A reference relationship construction and consistency control method in visual asset generation

Abstract

The present application belongs to the technical field of computer vision, and discloses a reference relationship construction and consistency control method in visual asset generation, which is executed by an embedded visual processor: geometric features, texture features and illumination features of dynamic visual assets are collected and normalized pretreatment is completed to obtain standardized time sequence feature data; a time sequence deviation quantization factor is calculated in combination with a single frame time sequence interval; a cross-domain mapping distortion coefficient is calculated by fusing a time sequence deviation, a time sequence interval and a double-domain feature component; a consistency control total weight is calculated by linking a time sequence parameter, a distortion coefficient and a cross-scale feature total amount; and a visual asset reference relationship is reconstructed based on the total weight and rendering output is completed. The present application introduces a physical time sequence dimension to quantify interframe deviation, realizes linkage correction of time sequence error and spatial distortion, solves the technical problem that traditional static feature matching cannot adapt to dynamic rendering scenes, and effectively guarantees dynamic visual asset feature consistency and rendering stability.

Description

Technical Field

[0001] This invention belongs to the field of computer vision technology and discloses a method for constructing reference relationships and controlling consistency in visual asset generation. Background Technology

[0002] In dynamic real-time visual rendering and digital visual asset generation scenarios, the construction of cross-frame reference relationships and feature consistency control of visual assets are core aspects of ensuring stable rendering. Existing visual asset generation technologies generally employ a purely numerical static feature matching method to construct inter-frame reference relationships. This method completes reference matching solely through simple comparison of feature values, without introducing a standard physical temporal dimension to quantify and correct inter-frame feature lag bias. This type of static matching method completely deviates from the physical temporal characteristics of dynamic rendering, failing to quantify the temporal lag error of inter-frame features in high-speed dynamic scenes and unable to adapt to the dynamic changes of continuous video frames. Furthermore, existing technologies cannot achieve the correlation and coupling between temporal physical bias and cross-domain mapping distortion in the feature domain. Mapping distortion generated during feature domain transformation in dynamic rendering cannot be synchronously compensated for in conjunction with temporal errors. Ultimately, this leads to broken visual asset reference relationships, failure of texture and lighting feature matching in high-speed dynamic scenes, low rendering fidelity, and poor temporal continuity, making it difficult to meet the requirements of high-precision dynamic visual rendering. Based on the above problems, there is an urgent need for a dynamic consistency control technology that combines physical time-series quantization and distortion coupling correction. Summary of the Invention

[0003] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method for constructing reference relationships and controlling consistency in visual asset generation, wherein the following steps are executed by an embedded vision processor: S1: Collect the geometric features, texture features, and lighting features of dynamic visual assets, and perform normalization preprocessing on the geometric features, texture features, and lighting features to obtain standardized temporal feature data; S2: Extract visual assets to generate single-frame temporal intervals, and calculate the temporal deviation quantification factor by combining standardized temporal feature data; S3: Calculate the cross-domain mapping distortion coefficient based on the temporal deviation quantization factor, the temporal interval of the visual asset generation single frame, the original feature domain reference feature components, and the target rendering domain reference feature components. S4: Calculate the total weight of reference relationship consistency control based on the temporal deviation quantization factor, the temporal interval of single frame generation of visual assets, the cross-domain mapping distortion coefficient, and the total amount of cross-scale normalized features of visual assets. S5: Based on the consistency of reference relationships, control the total weight and reconstruct the reference relationships of visual assets; S6: Based on the reference relationships of the reconstructed visual assets, perform visual asset rendering output operations.

[0004] Preferably, the geometric features, texture features, and lighting features are all visual feature data corresponding to consecutive video frames in a dynamic real-time rendering scene, and the geometric features, texture features, and lighting features are acquired synchronously.

[0005] Further preferred, the normalization preprocessing includes amplitude normalization and scale normalization. Amplitude normalization eliminates the interference of differences in visual feature amplitudes, while scale normalization unifies the numerical scale standard of visual features.

[0006] Further preferred, the input data used to calculate the temporal deviation quantification factor includes the normalized feature value at time t, the normalized feature value at time t minus the temporal interval of the visual asset generation single frame, the total number of visual reference feature paths, and the feature temporal weight coefficient.

[0007] Further preferably, the temporal deviation quantization factor is a physical temporal quantization parameter obtained based on the weighted summation of inter-frame feature differences, the normalization of the total number of features in the current frame, and the coupling of the temporal interval of a single frame. The temporal deviation quantization factor is used to characterize the degree of temporal lag deviation of visual reference features between consecutive video frames.

[0008] Further preferably, the cross-domain mapping distortion coefficient is a comprehensive distortion parameter obtained by fusing the temporal dimension distortion error and the feature domain cross-domain mapping error. The cross-domain mapping distortion coefficient is used to characterize the overall distortion degree during the conversion process from the original feature domain to the target rendering domain.

