A promotion method and system for ice powder marketing

Abstract

The application discloses a popularization method and system for ice powder marketing, and particularly relates to the technical field of food UGC intelligent screening and pushing, and is used for solving the problems of false screening of ice powder UGC materials and low reach rate of high-quality content in the prior art. A multimodal feature template is constructed by collecting ice powder material refraction disturbance basic parameters and ingredient texture distribution basic parameters. An optical consistency evaluation model containing refraction gradient matching rules and texture clustering rules is established based on the template. Material refraction disturbance tensors and ingredient texture vortex features are extracted from user UGC materials. A refraction deviation coefficient is generated by dynamic matching, and a texture matching degree is generated by combining vortex core density analysis. An optical consistency score is generated according to the two parameters, and candidate materials are screened. Based on the score, segmented correlation and differentiated marketing strategies are associated, user interaction heat data is corrected synchronously, and the pushing priority is adjusted, so that the identification accuracy of high-quality content is significantly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of food UGC intelligent screening and pushing, more specifically, the present application relates to a promotion method and system for ice powder marketing. BACKGROUND

[0002] In the marketing of ice powder products, user-generated content (UGC) can enhance brand influence. Consumers can effectively expand product coverage and enhance user stickiness by spontaneously spreading content such as shooting ice powder eating scenes, creative combinations, or production processes on social media platforms. However, the current screening and promotion of UGC content mainly rely on manual review or automated tools based on general image recognition technology. Due to the unique physical properties of ice powder, such as semi-transparent texture, multiple ingredient combinations, and light sensitivity, its visual presentation is easily affected by environmental light, differences in shooting equipment, and user operation level, resulting in uneven quality of materials. For example, brown sugar ice powder in a dark environment may be misidentified as a dark drink, and the layering of ingredients in a dynamic video may be blurred due to camera shaking.

[0003] In the prior art, the quality control method for ice powder UGC materials has defects: general image / video processing algorithms do not combine the physical properties of ice powder (such as the optical reflection law of semi-transparent materials and the color superposition effect of ingredients), resulting in insufficient accuracy and adaptability of content quality assessment, so that a large number of high-quality content that meets the brand tone cannot be screened and promoted in real time, and low-quality or irrelevant content may be incorrectly promoted, ultimately reducing the overall conversion efficiency of UGC marketing. SUMMARY

[0004] To overcome the above-mentioned defects of the prior art, the embodiments of the present application provide a promotion method and system for ice powder marketing to solve the problems raised in the background art.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0006] A promotion method for ice powder marketing, comprising:

[0007] S1, collecting material refraction disturbance basic parameters and ingredient texture distribution basic parameters of ice powder, and generating a multi-modal feature template;

[0008] S2, constructing an optical consistency evaluation model of ice powder UGC materials according to the multi-modal feature template, the optical consistency evaluation model comprising refraction gradient matching rules and texture clustering rules;

[0009] S3, extracting material refraction disturbance tensors and ingredient texture vortex features of consecutive frames from the ice powder UGC materials uploaded by users;

[0010] S4, dynamically match the material refraction disturbance tensor based on the refraction gradient matching rule to generate the refraction deviation coefficient, and analyze the vortex core density of the ingredient texture vortex feature based on the texture clustering rule to generate the texture matching degree;

[0011] S5, generate an optical consistency score according to the refraction deviation coefficient and the texture matching degree, and select candidate materials based on the optical consistency score;

[0012] S6, according to the optical consistency score of the candidate material and the user interaction heat data, push the candidate material to the target user and associate the marketing incentive strategy.

[0013] In a preferred embodiment, the material refraction disturbance basic parameters and the ingredient texture distribution basic parameters of the ice powder are collected to generate a multi-modal feature template, including:

[0014] The optical parameters of the semi-transparent material of the ice powder under dynamic refraction in a standard lighting environment are collected as the material refraction disturbance basic parameters, and the texture morphology parameters of the ice powder ingredients in a static distribution state are collected as the ingredient texture distribution basic parameters;

[0015] By quantifying the refraction rate change gradient sequence of the material refraction disturbance basic parameters and the spatial topological structure of the ingredient texture distribution basic parameters, a multi-modal feature template that integrates optical characteristics and texture characteristics is generated.

[0016] In a preferred embodiment, the material refraction disturbance basic parameters include a refraction angle offset dataset of the ice powder material in the visible spectrum range, and the ingredient texture distribution basic parameters include a geometric profile feature of the ingredient particles and a color superposition effect dataset.

[0017] In a preferred embodiment, an optical consistency evaluation model of the ice powder UGC material is constructed according to the multi-modal feature template, and the optical consistency evaluation model includes a refraction gradient matching rule and a texture clustering rule, including:

[0018] The refraction gradient matching rule is constructed based on the refraction rate change gradient sequence in the multi-modal feature template, and a dynamic optical correction function library is generated by establishing a mapping relationship between the gradient sequence and the refraction characteristics of the ice powder material under standard lighting conditions;

[0019] The texture clustering rule is constructed based on the spatial topological structure index and the color superposition effect dataset in the multi-modal feature template, and a texture feature classification decision tree is generated by defining a local aggregation threshold and a global distribution uniformity boundary of the ingredient particles;

[0020] The dynamic optical correction function library and the texture feature classification decision tree are integrated into a machine learning framework to train an optical consistency evaluation model, wherein a dimension of an input layer of the optical consistency evaluation model is consistent with a vector dimension of the multi-modal feature template, and an output layer is an optical consistency score.

[0021] In a preferred embodiment, material refraction disturbance tensors and ingredient texture vortex features of consecutive frames are extracted from user-uploaded ice powder UGC materials, including:

[0022] Dynamic optical flow field analysis is performed on user-uploaded ice powder UGC video materials to generate material refraction light variation trajectories of consecutive frames;

[0023] Based on the material refraction light variation trajectories, a refraction angle offset difference matrix of adjacent frames is calculated, and a three-dimensional material refraction disturbance tensor is constructed by time series stacking;

[0024] Motion trajectory tracking is performed on ingredient particles in consecutive frames to generate an ingredient displacement vector field;

[0025] Local rotating flow patterns are detected in the ingredient displacement vector field, and the curvature radius and angular velocity distribution of the rotating flow core region are extracted;

[0026] The reciprocal sequence of the curvature radius and the angular velocity distribution matrix are fused into ingredient texture vortex features.

