatiotemporal decoupling features. The image optimization module is used to input the spatiotemporal decoupling features into the physical constraint corrector. The physical constraint corrector generates a dynamic weight matrix based on the motion trajectory constraints, performs adaptive fusion and physical law correction on the spatiotemporal decoupling features, and optimizes the corrected features through residual connections and spatiotemporal transformation networks to generate video output with eliminated motion blur and ghosting.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the video stream image optimization method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the video stream image optimization method according to any one of claims 1 to 5.

Images