[0009] In a further preferred embodiment, the total weight of reference relationship consistency control is a dynamic control parameter obtained based on distortion attenuation control logic, temporal deviation coupling logic, and cross-scale feature nonlinear correction logic. The total weight of reference relationship consistency control is used to adaptively control the association strength of reference relationships in visual assets.

[0010] A further preferred method for reconstructing the reference relationships of visual assets includes controlling the total weight based on the consistency of the reference relationships and adjusting the correlation strength of geometric features, texture features, and illumination features of the visual assets.

[0011] In a further optimized manner, after completing the visual asset rendering output operation, the entire process of feature acquisition preprocessing, parameter calculation, reference relationship reconstruction, and rendering output is executed in a loop to achieve iterative processing of visual assets in consecutive frames.

[0012] In a further preferred embodiment, before performing the calculation of the temporal deviation quantization factor, the total number of visual reference feature paths and the total number of reference feature components are predetermined, and the initial configuration of the consistency basic control coefficient, distortion attenuation control factor, temporal deviation coupling factor, and scale feature correlation coefficient is completed.

[0013] Technical effects: This invention addresses the core problems of existing technologies, such as the lack of a physical temporal dimension quantization mechanism and the inability of static matching to adapt to dynamic rendering. It introduces a physical parameter representing the single-frame temporal interval to quantify inter-frame temporal deviations, achieving precise temporal dimension correction. Simultaneously, it integrates temporal distortion and feature domain mapping distortion to construct a comprehensive distortion coefficient, and combines this with cross-scale feature nonlinear control logic to generate dynamic consistency weights. This invention breaks through the technical limitations of traditional purely numerical static matching, establishing a dimensionally self-consistent multi-dimensional linkage control system. It effectively avoids the problem of broken reference relationships in visual assets under dynamic scenes, stably ensuring the cross-frame consistency of geometric, texture, and lighting features of visual assets, and adapting to various high-speed dynamic real-time rendering scenarios. Attached Figure Description

[0014] Figure 1 Flowchart of the reference relationship construction and consistency control method in the visual asset generation of this application; Figure 2 This is a sequence diagram of the reference relationship construction and consistency control method in the visual asset generation of this application. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0016] The technical problem with existing technologies is that they use static feature numerical matching to build reference relationships during the generation of dynamic visual assets, without introducing physical temporal parameters to quantify inter-frame lag bias.

[0017] Based on this, please refer to Figures 1-2 This embodiment provides a dynamic visual asset adaptive generation method, in which an embedded vision processor executes a defined closed-loop process. The embedded vision processor, as the execution entity, performs full-process processing of visual assets in dynamic real-time rendering scenarios. This method fully covers the entire closed-loop process of dynamic visual rendering temporal deviation quantization, distortion correction, weight adjustment, relation reconstruction, and rendering output, constructing a dynamic physical quantization control system that differs from traditional static matching.

[0018] The first step of this method involves collecting core feature data of dynamic visual assets and performing normalization preprocessing to form standardized temporal feature data. The geometric feature data, texture feature data, and lighting feature data collected in this step are all visual feature data corresponding to consecutive video frames in a dynamic real-time rendering scene. Geometric feature data reflects the morphological information of the visual asset in three-dimensional space, specifically including spatial geometric attributes such as vertex coordinates, patch structure, and contour curvature. Texture feature data reflects the texture distribution pattern of the visual asset surface, specifically including surface attributes such as texture density, texture direction gradient, and texture periodicity. Lighting feature data reflects the lighting interaction state of the visual asset in a dynamic scene, specifically including optical attributes such as brightness distribution, specular intensity, and shadow area proportion. The collection of these three types of core feature data is completed synchronously; that is, geometric features, texture features, and lighting features within the same video frame are extracted and recorded synchronously at the same time reference. This synchronous collection constraint ensures that the time reference of the three types of feature data is completely unified at the same time node, eliminating the problem of feature data temporal misalignment caused by differences in collection time points. The acquisition target is limited to continuous video frames because the core characteristic of dynamic rendering scenes is the temporal continuity between frames. Only the feature data of continuous video frames can truly reflect the dynamic changes of visual assets. Static single-frame data cannot reflect the temporal evolution of inter-frame lag and dynamic distortion. The dual limitation of synchronous acquisition and continuous frame acquisition ensures the temporal validity and benchmark consistency of the input data, providing reliable data support for subsequent temporal deviation quantification and distortion correction.