[0027] In a preferred embodiment, the material refraction disturbance tensor is dynamically matched based on a refraction gradient matching rule to generate a refraction deviation coefficient, including:

[0028] The material refraction disturbance tensor is decomposed into a sequence of consecutive tensors along the time dimension, and each sub-tensor is pattern-matched with the dynamic optical correction function library in the refraction gradient matching rule through a dynamic time series alignment window;

[0029] The residual norm of the sub-tensor feature channel and the reference mode of the dynamic optical correction function library is calculated, and the refraction deviation coefficient is generated by weighted aggregation of the residual norm. The weight value is allocated according to the illumination stability of the sub-tensor on the time axis.

[0030] In a preferred embodiment, the ingredient texture vortex feature is analyzed based on a texture clustering rule to generate a texture matching degree, including:

[0031] An effective region with a vortex core density exceeding a preset density threshold is selected in the ingredient texture vortex feature, and a vorticity distribution matrix of the effective region is extracted;

[0032] The vorticity distribution matrix is input into a texture feature classification decision tree in the texture clustering rule, and the decision tree nodes are traversed to determine the path;

[0033] Generate a texture matching degree according to the leaf node belonging probability distribution of the decision path.

[0034] In a preferred embodiment, an optical consistency score is generated according to the refraction deviation coefficient and the texture matching degree, and candidate materials are screened based on the optical consistency score, including:

[0035] A product relationship model of the refraction deviation coefficient and the texture matching degree is established, and the product value is mapped to the optical consistency score through a nonlinear transformation function, and the parameters of the nonlinear transformation function are determined according to the score distribution fitting of historical samples;

[0036] A double-threshold screening mechanism is set up by setting the allowed upper limit of the refraction deviation coefficient and the allowed lower limit of the texture matching degree, and only when the candidate material meets the condition that the refraction deviation coefficient does not exceed the allowed upper limit and the texture matching degree is not lower than the allowed lower limit at the same time, it enters the candidate set;

[0037] The candidate materials in the candidate set are arranged in descending order of optical consistency score, and a preset number of materials with high ranking are selected as the final output.

[0038] In a preferred embodiment, according to the optical consistency score of the candidate material and the user interaction heat data, the candidate material is pushed to the target user and associated with the marketing incentive strategy, including:

[0039] The candidate materials are divided into high-score section, medium-score section and low-score section according to the optical consistency score, and the division boundary is dynamically adjusted according to the statistical distribution characteristics of historical score samples;

[0040] Material preference features of the target user are extracted from the user interaction heat data, and the push priority order of the materials in each section is adjusted through a heat correction factor;

[0041] A mapping rule between the section and the marketing incentive strategy is established, the high-score section is associated with a brand exposure type incentive strategy, the medium-score section is associated with a conversion guide type incentive strategy, and the low-score section is associated with a user feedback type incentive strategy;

[0042] The candidate materials are pushed to the target user according to the corrected push priority, and the associated marketing incentive strategy execution component is activated synchronously.

[0043] On the other hand, the present application provides a promotion system for ice powder marketing, comprising:

[0044] A material modeling module for collecting the material refraction disturbance basic parameters and the ingredient texture distribution basic parameters of the ice powder, and generating a multi-modal feature template;

[0045] A model evaluation module for constructing an optical consistency evaluation model of the ice powder UGC material according to the multi-modal feature template, the optical consistency evaluation model comprising a refraction gradient matching rule and a texture clustering rule;

[0046] A motion feature module is used to extract the material refraction disturbance tensor and ingredient texture vortex feature of continuous frames from the user-uploaded ice powder UGC material;

[0047] A double-matching module is used to dynamically match the material refraction disturbance tensor based on a refraction gradient matching rule to generate a refraction deviation coefficient, and to analyze the vortex core density of the ingredient texture vortex feature based on a texture clustering rule to generate a texture matching degree;

[0048] A screening candidate module is used to generate an optical consistency score according to the refraction deviation coefficient and the texture matching degree, and to screen candidate materials based on the optical consistency score;

[0049] A strategy execution module is used to push the candidate materials to target users and associate marketing incentive strategies according to the optical consistency score of the candidate materials and user interaction heat data.

[0050] Compared with the prior art, the present application has the following beneficial effects:

[0051] 1. By converting the physical and optical properties of ice powder and the ingredient motion features into quantifiable analysis models, the accuracy and adaptability of UGC material quality evaluation are significantly improved. For the light-sensitive characteristics of semi-transparent materials, the material refraction disturbance tensor dynamic matching mechanism effectively overcomes the problem of material misjudgment caused by changes in ambient light, ensuring that dark-colored materials such as brown sugar ice powder can still be accurately identified in dark environments. Based on the texture vortex core density analysis of fluid mechanics principles, the hierarchical structure and motion trajectory of ingredients in dynamic videos are accurately captured, eliminating texture blur interference caused by camera shaking. This dual-path analysis model that integrates physical properties solves the adaptability defects of traditional image recognition algorithms in the ice powder UGC scene.

[0052] 2. The hierarchical screening and pushing mechanism based on the optical consistency score realizes the deep coupling of content quality and user preferences. By dynamically associating quality scores with user interaction heat, the system intelligently corrects the pushing priority to avoid burying high-quality low-heat materials. At the same time, according to the score partition matching differential marketing strategies (high partition intensifies brand exposure, medium partition guides conversion, and low partition collects feedback), the touch efficiency and conversion rate of high-quality content are significantly improved, not only solving the efficiency bottleneck of manual review, but also breaking through the recognition limitations of general algorithms for special material foods. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 The flowchart of the present application is a promotion method for ice powder marketing;

[0054] Figure 2 The structural schematic diagram of the present application is a promotion system for ice powder marketing. DETAILED DESCRIPTION

[0055] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work are within the protection scope of the present application.

[0056] Embodiment 1 Figure 1 A promotion method for ice powder marketing is given, comprising:

[0057] S1, collecting material refraction disturbance basic parameters and ingredient texture distribution basic parameters of the ice powder, and generating a multi-modal feature template;

[0058] S2, constructing an optical consistency evaluation model of the ice powder UGC material according to the multi-modal feature template, the optical consistency evaluation model comprising a refraction gradient matching rule and a texture clustering rule;

[0059] S3, extracting continuous frame material refraction disturbance tensors and ingredient texture vortex features from the ice powder UGC material uploaded by the user;

[0060] S4, generating a refraction deviation coefficient by dynamically matching the material refraction disturbance tensors based on the refraction gradient matching rule, and generating a texture matching degree by vortex core density analysis of the ingredient texture vortex features based on the texture clustering rule;

[0061] S5, generating an optical consistency score according to the refraction deviation coefficient and the texture matching degree, and screening candidate materials based on the optical consistency score;

[0062] S6, pushing the candidate materials to target users and associating marketing incentive strategies according to the optical consistency score of the candidate materials and user interaction heat data.