[0019] After feature data acquisition, normalization preprocessing is performed on the three types of core feature data. Normalization preprocessing specifically includes amplitude normalization and scale normalization. Amplitude normalization addresses the amplitude differences between feature data from different frames and dimensions. These amplitude differences refer to non-physical fluctuations in the numerical values ​​of the same type of feature at different time points or under different scene conditions. For example, the overall shift in illumination feature data due to changes in the overall brightness of the scene, and the density scaling of texture feature data due to adjustments in rendering precision. These amplitude fluctuations do not originate from changes in the physical state of the visual asset itself. If directly used for subsequent difference calculations and weighted operations, they will introduce invalid interference and compromise the accuracy of the quantization results. Amplitude normalization eliminates the aforementioned redundant amplitude interference by mapping the numerical range of each feature data to a unified preset interval, retaining only the effective information reflecting the dynamic physical changes of the visual asset. Scale normalization addresses the problem of inconsistent numerical scales between different categories of features. The numerical range of geometric feature data may differ by orders of magnitude from that of illumination feature data, and the numerical scale of texture feature data differs from both. The raw values ​​of these three types of feature data cannot be compared and weighted under a unified numerical standard. Scale normalization transforms feature data from different categories to the same numerical scale standard, providing the foundation for lateral computation of multi-dimensional features. Two types of normalization preprocessing work together to remove invalid interference information from the original feature data, retaining only the dynamic physical change patterns of the visual assets, forming standardized temporal feature data. This standardized temporal feature data serves as the compliant data foundation for all subsequent quantization operations, meeting the requirements of consistency and purity of input data for multi-dimensional quantization operations.

[0020] The second step of this method involves extracting the physical parameter of the single-frame temporal interval from hardware rendering and combining it with standardized temporal feature data to quantify the temporal deviation. The core of this step is to transform the traditional purely numerical feature differences into deviation parameters with clear physical temporal meaning. Before implementation, two preparatory operations need to be completed. The first preparatory operation is to determine the total number of visual reference feature paths. and the total number of reference feature components Total number of visual reference features This refers to the total number of standardized feature vectors involved in the time-series bias quantization calculation. Each feature vector corresponds to an independent feature dimension. For example, vertex curvature in geometric features can be considered as an independent feature, and directional gradient in texture features can be considered as another independent feature. Features of different dimensions together constitute... Road visual reference features. Total number of reference feature components. This refers to the total number of feature components contained in both the original feature domain and the target rendering domain during subsequent distortion coefficient calculations. Each feature component corresponds to a numerical representation on a specific feature dimension. The values ​​are determined based on the feature dimension partitioning scheme in the actual application. Pre-determining the above two feature quantity parameters is to unify the feature dimension benchmark for calculation across frames, avoiding quantization deviation fluctuations caused by inconsistencies in the number of feature paths or feature components between different frames, thus ensuring the stability of the calculation dimension from the source. The second pre-operation is to complete the standardization initialization of the control coefficients, specifically including the feature temporal weight coefficients. Temporal distortion coupling coefficient Distortion attenuation regulation factor Consistent basic control coefficient Timing deviation coupling factor and scale feature correlation coefficient Feature time-series weighting coefficients for A dimensional vector, where each dimension corresponds to a visual reference feature, is used to distinguish the differences in sensitivity of different feature dimensions to temporal lag bias. For example, dynamically changing illumination features have higher temporal sensitivity, corresponding to... The value can be relatively increased; for geometric features with relatively gradual changes, its time sensitivity is low, and the corresponding... The value can be relatively reduced. All The sum of the values ​​must satisfy a preset normalization constraint to ensure a consistent standard for weighted operations. (Temporal distortion coupling coefficient) Used to control the intensity of the transformation of timing bias into distortion in subsequent distortion coefficient calculations. Distortion attenuation control factor. Used to control the impact of distortion on the decay rate of consistency weights. Consistency base adjustment coefficient. A baseline level used to set consistency weights. Timing deviation coupling factor. Used to adjust the strength of the influence of time series deviations on the weighted denominator correction term. Scale characteristic correlation coefficient. This is used to regulate the influence of cross-scale feature differences on the weight denominator correction term. The values ​​of all the above regulation coefficients are determined during the system initialization phase based on the rendering scene type, visual asset attributes, and performance requirements, through a limited number of experiments or empirical rules, and are stored in the parameter configuration unit of the embedded vision processor to ensure that the regulation rule benchmark is consistent for different rendering cycles and different dynamic scenes.

[0021] After completing the preliminary operations, timing deviation quantization is performed. The input data for this step includes four types of core parameters: Time-normalized eigenvalues , Time-normalized eigenvalues Total number of visual reference features and feature time-series weight coefficients .in, The time frame refers to the time sequence number of the currently processed consecutive video frames. The single-frame timing interval is the time interval between two adjacent frames. This parameter is directly extracted from the frame rate parameter of the hardware rendering engine. For example, when the rendering frame rate is 60 frames per second, the single-frame timing interval is... The value is Second, the unit of measurement is time. The moment is The time sequence node of the previous frame. express At this moment Feature values ​​of standardized time-series feature data. This represents the feature value of the same standardized temporal feature data from the previous frame; both are dimensionless values ​​after normalization preprocessing. The formula for calculating the temporal deviation quantization factor used in this step is: ; The formula is designed based on the fact that the essence of inter-frame lag bias is the rate of change of visual features per unit time. Simple feature differences only reflect the absolute amount of change and cannot represent the speed of change. It must be coupled with the temporal physical interval to be transformed into a physically meaningful bias quantification parameter. The overall calculation logic of the formula adopts a three-layer progressive structure. The first layer is the inter-frame feature difference calculation, which is performed by obtaining... Time and absolute difference of eigenvalues ​​of the same path at time 1 The use of absolute difference calculations avoids the problem of positive and negative biases canceling each other out during the summation process, thus fully preserving the lag bias data of all feature dimensions. The second layer is a weighted normalization operation, calculated by first using the feature time-series weight coefficients. The absolute differences of each feature are weighted, and then... By summing the weighted differences, we obtain the weighted sum term in the numerator. The purpose of weighted processing is to adapt to the differences in temporal sensitivity of different visual features, so that feature dimensions that are more sensitive to temporal lag have a greater contribution weight in bias quantification. Subsequently, the weighted sum is divided by the current... All moments Sum of road eigenvalues The normalization process is then completed. Normalization eliminates numerical interference caused by the total feature amplitude, ensuring a unified comparison benchmark for deviations between different frames and in different scenes. The third layer is a physical-temporal coupling operation, which calculates the normalized dimensionless value by multiplying it by the single-frame temporal interval. reciprocal This process completes the dimensional transformation. Through this operation, the originally dimensionless pure numerical deviation is transformed into a time-series deviation change rate with temporal physical meaning.

[0022] The following section provides a detailed explanation of the dimensions and physical meanings of each parameter in the formula for calculating the timing deviation quantization factor. (Parameters in the formula) for Time of the first The eigenvalues ​​of the standardized time-series characteristic data have undergone amplitude and scale normalization. These values ​​are dimensionless pure numerical values, having eliminated amplitude and scale differences, and their dimension is one. .parameter for The feature values ​​of the standardized time-series feature data at the same time point are also dimensionless pure numerical values, with dimensions of . Total number of visual reference features Let be a dimensionless counting parameter, representing the total number of feature dimensions involved in the calculation, with dimensions of . Feature time-series weighting coefficients A pre-configured dimensionless weighting factor, used to characterize the first... The sensitivity of road characteristics to timing deviations, with dimensions of Since all the above parameters have the same dimension, the dimensions remain constant throughout the entire process of characteristic difference operation, weighted summation operation, and normalized fraction operation. .parameter The single-frame timing interval is a timing parameter with a clear physical meaning. Its value is directly obtained from the rendering hardware's operating parameters, and its unit is time, seconds. After completing the normalized fractional operation, a dimensionless intermediate result is obtained, reflecting the relative magnitude of the feature deviation between two adjacent frames. This dimensionless intermediate result is then multiplied by... Later, due to The dimensions are its reciprocal The dimensions are That is, per second, therefore the final output timing deviation quantization factor The dimensions are Its SI unit is per second. This dimensional design perfectly matches the definition of physical temporal change rate, characterizing the rate of change of visual reference feature hysteresis bias per unit time. Assuming that in a specific embodiment, the single-frame temporal interval... Values Seconds, corresponding to a rendering frame rate of 60 frames per second; the calculated normalized weighted relative deviation is... If the characteristic change rate is 3%, then the time series deviation quantification factor is... The calculation result is Divide by , approximately equal to per second indicates the rate of change of the feature hysteresis bias between the current two frames is approximately per second. The relative characteristic change per unit. This quantification result with clear physical dimensions can be directly used for subsequent distortion correction calculations, providing a precise physical timing basis for the adaptive control of dynamic rendering.