[0063] S1, collecting material refraction disturbance basic parameters and ingredient texture distribution basic parameters of the ice powder, and generating a multi-modal feature template, which is specifically implemented as:

[0064] The optical parameters of the ice powder semi-transparent material under standard light environment are collected as the basic parameters of the refractive disturbance. The method for establishing the standard light environment is to configure a programmable LED light source array, monitor the output characteristics of the light source in real time through a spectrum analyzer, dynamically adjust the current to stabilize the color temperature of the light source at 5500-6500 K, and use an illuminometer to calibrate the light source intensity to ensure that the sample surface illuminance is always not less than 1000 lux. The process of obtaining dynamic refraction parameters is to fix the 10 mm thick ice powder standard sample on a precision electric rotating platform, control the platform to rotate at a constant angular velocity of 5 degrees per second, and change the incident light angle in the range of 0-180 degrees. During this process, the fiber optic spectrometer captures the refraction response data of the ice powder material to the visible spectrum of 380-780 nm at a sampling frequency of 10 times per second, records a set of refraction angle measurement values for every 1 degree change in incident angle, and finally forms a refraction angle offset data set containing 181 data points, which completely covers the dynamic optical characteristics of the ice powder semi-transparent material.

[0065] The collection of ice powder ingredient texture distribution basic parameters is carried out in the same standard light environment as described above. The method for realizing the static distribution state is to uniformly spread the typical ingredients such as brown sugar, fruit particles and nut fragments on the ice powder surface in a 90 mm diameter petri dish according to the mass ratio of 1:0.5:0.3, and stand still for 300 seconds until the particles are completely stationary. The acquisition of texture morphology parameters is completed by a high-resolution digital camera equipped with a 100 mm fixed-focus lens, which is installed on a three-axis fine adjustment gimbal, and the lens plane is kept parallel to the ice powder surface at a constant distance of 300 mm through a laser range finder. The extraction method of geometric contour features is to perform Gaussian filtering on the obtained ingredient image, and then use an edge detection algorithm based on pixel gradient intensity. This algorithm identifies the particle boundary by calculating the change rate of gray value in a 3x3 pixel neighborhood, and the gray value change rate threshold for determining the existence of the boundary is set to 50-150 gradient intensity according to the average contrast value between the ice powder matrix and the ingredients measured in the pre-experiment. The generation method of color superposition effect data set is to convert the original red-green-blue three-channel image to hue-saturation-lightness color space, and then calculate the hue distribution dispersion, saturation average value and lightness fluctuation degree of the ingredient area, respectively. The hue distribution dispersion is obtained by calculating the sum of the square deviations of the hue values of all pixels in the region from their arithmetic mean value and dividing the total number of pixels.

[0066] The quantification process of the material refraction disturbance basic parameter is specifically implemented as follows: sequentially differential processing is performed on the refraction angle offset data set, the ratio of the refraction angle change amount corresponding to two adjacent incidence angle points to the incidence angle change amount is calculated in the arrangement order from small to large incidence angle, wherein the incidence angle change amount is fixed as 1 degree, and the refraction angle change amount is the difference of the actual measured value, and finally a refraction index change gradient sequence containing 180 continuous ratios is generated. The sequence is stored in an array with the incidence angle as the index, and each gradient value is a dimensionless value.

[0067] The spatial topology quantification implementation method of the ingredient texture distribution basic parameter is as follows: first, the center points of the minimum circumscribed rectangles of all ingredient particles are calculated as the centroid coordinates based on the extracted geometric contour features, and then the Delaunay triangulation algorithm of a plane point set is applied to construct a topology network. The algorithm connects three adjacent centroids to form non-overlapping triangular cells, ensuring that all particles are connected in the network. The generation of the spatial topology index includes: calculating the arithmetic mean of the lengths of the three edges of each triangular cell as the local aggregation index, and calculating the standard deviation of the lengths of all triangular cell edges as the global distribution uniformity index.

[0068] The generation of the multi-modal feature template is realized by a feature fusion engine: the refraction index change gradient sequence, the local aggregation index and the global distribution uniformity index of the spatial topology structure, and the three channel values of the color superposition effect data set are integrated. Before integration, all feature values are processed by a data standardization processor to be dimensionless. The standardization method is to apply a linear transformation formula to the numerical value of each feature dimension: standardized value=(original value-historical minimum value) / (historical maximum value-historical minimum value), wherein the historical minimum value and the historical maximum value are derived from the 1000-group ice powder sample feature extreme value database pre-stored in the laboratory. The final output multi-modal feature template is a 185-dimensional feature vector, of which the first 180 dimensions are the refraction index change gradient sequence, the 181st to 182nd dimensions are the spatial topology index, and the 183rd to 185th dimensions are the color channels.

[0069] S2, constructing an optical consistency evaluation model of ice powder UGC materials according to the multi-modal feature template, the optical consistency evaluation model comprising refraction gradient matching rules and texture clustering rules, and the specific implementation is as follows:

[0070] The refractive gradient matching rule is constructed based on the refractive index variation gradient sequence in the multi-modal feature template. The specific implementation method for establishing the mapping relationship between the gradient sequence and the refractive characteristics of the ice powder material under standard lighting conditions is as follows: a plurality of groups of refractive index variation gradient sequences under different lighting angles pre-stored in the laboratory are called as training samples, and each group of sequences includes 180 continuous gradient values. The mapping relationship is established by a least square fitting algorithm, which iteratively optimizes parameters to minimize the sum of squares of the difference between the training sample gradient sequence values and the corresponding points of the standard refractive characteristic curve, thereby generating a correlation coefficient matrix. The construction process of the dynamic light correction function library is as follows: for the case where the ambient color temperature is lower than the standard value of 5500K, a linear compensation function is designed, and the correction coefficient of the function is composed of a benchmark coefficient and a lighting attenuation factor product, wherein the benchmark coefficient is an empirical constant 1.0, and the lighting attenuation factor is calculated according to the ratio of the ambient color temperature value measured by the spectrophotometer to the standard color temperature 5500K; when the measured ambient color temperature is lower than 3000K, a fixed compensation mode is enabled, which uses a predefined compensation parameter 0.8.

[0071] The texture clustering rule is constructed based on the spatial topology structure index and the color superposition effect dataset in the multi-modal feature template. The setting method of the local aggregation degree threshold is as follows: 500 qualified ice powder sample images are collected, the local aggregation degree index value distribution of these samples is counted, and the 5th percentile value of the value distribution is taken as the lower threshold and the 95th percentile value is taken as the upper threshold. The determination of the global distribution uniformity boundary is realized by Monte Carlo simulation: the centroid points of the ingredient particles are randomly generated on a virtual ice powder surface with an area of 100 square centimeters, and the number of particles obeys a normal distribution. The global distribution uniformity index is calculated for each distribution model generated 1000 times, and finally the 3 times of the standard deviation of all results is taken as the boundary tolerance threshold. The generation process of the texture feature classification decision tree is as follows: a three-level decision structure is constructed by using a classification regression tree algorithm, the first level decision node judges whether the local aggregation degree index is within the threshold lower limit to the threshold upper limit interval, the second level decision node judges whether the global distribution uniformity index is less than the boundary tolerance threshold, and the third level decision node judges whether the hue distribution dispersion in the color superposition effect dataset is less than 25 degrees; wherein the 25 degree threshold value is determined by testing the classification accuracy corresponding to different threshold values on the validation set to select the value that makes the accuracy highest.

[0072] The dynamic optical correction function library and the texture feature classification decision tree are integrated into the machine learning framework. The integration method is to build a fully connected neural network model, and the input layer of the model is set to 185 neuron nodes, which strictly corresponds to the 185-dimensional feature vector of the multi-modal feature template generated in the previous step. The neural network hidden layer adopts a double-branch structure: the optical feature analysis branch contains 3 one-dimensional convolution layers with a convolution kernel size of 5 and a step size of 1, which are specifically used to process the first 180-dimensional refractive index gradient sequence; the texture feature analysis branch contains 2 fully connected layers with 64 and 32 neurons, respectively, which process the 5-dimensional encoding vector output by the texture feature classification decision tree. The training process adopts a supervised learning method: 2000 groups of labeled sample sets are prepared, which are divided into training set and validation set in the ratio of 8:2, and each group of samples contains complete 185-dimensional feature vector and artificial labeled optical consistency score label. The optimizer selects the Adam algorithm, the initial learning rate is set to 0.001, the learning rate is multiplied by 0.5 to decay every 10 training periods, and the training termination condition is that the mean square error of the validation set changes by less than 0.5% for 5 consecutive periods.

[0073] The output layer of the optical consistency evaluation model maps the original output value of the neural network to the range of 0 to 100 points through the Sigmoid activation function. An exception handling mechanism is set in the model deployment stage: when the dimension of the input feature vector is not equal to 185, trigger a dimension error interrupt and return error code 1001; when more than 18 abnormal outliers are detected in the refractive index gradient sequence, automatically activate the robust evaluation mode, which uses the median instead of the arithmetic mean to calculate the gradient statistical feature quantity.

[0074] S3, extract the material refraction disturbance tensor and the ingredient texture vortex feature of the continuous frames from the user uploaded ice powder UGC material, and the specific implementation is:

[0075] Perform dynamic optical flow field analysis on the user uploaded ice powder UGC video material to generate material refraction light variation trajectories of continuous frames. The specific implementation method is: use the light flow calculation method based on brightness conservation and spatial smoothing constraint to process the video frame sequence. The brightness conservation constraint means that the brightness of the same material area between adjacent frames remains unchanged, and the displacement vector is solved by establishing a brightness change equation; the spatial smoothing constraint means that the displacement vectors of adjacent pixels are forced to change continuously, and the smoothing optimization is realized through the Laplace operator. The setting basis of the brightness change threshold is: under the standard light environment of 5500 Kelvin, measure the maximum allowed brightness fluctuation range of 100 groups of ice powder samples in video shooting, and take the median value of 18 gray levels as the judgment threshold. The generation of material refraction light variation trajectories is realized through the spatiotemporal continuity constraint: taking the original frame rate of the video 30 frames / second as the benchmark, setting a 2-second analysis time window, and constructing a trajectory chain containing 60 continuous displacement vectors for each pixel point.

[0076] The refraction angle offset difference matrix of adjacent frames is calculated based on the refraction light trajectory of the material, and a three-dimensional material refraction disturbance tensor is constructed by time sequence stacking. The calculation principle of the refraction angle offset is as follows: according to Snell's law, the direction change of the displacement vector and the change of the refraction angle satisfy the geometric projection relationship, and the relative change of the refraction angle is inversely calculated by the vector decomposition method. The generation method of the difference matrix is as follows: the video frame is divided into 5*5 pixel space units, the standard deviation and kurtosis coefficient of the refraction angle change of all trajectory points in each unit are calculated, the standard deviation reflects the local change intensity, and the kurtosis coefficient represents the change distribution form. The construction of the three-dimensional material refraction disturbance tensor is realized by sequentially stacking the difference matrix of 15 continuous frames along the time axis, and the tensor dimension structure is time dimension * space unit dimension * feature channel dimension, wherein the feature channel includes three channels of standard deviation, kurtosis coefficient and change gradient.

[0077] Motion trajectory tracking is performed on the ingredients particles in the continuous frames to generate an ingredient displacement vector field. The tracking method uses an improved Kanade-Lucas-Tomasi feature point matching algorithm: first, the corner features of the ingredients particles are detected in the initial frame, and the corner selection standard is that the Harris response value is greater than 0.05 local maximum point, and the response threshold is determined by analyzing the average texture contrast of 100 sample images of the ingredients and ice powder matrix. In the subsequent frame, the matching points are searched by Newton-Raphson iteration method, and the matching similarity threshold is set to 0.7, which is determined by optimizing the tracking accuracy on the test set. The generation of the ingredient displacement vector field is realized by calculating the displacement vector of each feature point between continuous frames, and the spatial resolution of the vector field is consistent with the original resolution of the video, which is 1920*1080.

[0078] In the ingredient displacement vector field, a local rotational flow pattern is detected, and the curvature radius and angular velocity distribution of the rotational flow core area are extracted. The detection principle of the rotational flow pattern is as follows: the rotational center is identified by calculating the vorticity field of the displacement vector, and the vorticity value is calculated by the circulation density of the displacement vector in the local neighborhood. The vorticity threshold is set to 0.1 radian / frame, which is based on the average rotation intensity of the ingredients particles measured in 50 groups of typical stirring action videos. The extraction method of the curvature radius is as follows: the displacement vector in the core area is subjected to least square circular arc fitting, and the fitting circular arc radius is taken as the curvature radius. The calculation of the angular velocity distribution is as follows: the instantaneous angular velocity is obtained by dividing the azimuth angle change of the feature point by the frame interval of 0.033 seconds, and the time interval is determined by the reciprocal of the video frame rate of 30 frames / second.

[0079] The sequence of the reciprocal of the radius of curvature is fused with the angular velocity distribution matrix as the ingredient texture vortex feature. The fusion method adopts a feature concatenation strategy: for each rotating flow core area, the 60-frame time sequence average value of the reciprocal of the radius of curvature is calculated as the first dimension of the feature vector; the upper quartile, median and lower quartile of the angular velocity distribution matrix are extracted as the second to fourth dimensions. When there are multiple rotating flow areas in the video, a weighted average algorithm is used to calculate the global feature vector, and the weight value is determined according to the product of the area ratio and the average vorticity value, and the formula is: weight=area ratio x vorticity value / max vorticity value.