[0023] The third step of this method integrates temporal lag parameters, temporal physical intervals, and component differences between the two feature domains to achieve coupled quantization of temporal and spatial two-dimensional distortions, yielding cross-domain mapping distortion coefficients. The distortion quantization system constructed in this step breaks through the single-dimensional limitation of traditional techniques that only correct spatial mapping distortions in the feature domain. Traditional solutions completely ignore temporal distortions caused by temporal lag biases, and the two types of distortions overlap and interfere with each other during dynamic rendering, resulting in incomplete distortion correction. This step incorporates temporal dimension distortion errors and cross-domain mapping errors in the same quantization system, obtaining comprehensive distortion quantization parameters through fusion calculations. The formula for calculating the cross-domain mapping distortion coefficients used in this step is: ; The formula is designed based on the fact that the overall distortion in dynamic rendering scenarios originates from two relatively independent dimensions: one is the temporal distortion caused by inter-frame temporal lag, and the other is the spatial distortion during feature domain mapping. These two dimensions are orthogonal in their physical causes and mechanisms, making a square root approach suitable for fusion and coupling. The formula consists of two main computational modules: a temporal distortion conversion module and a spatial mapping distortion quantization module. The first module performs the temporal distortion conversion operation, which is calculated by first taking the temporal deviation quantization factor. With single-frame timing interval The product of these factors is then combined with the timing distortion coupling coefficient. Multiply, we get Finally, the term is squared. The purpose of this operation is to determine the time series deviation quantization factor. The dimensions are This reflects the rate of change of temporal deviation, not a direct distorted variable, and cannot be directly fused with spatial mapping distortion. and Multiply, The dimensions are The dimensions cancel each other out when the two are multiplied together. This completes the dimensional transformation from the rate of change of time series to the dimensionless time series distortion variable. Time series distortion coupling coefficient. This is a dimensionless adjustment factor that scales the temporal deviation values ​​after dimensional transformation to a range matching the spatial mapping distortion values, ensuring that the two modules do not become inaccurate due to differences in magnitude during fusion. After squaring, the output of the first module is the dimensionless temporal distortion component value. The second module performs spatial mapping distortion scaling, calculating the feature component value of the original feature domain for each dimension of the feature component from the original feature domain to the target rendering domain. Feature component values ​​of the target rendering domain The difference is obtained by squaring the difference. Then put all The numerator is obtained by summing the squared differences of the characteristic components. .in, Represents the first in the original feature domain The values ​​of each feature component, This represents the numerical value of the corresponding feature component in the target rendering domain; both are dimensionless parameters that have undergone normalization. For reference, the total number of feature components, The maximum normalization threshold is a pre-defined normalization constant based on the range of feature values, used to map the sum of squared differences to a standardized numerical range. This is achieved by dividing by... and The product of these values ​​performs a dual normalization process on the number of feature components and the extrema of the features, eliminating the interference of feature dimension and amplitude range on the distortion values. After the above calculations, the output of the second module is also a dimensionless, standardized spatial mapping distortion value. The values ​​from the two modules are squared and summed, and then the square root of the sum is taken to finally output the cross-domain mapping distortion coefficients. Dimensions are This is a dimensionless distortion coefficient, a purely numerical value. This coefficient comprehensively and accurately reflects the overall distortion level under different dynamic scenarios, including both reference relationship deviation distortion caused by temporal lag and mapping error distortion caused by feature domain transformation. Assuming that in a certain embodiment, the temporal distortion coupling coefficient... Value Time-series deviation quantization factor for Seconds, single frame time interval for The time distortion component is calculated to be seconds. Meanwhile, the spatial mapping distortion module calculates the normalized spatial distortion component values ​​as follows: The cross-domain mapping distortion coefficient is obtained by summing the squares and taking the square root. The calculation result is This value comprehensively reflects the overall distortion level of the current frame's dynamic rendering scene, providing precise quantitative distortion data support for the subsequent dynamic adjustment of consistency weights.

[0024] The fourth step of this method involves linking temporal deviation parameters, distortion coefficients, and cross-scale feature parameters to generate an adaptive total weight for consistency control through nonlinear computation. This step constructs a dynamic weight adjustment model with multi-factor nonlinear linkage, fundamentally addressing the technical problem that traditional fixed weights cannot adapt to real-time changes in dynamic scenes. Traditional visual asset consistency control uses fixed adjustment weights, which cannot be adjusted according to the real-time state of temporal deviation, mapping distortion, and cross-scale features. Fixed weights over-adjust in some scenarios and under-adjust in others, failing to achieve dynamic adaptation of reference relationship consistency. This step integrates the three factors—distortion degree, temporal deviation magnitude, and cross-scale feature differences—into a single weight calculation formula through nonlinear computation, generating a total control weight that can adaptively change according to the real-time state of each frame. The formula for calculating the total weight for reference relationship consistency control used in this step is: ; The formula is designed based on the fact that the overall weight for consistency control needs to simultaneously respond to three different regulatory demands: distortion suppression, temporal correction, and cross-scale feature correction. These three demands have different directions and mechanisms of influence on the weights; therefore, a nonlinear function architecture with a numerator and denominator separation is suitable for achieving multi-factor linkage. The formula is divided into a distortion negative feedback regulation module in the numerator and a multi-dimensional nonlinear correction module in the denominator. The numerator adopts an exponential decay function architecture, and the calculation method involves using the distortion decay regulation factor... With cross-domain mapping distortion coefficient Multiplication, with the product used as the independent variable of the exponential function, is calculated. Then the calculation results are compared with the consistency basic control coefficient. Multiplication. The core characteristic of the exponential decay function is that when the independent variable... As the distortion increases, the function value decreases exponentially. Applying this characteristic to distortion control can achieve a negative feedback mechanism where the larger the distortion, the lower the consistency weight. When the distortion level of a dynamically rendered scene intensifies, the distortion coefficients are mapped across domains. The value increases, The value of the numerator decreases accordingly, reducing the overall value of the numerator and ultimately leading to a decrease in the total weight of the consistency control. This negative feedback mechanism automatically adapts the correlation strength of the reference relationship to the degree of scene distortion, proactively weakening the constraint strength of the reference relationship in high-distortion scenarios, effectively suppressing reference relationship matching errors caused by distortion. (Consistency base adjustment coefficient) Its function is to set the basic benchmark for weights; when the distorted variable approaches zero, the numerator approaches zero. This ensures that consistency control maintains a reasonable baseline strength even in low-distortion scenarios. The denominator implements a nonlinear joint correction for timing bias and cross-scale features. The denominator is calculated by first calculating the timing bias coupling factor. Timing deviation quantization factor With single-frame timing interval The product of the three Then calculate the scale feature correlation coefficient. Total normalized features across scales product ,right Plus Taking the natural logarithm of the result, we get Finally, multiply the two product terms together to get... , and then with The sums form the denominator. Among them, the total normalized features across scales... It is a dimensionless comprehensive index obtained by comprehensively statistically analyzing visual feature data at different scales according to preset fusion rules, used to reflect the total amount of cross-scale feature information of the visual asset in the current frame. Scale feature correlation coefficient This is a dimensionless adjustment factor used to control the sensitivity of cross-scale feature differences to weight adjustments. Natural logarithm function. Its core characteristic lies in its non-linear and gradually flattening growth trend. When increasing from zero, the logarithmic value grows at a relatively rapid rate, followed by a gradual slowdown in the growth rate. This non-linear characteristic ensures a smooth transition in the adjustment of weights based on cross-scale feature differences, avoiding drastic fluctuations in weights caused by abrupt changes in scale characteristics and guaranteeing the stability of the reference relationship adjustment. The denominator dynamically adjusts the baseline weights calculated in the numerator using the product of the time series deviation and the cross-scale feature differences. As the time series deviation and cross-scale feature differences increase, the denominator increases, resulting in an increase in the total weight. This results in a corresponding reduction, achieving adaptive regulation through the combined action of multiple factors. At the dimensional level, the exponential function... independent variable As a dimensionless quantity, the function value itself is also a dimensionless pure numerical value; the natural logarithm function independent variable Both are dimensionless quantities, and function values ​​are dimensionless. Dimensions are , Dimensions are The dimensions cancel each other out when the two are multiplied together. , Since the coefficients are dimensionless, all operational terms in the denominator have dimensions. The dimensions of the numerator and denominator remain unchanged throughout the entire calculation process. The final output reference relationship consistency control total weight The dimensions are In one embodiment, the consistency-based control coefficient Set as Distortion attenuation regulator Set as Cross-domain mapping distortion coefficients for The calculated molecule is Timing deviation coupling factor Set as Time-series deviation quantization factor for Seconds, single frame time interval for Seconds, scale feature correlation coefficient Set as Total cross-scale normalized features The statistical value in the current frame is Calculations yielded The denominator is The final reference relationship consistency control total weight The calculation result is Divide by , approximately equal to This weight value indicates that under the dynamic scene conditions of the current frame, the overall distortion level is low, and the temporal deviation and cross-scale feature differences are within the normal range, and the correlation strength of the visual asset reference relationship remains at a high level close to the baseline.