[0080] The design of S3 step aims at the light sensitivity of ice powder translucent material and the dynamic blur problem of ingredients, captures the dynamic changes of material refraction through optical flow field analysis, combines with rotating flow pattern analysis to solve the problem of physical property adaptation which is not considered by general image recognition algorithm. Compared with the prior art, the combination of optical refraction trajectory and fluid mechanics vorticity analysis significantly improves the accuracy of material recognition in dark light environment and the resolvability of ingredient levels in dynamic video, overcoming the two defects of "red sugar ice powder misidentification" and "blurring caused by camera shaking". The physical properties of food are converted into a calculable model, providing a new paradigm for special material UGC quality control.

[0081] S4, dynamically match the material refraction disturbance tensor to generate the refraction deviation coefficient based on the refraction gradient matching rule, and analyze the vortex core density of the ingredient texture vortex feature based on the texture clustering rule to generate the texture matching degree, which is specifically implemented as:

[0082] The material refraction disturbance tensor is decomposed into a sequence of continuous sub-tensors along the time dimension, and each sub-tensor is matched with the dynamic optical correction function library in the refraction gradient matching rule through dynamic time sequence alignment window. The specific implementation method is: according to the video frame rate characteristics, the time dimension length of the sub-tensor is determined, the principle is to ensure that each sub-tensor covers the complete material refraction change period, and the period length is obtained by pre-experiment analysis of the minimum stable period of refraction characteristics in typical ice powder video. The construction of dynamic time sequence alignment window adopts sliding window mechanism, and the window size is strictly matched with the time dimension of the sub-tensor, and the sliding step is set to 1 frame to ensure the time sequence continuity. The mode matching process uses dynamic time warping algorithm to align the sub-tensor with the reference mode, which realizes the nonlinear alignment of time axis by finding the path with the smallest cumulative distance, and the search width is adjusted adaptively according to the dynamic change intensity of video content, and the adjustment basis is to calculate the average difference rate of feature values between adjacent frames of the sub-tensor, and the difference rate threshold is set to 0.15, and when the difference rate exceeds the threshold, the search width is expanded to 30% of the window length.

[0083] The residual norm of the sub-tensor feature channel and the dynamic optical correction function library reference mode is calculated, and the refractive deviation coefficient is generated by weighted aggregation of the residual norm. The residual norm calculation adopts the Euclidean distance measurement method, that is, the square sum of the difference of the three feature channels is calculated and the square root is calculated. The three feature channels correspond to the standard deviation channel, the kurtosis coefficient channel and the change gradient channel in the material refractive disturbance tensor. The core of the weight value allocation mechanism lies in the illumination stability evaluation: first, the brightness variance of each video frame in the corresponding period of the sub-tensor is calculated, and the variance value reflects the degree of illumination fluctuation; then the variance value is converted into a weight coefficient through a stability mapping function, and the mapping function is designed as a piecewise linear function. When the variance is lower than the stability threshold, the maximum weight 1.0 is given, and when the variance exceeds the threshold, the weight is reduced according to the linear decay formula, and the decay slope is set to 0.2 weight value per unit variance. The determination of the stability threshold is to measure the brightness fluctuation limit value of 100 qualified samples under the standard laboratory 5500 Kelvin illumination environment, and the 5th percentile of the limit value distribution is taken as the reference threshold.

[0084] In the ingredient texture vortex feature, the effective area with vortex core density exceeding the preset density threshold is screened, and the vorticity distribution matrix of the effective area is extracted. The setting method of the density threshold is: collect 500 qualified ice powder UGC materials, count the distribution density value of the vortex core in each material, the density value calculation unit is the number of cores per square centimeter, and the 90th percentile of the density value distribution is taken as the reference threshold. The effective area screening adopts spatial grid analysis method: the video frame is divided into 10x10 pixel fixed size grid units, the ratio of the number of vortex cores in each unit to the unit area is calculated, and the unit with the ratio exceeding the reference threshold is selected as the effective area. The construction method of the vorticity distribution matrix is: a vorticity value statistical histogram is generated for each effective area, the histogram binning strategy adopts the equal frequency binning method based on data distribution, and the number of bins is adaptively determined according to the area and the vorticity value range. The determination principle is to meet the constraints of minimum value 5 and maximum value 20, and each bin contains data with a difference of not more than 10%.

[0085] The vorticity distribution matrix is input into the texture feature classification decision tree in the texture clustering rule, and the path is traversed. The traversal process is executed according to the hierarchical rule predefined by the decision tree: starting from the root node, the branch path is selected according to the statistical feature value of the vorticity distribution matrix, wherein the first level node is determined according to whether the average vorticity value is in the interval of 0.1 to 0.5 rad / frame, the second level node is determined according to whether the vorticity distribution skewness is less than 0.3, and the third level node is determined according to whether the vorticity peak value is less than 1.0 rad / frame. The record of the determination path is realized by storing the node identification sequence passed, and the sequence storage format is a 32-bit integer array.

[0086] The texture matching degree is generated according to the leaf node attribution probability distribution of the decision path. The probability distribution data is derived from the leaf node classification statistical information accumulated in the decision tree training stage, and the data storage format is a single-precision floating-point number vector. The calculation of the texture matching degree adopts the path weighted probability method: first, the highest category probability value corresponding to the leaf node is obtained as the basic matching degree; then a path depth correction factor is introduced to calibrate the basic value, and the correction factor calculation formula is 1.0 plus the path depth multiplied by the gain coefficient, and the gain coefficient is set to 0.03 through decision tree validation set optimization. The final matching degree is normalized to the standard interval of 0 to 1 through linear transformation, and the transformation formula is: matching degree = (original value - historical minimum value) / (historical maximum value - historical minimum value), wherein the historical extreme value is derived from the training data set statistics.

[0087] The design of the S4 step aims at the optical characteristic distortion and ingredient motion blur problems of the ice powder UGC dynamic scene, solves the problem that the general quality evaluation method is not adapted to the dynamic response of the semi-transparent material by matching the refraction characteristics and vortex topology decision through time sequence decomposition, and solves the problem that the general quality evaluation method is not adapted to the dynamic response of the semi-transparent material. Compared with the prior art, the dynamic time sequence alignment mechanism is combined with the curl distribution decision tree mapping, the refraction evaluation robustness under the video frame rate fluctuation and the motion ingredient level analysis precision are effectively improved, and the defects of "dark light refraction misjudgment" and "vortex texture blur" are overcome.

[0088] S5, generating optical consistency score according to refraction deviation coefficient and texture matching degree, and screening candidate materials based on optical consistency score, specific implementation is:

[0089] A product relationship model of the refractive deviation coefficient and the texture matching degree is established, and the product value is mapped to the optical consistency score through a nonlinear transformation function. The specific implementation method is: the core principle of constructing the product relationship model is to forcibly require that the optical material characteristics and the ingredient texture characteristics must meet the quality standard at the same time, the model directly multiplies the refractive deviation coefficient and the texture matching degree to obtain the original product value, and this design ensures that any significant defect of a parameter will cause the product value to be greatly reduced. The construction process of the nonlinear transformation function is: collect the refractive deviation coefficient and the texture matching degree product value and the corresponding artificial quality score of not less than 1000 groups of materials in the historical sample library, and solve the transformation parameters through the least square fitting algorithm. The fitting process adopts an iterative optimization strategy: the initial transformation parameters are a linear mapping relationship, then the mean square error of the transformation function output value and the artificial score under the current parameters is calculated, the parameters are adjusted by the gradient descent method to gradually reduce the error, and the gradient descent step length is initially set to 0.01. The step length is attenuated according to the cosine annealing strategy after each iteration. The iteration termination condition is that the absolute value of the error change of 10 consecutive iterations is less than 0.001. The nonlinear transformation function adopts a piecewise function form, an exponential function is used to enhance the quality difference discrimination degree when the product value is lower than 0.3, and a logarithmic function is used to avoid score saturation when the product value is higher than 0.7. The function form switching point is determined through sample data distribution inflection point detection, and the inflection point detection method is to calculate the zero point of the second derivative.

[0090] The allowable upper limit of the refractive deviation coefficient and the allowable lower limit of the texture matching degree constitute a double-threshold screening mechanism. The determination method of the allowable upper limit is: analyze the numerical distribution of the refractive deviation coefficient in the historical qualified sample library, calculate the statistical percentile value of the numerical distribution, and take the 95th percentile value as the reference upper limit value. The selection of the percentile is based on covering the refractive characteristic fluctuation range of 95% of the qualified samples. The allowable lower limit is set by similar logic: the numerical distribution range of the texture matching degree of the qualified samples is calculated, and the 5th percentile value is taken as the reference lower limit value. The execution mechanism of the double-threshold screening is: two independent judgment processes are executed in parallel for each candidate material, the first process checks whether the refractive deviation coefficient is not greater than the allowable upper limit, and the second process checks whether the texture matching degree is not less than the allowable lower limit. Only when both processes return the pass judgment, the material can enter the candidate set. The threshold dynamic updating mechanism is: collect not less than 100 groups of new qualified sample data every month, recalculate the percentile value and update the threshold, and use the sliding window mechanism to retain the latest 300 groups of samples when the sample quantity is insufficient.

[0091] The materials entering the candidate set are ranked in descending order of optical consistency score, and a preset number of materials with high ranking are selected as the final output. The sorting algorithm adopts a two-axis quick sorting optimization method, which improves the comparison operation according to the single-precision floating-point number characteristics of the optical consistency score, and sets the floating-point comparison tolerance to 0.0001 to avoid sorting displacement caused by precision errors. The determination principle of the preset number is: dynamically configured according to the application scene requirement, fixedly select the top 5 materials with the highest score in the advertisement recommendation scene, and select the top 20% materials with the highest score in the quality monitoring scene. The final output is executed before the quality review mechanism: when the difference between the highest score and the lowest score in the candidate set exceeds 3 times the standard deviation of the historical sample score, the manual review process is automatically triggered, and the materials with a score in the lowest 10% interval are reviewed by the reviewer.

[0092] The design of step S5 aims at the defect of decoupling determination of material and texture characteristics in ice powder UGC material quality evaluation. Through the product model to force the coupling of double parameters and the double threshold physical screening mechanism, the problem of unqualified comprehensive quality although the single feature meets the standard is solved, which is not handled in the prior art. Compared with the traditional method, the collaborative constraint relationship between optical properties and texture fidelity is established, which effectively avoids the misjudgment scene of "material light transmission qualified but ingredient blurred" or "texture clear but refraction distorted". The correlation of physical properties of food images is converted into a mathematical constraint model, and the efficiency of dynamic evaluation of translucent materials is improved.

[0093] S6, according to the optical consistency score of the candidate material and the user interaction heat data, the candidate material is pushed to the target user and the marketing incentive strategy is associated, and the specific implementation is:

[0094] The candidate materials are divided into high score section, medium score section and low score section according to the optical consistency score, and the partition boundary is dynamically adjusted according to the statistical distribution characteristics of the historical score samples. The specific implementation method is: continuously collect the optical consistency score data of all candidate materials in the historical period, and construct a score value distribution histogram. The partition boundary determination principle is based on the statistical quartile method: the lower limit value of the high score section is the seventy-fifth percentile value of the score distribution, the lower limit value of the medium score section is the twenty-fifth percentile value, and the low score section includes all materials lower than the twenty-fifth percentile value. The core of the dynamic adjustment mechanism is: the percentile values of the historical score samples are recalculated every week, and the boundary is updated when the difference between the new percentile and the current boundary exceeds the allowed fluctuation threshold. The allowed fluctuation threshold is set to 0.5 times the score standard deviation, and the standard deviation is obtained by calculating the historical samples. After the partition section is updated, the boundary smoothing processing is performed: the moving average algorithm is used to eliminate the influence of short-term fluctuations, and the moving window length is four adjustment periods, each period is seven days.

[0095] The material preference feature of the target user is extracted in the user interaction heat data, and the push priority order of the material in each partition is adjusted through a heat correction factor. The extraction method of the material preference feature is: analyzing the complete interaction behavior log of the target user in the last thirty days, and the interaction behavior dimensions include click operation, stay time operation and sharing operation. The feature extraction process includes: calculating the click rate deviation value of the user to the ice powder content, the deviation value calculation formula is the user click rate divided by the platform user average click rate; the average stay time ratio of the user watching the ice powder video is counted, the ratio value is the user stay time divided by the average stay time of the same type of video; the frequency coefficient of the user sharing the ice powder content is recorded, the coefficient value is the user sharing times divided by the content receiving times. The generation method of the heat correction factor is: the three feature values are normalized by minimum and maximum, and then weighted sum is calculated according to the weight coefficient, the weight distribution is that the click rate deviation weight is zero point four, the stay time ratio weight is zero point three, and the sharing frequency coefficient weight is zero point three, and the weight value is determined through historical three-month conversion data regression analysis. The priority adjustment rule is: on the basis of the original sorting within the partition, the sorting position of each material is multiplied by the heat correction factor to obtain a new position sequence number, and finally the corrected queue is generated by reordering according to the new sequence number in ascending order.