[0025] The fifth step of this method is to control the total weight based on the consistency of the reference relationship. This step involves adjusting the correlation strength of geometric features, texture features, and lighting features of the visual assets to reconstruct the reference relationships. Geometric features, texture features, and lighting features are the three core feature dimensions that constitute the rendering effect of visual assets. These three types of features differ significantly in terms of dynamic change patterns and distortion sensitivity. Geometric features typically change relatively smoothly and have relatively low distortion sensitivity; lighting features are most significantly affected by dynamic scenes and have high distortion sensitivity; texture features fall somewhere in between. If a uniform adjustment is used, it is impossible to accurately optimize for the differentiated distortion states of different feature dimensions, which can easily lead to the failure of consistency for single-dimensional features. The specific reconstruction operation in this step is to multiply the original correlation strength of the geometric features by the total weight of the reference relationship consistency control. This yields the adjusted geometric feature association strength; the original association strength of the texture features is then multiplied by... This yields the adjusted texture feature correlation strength; the original correlation strength of the illumination features is then multiplied by... The adjusted illumination feature correlation intensity is obtained. Because... It is a dynamically changing weight factor that adapts to the scene state. The correlation strength of the three types of features is adjusted under a unified overall weight, but each maintains its own adjustment ratio due to the difference in the original correlation strength. This reconstruction method achieves an organic combination of unified control of the overall weight and precise adaptation of each dimension, ensuring the overall coordination of the reference relationship of the three core features while preserving the inherent differences in the features of each dimension. The reconstructed reference relationship can maintain a consistent match between the geometric shape, texture details, and lighting effects of the visual assets under different distortion states and different temporal deviations, avoiding rendering distortion caused by the breakage of local dimensional reference relationships.

[0026] The sixth step of this method involves rendering the visual assets based on the reconstructed reference relationships. The rendering process is integrated with the embedded vision processor's rendering pipeline. The reconstructed geometric feature association strength, texture feature association strength, and lighting feature association strength are fed into the corresponding processing stages of the rendering pipeline to guide the geometric model reconstruction, texture mapping, and lighting calculation of the visual assets. In the geometric model reconstruction stage, the adjusted geometric feature association strength controls the matching accuracy of geometric features at different detail levels of the visual assets. In the texture mapping stage, the adjusted texture feature association strength controls the association matching range of texture sampling and filtering. In the lighting calculation stage, the adjusted lighting feature association strength controls the degree of association between lighting model parameters and the surface attributes of the visual assets. The rendering pipeline generates the final image of the visual assets frame by frame based on the updated reference relationships.

[0027] After a single frame is rendered, the method enters a closed-loop iterative mode. The embedded vision processor cyclically executes all steps from feature acquisition and preprocessing, parameter calculation, reference relationship reconstruction to rendering output, iteratively processing the visual assets of consecutive video frames. The cyclic iterative processing mode ensures that each frame completes parameter updates and relationship reconstruction based on the latest dynamic scene data, continuously adapting to real-time changes in the scene and guaranteeing a smooth transition of reference relationships between adjacent frames. During inter-frame iteration, the feedback information from the rendering output of the previous frame can be used as a calibration reference for feature acquisition in the next frame, helping to improve the feature extraction accuracy of subsequent frames. This closed-loop iterative logic constructs a continuously dynamic consistency control system, ensuring the temporal continuity and stability of dynamically rendered images, and avoiding the problems of screen flickering, feature jumps, and temporal breaks caused by the traditional single-frame independent processing mode.

[0028] As a further improvement, the embedded vision processor in this embodiment incorporates an unsupervised learning model. This unsupervised learning model is trained using massive amounts of dynamic visual asset rendering video data to generate a visual feature extraction network. The number of nodes in the input layer of the visual feature extraction network is consistent with the dimension of the output standardized temporal feature data, the number of nodes in the output layer is the same as the input layer, and the number of nodes in the hidden layer is dynamically adjusted according to the dimension of the input data. Specifically, the adjustment mechanism involves real-time detection of the dimension of the currently input standardized temporal feature vector. Set the number of hidden layer nodes to That is, the ratio of dimension to the number of hidden layer nodes is fixed. This proportional relationship has been experimentally verified, ensuring both feature compression and reconstruction accuracy while maintaining network computational efficiency. When a change in the dimension of the input feature vector is detected, the network configuration unit automatically generates and loads a new network structure configuration file within a preset adjustment time threshold, without affecting the continuous operation of real-time feature extraction. The total training data for the unsupervised learning model exceeds 100,000 hours of dynamically rendered video data, covering rich samples under different types of visual assets, varying dynamic scene complexities, and different rendering parameter conditions. The initial learning rate of the model is set to... Batch size is The training rounds are The optimizer uses the Adam optimizer, and the loss function is the mean squared error function. Model parameter optimization is achieved by minimizing the reconstruction error between the input and output layer data. After training, the network has the ability to automatically extract semantically consistent standardized temporal features from raw visual frame data, providing high-quality model support for front-end feature acquisition and normalization preprocessing. This unsupervised learning model also has a periodic update mechanism. When the newly accumulated dynamic rendering video data reaches a preset data volume threshold, the model retraining process is automatically triggered to continuously optimize the network's extraction accuracy and adapt to the iterative development of dynamic rendering technology.

[0029] As an alternative improvement, differentiated normalization strategies can be employed for different categories of visual features during the normalization preprocessing stage. For geometric feature data, since its numerical distribution is usually relatively concentrated and lacks extreme outliers, linear normalization can be used to linearly map the feature values ​​to a preset normalization interval. For illumination feature data, due to the influence of light source changes in dynamic scenes, the numerical distribution may exhibit significant skewness. A quantile-based normalization method can be used, standardizing the data based on the median and interquartile range to effectively suppress the interference of extreme illumination changes on the normalization results. For texture feature data, a variance-based standardization method can be used, subtracting the mean from the feature value and dividing by the standard deviation, allowing different texture features to be compared at a uniform scale. These differentiated normalization strategies further improve the quality of standardized temporal feature data, providing more accurate data input for subsequent quantization operations.

[0030] Characteristic temporal weight coefficient in the formula for calculating the temporal deviation quantization factor The specific value can be optimized and determined through statistical analysis of historical rendering data. Specifically, feature data from multiple consecutive video frames within the same dynamic rendering scene over a preset time period is collected. The ratio of the standard deviation to the mean of each feature data point between adjacent frames is calculated to obtain the temporal fluctuation coefficient for each feature. The temporal fluctuation coefficients of each feature are then normalized so that their sum satisfies a preset summation constraint. The normalized result is the value corresponding to each feature. Value assignment. This method assigns weights based on the statistical characteristics of actual dynamic scene data, making... The value of can truly reflect the differences in temporal sensitivity of different features in the actual rendering environment. Compared with human experience-based assignment, it has stronger data basis and scene adaptability.

[0031] Maximum normalized threshold in the cross-domain mapping distortion coefficient calculation formula Its value can be set differently depending on the type of visual asset being rendered. For industrial visual asset rendering scenarios where geometric accuracy is the core requirement, the numerical accuracy of geometric feature components is required to be high. It can be set to a preset upper quantile value for the geometric feature numerical distribution to finely distinguish subtle differences in geometric feature domain mapping distortion. For film-level visual asset rendering scenes where visual realism is a core requirement, the importance of lighting and texture features increases. Different settings can be applied to different types of feature components, such as the illumination feature component. With texture feature components Each feature can be configured independently based on its own numerical range to achieve refined quantification of distortion by type.