[0096] The mapping rule of the partition and the marketing incentive strategy is established, the high partition is associated with the brand exposure type incentive strategy, the medium partition is associated with the conversion guide type incentive strategy, and the low partition is associated with the user feedback type incentive strategy. The specific implementation content of the brand exposure type incentive strategy is: when the user watches the video content completely, the brand trademark display component is triggered and the integral reward system is activated, and the integral number is set according to the user historical activity level. The execution mode of the conversion guide type incentive strategy is: a time-limited coupon floating window component is popped up when the video is played to the fifteenth second, the user clicks to jump to the purchase page and automatically deducts the corresponding amount. The operation mechanism of the user feedback type incentive strategy is: a feedback questionnaire component is displayed when the video is played, and the lottery wheel component is activated after the user completes more than three questionnaire evaluations. The activation control of the mapping rule is: the high partition material is bound with the brand integral distribution component, the medium partition material is associated with the coupon generation component, and the low partition material is connected with the questionnaire distribution component, and the binding relationship is stored in the strategy mapping relationship database.

[0097] According to the modified push priority, the candidate material is pushed to the target user, and the associated marketing incentive strategy execution component is activated synchronously. The push execution process is: a layered and progressive push strategy is adopted, the first push is the top three materials of the high segment priority, if the user does not interact within two hours, the top two materials of the medium segment are added, and if there is no interaction within twenty-four hours, the first material of the low segment is added. The push quantity control mechanism is: based on the user receiving capacity model, the maximum push quantity of the day is calculated, the capacity model input parameters include the user's seven-day average activity, historical opening rate and device performance score, and the output is an integer push quantity. The synchronous activation of the marketing incentive strategy is realized through the event bus mechanism: the video playback state change event triggers the strategy detector, and when the preset condition is met, the corresponding executor component is called. The whole process data closed loop mechanism is: all user interaction behaviors are written into the interaction log database in real time, and the log data is used to update the heat correction factor calculation model every day.

[0098] The design of step S6 aims at the problem that the quality evaluation of UGC material is not consistent with user preferences, and solves the defect that the existing push system does not consider both content quality and user interest by fusing optical quality grading and user behavior heat through a dynamic partition mechanism. Compared with the traditional method, the heat correction factor is established to dynamically adjust the push priority, and the quality layer is matched to combine the differentiated marketing strategy, which effectively avoids the burying of high-quality low-heat materials and the invalid incentive of low-quality materials, and three-dimensionally couples the content physical property evaluation, user behavior analysis and business target realization, thereby providing a systematic solution for food UGC precise marketing.

[0099] The technical solutions of steps S1 to S6 realize cross-field technology fusion: the physical and optical characteristics (refraction disturbance tensor) of the semi-transparent material of the ice powder and the fluid motion characteristics (texture vortex) of the ingredients are converted into a calculable model, and a dynamic coupling mechanism of the optical characteristics and user behavior data is established. The space-time continuity characteristics of the material refraction trajectory can represent the freshness of the ice powder, and the ingredient vorticity distribution can reflect the production process level. According to the three-level linkage mechanism of “physical characteristics-quality score-user response” constructed accordingly, the problem of the separation of quality evaluation and user preference in the UGC scene is solved. The quality of the food image is quantitatively evaluated based on the optical physical principle, so that the system can simultaneously optimize the content quality and the user touch efficiency.

[0100] Embodiment 2: Figure 2 A structure diagram of a popularization system for ice powder marketing is given, and the popularization system for ice powder marketing comprises:

[0101] The material sampling module is used for collecting the material refraction disturbance basic parameters and the ingredient texture distribution basic parameters of the ice powder, and generating a multi-modal feature template;

[0102] The model building evaluation module is configured to construct an optical consistency evaluation model of the ice powder UGC material according to a multi-modal feature template, and the optical consistency evaluation model comprises a refraction gradient matching rule and a texture clustering rule.

[0103] The dynamic feature extraction module is configured to extract a material refraction disturbance tensor and an ingredient texture vortex feature of continuous frames from the ice powder UGC material uploaded by the user.

[0104] The dual-matching module is configured to generate a refraction deviation coefficient by dynamically matching the material refraction disturbance tensor based on the refraction gradient matching rule, and generate a texture matching degree by vortex core density analysis on the ingredient texture vortex feature based on the texture clustering rule.

[0105] The evaluation and screening module is configured to generate an optical consistency score according to the refraction deviation coefficient and the texture matching degree, and screen candidate materials based on the optical consistency score.

[0106] The strategy pushing and executing module is configured to push the candidate materials to target users and associate marketing incentive strategies according to the optical consistency score of the candidate materials and user interaction heat data.

[0107] The material sampling module collects the ice powder refraction angle and the ingredient texture distribution in a standard lighting environment to generate a multi-modal feature template; the model building evaluation module constructs an optical consistency evaluation model based on the template, which comprises a refraction gradient matching rule that quantifies material deviation through a dynamic optical correction function library, and a texture clustering rule that analyzes vortex distribution characteristics through a decision tree. The dynamic feature extraction module extracts a material refraction disturbance tensor (spatiotemporal three-dimensional structure) and an ingredient texture vortex feature (vorticity matrix) of continuous frames from the user video to provide structured input for the dual-matching module. The dual-matching module performs dynamic time alignment to generate a refraction deviation coefficient, and simultaneously outputs a texture matching degree through vortex core density analysis, both of which are parallel parameter inputs to the evaluation and screening module. The evaluation and screening module maps the two parameters into an optical consistency score through a product coupling model, and screens a candidate material set based on a dual-threshold mechanism. The strategy pushing and executing module associates differentiated marketing strategies according to score partitions, dynamically corrects the pushing priority through user interaction heat, and finally realizes precise reach of the candidate materials and synchronous activation of the incentive strategies.

[0108] In the embodiments, all calculations are dimensionless numerical calculations, and the preset parameters and threshold values in the calculations are set by a person skilled in the art according to actual conditions.

[0109] It should be noted that the application can be deployed on a device itself to realize embedded application, or run on a PC terminal or other terminal with a user interface, thereby meeting various hardware environments and use requirements.

[0110] The above-described embodiments can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented by software, the above-described embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in the embodiments of the present application are wholly or partially generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transferred from one website, computer, server, or data center to another website, computer, server, or data center through a wired (for example, infrared, wireless, microwave, etc.) manner. The computer-readable storage medium can be any available medium accessible by a computer or a data storage device such as a server, data center, etc. containing a set of one or more available media. The available medium can be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state disk.