[0032] The total amount of cross-scale normalized features in the formula for calculating the total weight of reference relationship consistency control The calculation method can be based on hierarchical fusion of multi-scale feature pyramids. Specifically, the standardized temporal feature data of the current frame is decomposed into multiple levels of feature components according to different spatial resolutions or frequency domain scales. The statistical feature values ​​of each level's feature components are calculated separately, and then weighted and fused according to preset hierarchical weights to obtain the final cross-scale normalized feature total. Under this calculation method, The numerical value can comprehensively reflect the cross-scale feature information of visual assets from detailed texture to overall shape, providing more comprehensive scale feature parameters for the denominator nonlinear correction module, and further improving the accuracy and robustness of weight control.

[0033] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for constructing reference relationships and controlling consistency in visual asset generation, characterized in that, The embedded vision processor performs the following steps: S1: Collect the geometric features, texture features, and lighting features of dynamic visual assets, and perform normalization preprocessing on the geometric features, texture features, and lighting features to obtain standardized temporal feature data; S2: Extract visual assets to generate single-frame temporal intervals, and calculate the temporal deviation quantification factor by combining standardized temporal feature data; S3: Calculate the cross-domain mapping distortion coefficient based on the temporal deviation quantization factor, the temporal interval of the visual asset generation single frame, the original feature domain reference feature components, and the target rendering domain reference feature components. S4: Calculate the total weight of reference relationship consistency control based on the temporal deviation quantization factor, the temporal interval of single frame generation of visual assets, the cross-domain mapping distortion coefficient, and the total amount of cross-scale normalized features of visual assets. S5: Based on the consistency of reference relationships, control the total weight and reconstruct the reference relationships of visual assets; S6: Based on the reference relationships of the reconstructed visual assets, perform visual asset rendering output operations.

2. The method for constructing reference relationships and controlling consistency in visual asset generation according to claim 1, characterized in that, Geometric features, texture features, and lighting features are all visual feature data corresponding to consecutive video frames in a dynamic real-time rendering scene. Geometric features, texture features, and lighting features are acquired synchronously.

3. The method for constructing reference relationships and controlling consistency in visual asset generation according to claim 2, characterized in that, Normalization preprocessing includes amplitude normalization and scale normalization. Amplitude normalization eliminates the interference of differences in visual feature amplitudes, while scale normalization unifies the numerical scale standard of visual features.

4. The method for constructing reference relationships and controlling consistency in visual asset generation according to claim 3, characterized in that, The input data used to calculate the temporal deviation quantization factor includes the normalized feature value at time t, the normalized feature value at time t minus the temporal interval of the visual asset generation single frame, the total number of visual reference features, and the feature temporal weight coefficient.

5. The method for constructing reference relationships and controlling consistency in visual asset generation according to claim 4, characterized in that, The temporal deviation quantization factor is a physical temporal quantization parameter obtained by weighted summation of inter-frame feature differences, normalization of the total number of features in the current frame, and coupling of single-frame temporal intervals. The temporal deviation quantization factor is used to characterize the degree of temporal lag deviation of visual reference features between consecutive video frames.

6. The method for constructing reference relationships and controlling consistency in visual asset generation according to claim 5, characterized in that, The cross-domain mapping distortion coefficient is a comprehensive distortion parameter obtained by fusing temporal dimension distortion error and feature domain cross-domain mapping error. The cross-domain mapping distortion coefficient is used to characterize the overall distortion degree during the conversion process from the original feature domain to the target rendering domain.

7. The method for constructing reference relationships and controlling consistency in visual asset generation according to claim 6, characterized in that, The total weight of reference relationship consistency control is a dynamic control parameter obtained based on distortion attenuation control logic, temporal deviation coupling logic, and cross-scale feature nonlinear correction logic. The total weight of reference relationship consistency control is used to adaptively control the association strength of reference relationships in visual assets.

8. The method for constructing reference relationships and controlling consistency in visual asset generation according to claim 7, characterized in that, The operation of reconstructing the reference relationships of visual assets includes controlling the total weight based on the consistency of the reference relationships and adjusting the correlation strength of geometric features, texture features, and lighting features of the visual assets.

9. The method for constructing reference relationships and controlling consistency in visual asset generation according to claim 8, characterized in that, After completing the visual asset rendering output operation, the entire process of feature acquisition preprocessing, parameter calculation, reference relationship reconstruction, and rendering output is executed in a loop to achieve iterative processing of visual assets in consecutive frames.

10. The method for constructing reference relationships and controlling consistency in visual asset generation according to claim 9, characterized in that, Before performing the calculation of the temporal deviation quantization factor, the total number of visual reference feature paths and the total number of reference feature components are determined in advance, and the initial configuration of the consistency basic control coefficient, distortion attenuation control factor, temporal deviation coupling factor and scale feature correlation coefficient is completed.

Images