[0111] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described system, device and module can refer to the corresponding processes in the foregoing method embodiments, which will not be described here.

[0112] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other ways. For example, the above-described device embodiments are only schematic, for example, the division of the modules is only a logical function division, and actual implementation can have another division manner, for example, a plurality of modules or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed each other can be indirect coupling or communication connection through some interfaces, devices or modules, which can be electrical, mechanical or other forms.

[0113] The modules described as separate components can or can not be physically separated, and the components displayed as modules can or can not be physical modules, which can be located in one place or distributed on a plurality of network modules. Part or all of the modules can be selected according to actual needs to achieve the purpose of the embodiments.

[0114] In addition, each functional module in the various embodiments of the present application can be integrated in one processing module, or each module can exist physically independently, or two or more modules can be integrated in one module.

[0115] If the functions are implemented in the form of software functional modules and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application can be embodied in the form of a software product, and the computer software product is stored in a storage medium, and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0116] The above is merely specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0117] Finally: the above is merely preferred embodiments of the present application, and is not used to limit the present application, and any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A promotional method for the marketing of ice jelly, characterized in that, include: S1. Collect the basic parameters of material refraction disturbance and ingredient texture distribution of ice powder, and generate a multimodal feature template; S2. Construct an optical consistency evaluation model for ice powder UGC materials based on multimodal feature templates. The optical consistency evaluation model includes refraction gradient matching rules and texture clustering rules. S3. Extract the material refraction perturbation tensor and ingredient texture vortex features of consecutive frames from user-uploaded ice jelly UGC materials, including: Dynamic optical flow field analysis is performed on user-uploaded ice jelly UGC video materials to generate continuous frame material refraction light change trajectories; The refraction angle offset difference matrix between adjacent frames is calculated based on the material refraction light curve trajectory, and a three-dimensional material refraction perturbation tensor is constructed by temporal stacking. Motion trajectory tracking is performed on the ingredients particles in consecutive frames to generate an ingredient displacement vector field; Local rotating flow patterns are detected in the batching displacement vector field, and the radius of curvature and angular velocity distribution of the core region of the rotating flow are extracted. The inverse sequence of the radius of curvature is fused with the angular velocity distribution matrix to form the vortex feature of the ingredient texture; S4. Based on the refraction gradient matching rule, the material refraction perturbation tensor is dynamically matched to generate the refraction deviation coefficient, and based on the texture clustering rule, the vortex core density of the ingredient texture vortex feature is analyzed to generate the texture matching degree. S5. Generate an optical consistency score based on the refractive deviation coefficient and texture matching degree, and select candidate materials based on the optical consistency score; S6. Based on the optical consistency score and user interaction popularity data of the candidate materials, push the candidate materials to the target users and associate them with marketing incentive strategies.

2. The promotion method for marketing ice jelly according to claim 1, characterized in that, Collect the basic parameters of material refraction perturbation and ingredient texture distribution of ice jelly, and generate a multimodal feature template, including: The optical parameters of the translucent ice powder material under standard lighting conditions were collected as the basic parameters of material refraction perturbation. At the same time, the texture morphology parameters of the ice powder ingredients under static distribution were collected as the basic parameters of ingredient texture distribution. By quantifying the gradient sequence of refractive index change of the basic parameters of material refraction perturbation and the spatial topology of the basic parameters of ingredient texture distribution, a multimodal feature template that integrates optical and textural properties is generated.

3. The promotion method for marketing ice jelly according to claim 2, characterized in that, The basic parameters of material refraction perturbation include the refraction angle offset dataset of ice powder material in the visible spectrum range, and the basic parameters of ingredient texture distribution include the geometric contour features of ingredient particles and the color superposition effect dataset.

4. The promotion method for marketing ice jelly according to claim 1, characterized in that, An optical consistency evaluation model for ice powder UGC materials is constructed based on multimodal feature templates. The optical consistency evaluation model includes refraction gradient matching rules and texture clustering rules, including: Based on the refractive index change gradient sequence in the multimodal feature template, a refractive gradient matching rule is constructed. By establishing the mapping relationship between the gradient sequence and the refractive properties of the ice powder material under standard illumination conditions, a dynamic optical correction function library is generated. Based on the spatial topology index and color superposition effect dataset in the multimodal feature template, texture clustering rules are constructed. By defining the local aggregation degree threshold and global distribution uniformity boundary of the ingredients particles, a texture feature classification decision tree is generated. The dynamic optical correction function library and texture feature classification decision tree are integrated into the machine learning framework to train and generate an optical consistency evaluation model. The input layer dimension of the optical consistency evaluation model is consistent with the vector dimension of the multimodal feature template, and the output layer is the optical consistency score.

5. A promotional method for marketing ice jelly according to claim 1, characterized in that, Based on the refraction gradient matching rule, the material refraction perturbation tensor is dynamically matched to generate the refraction deviation coefficient, including: The material refractive perturbation tensor is decomposed into a continuous sequence of sub-tensors along the time dimension, and each sub-tensor is pattern-matched with the dynamic optical correction function library in the refractive gradient matching rule through a dynamic temporal alignment window. The residual norms of the subtensor feature channels and the reference modes of the dynamic optical correction function library are calculated. The weighted aggregated residual norms are used to generate the refractive deviation coefficients, and the weight values ​​are assigned according to the illumination stability of the subtensor on the time axis.

6. The promotion method for marketing ice jelly according to claim 1, characterized in that, Based on texture clustering rules, vortex core density analysis is performed on the vortex features of ingredient textures to generate texture matching degree, including: In the vortex features of the ingredient texture, select the effective regions where the vortex core density exceeds a preset density threshold, and extract the curl distribution matrix of the effective regions. Input the curl distribution matrix into the texture feature classification decision tree in the texture clustering rules, and traverse the decision tree node decision path; Texture matching degree is generated based on the probability distribution of leaf node affiliation in the determination path.

7. The promotion method for marketing ice jelly according to claim 1, characterized in that, An optical consistency score is generated based on the refractive deviation coefficient and texture matching degree, and candidate materials are selected based on the optical consistency score, including: A product relationship model between the refractive deviation coefficient and the texture matching degree is established. The product value is mapped to the optical consistency score through a nonlinear transformation function. The parameters of the nonlinear transformation function are determined by fitting the score distribution of historical samples. A dual-threshold screening mechanism is established by setting an upper limit for the refractive deviation coefficient and a lower limit for the texture matching degree. Only when the candidate material simultaneously satisfies that the refractive deviation coefficient does not exceed the upper limit and the texture matching degree is not lower than the lower limit will it be included in the candidate set. Candidate materials entering the candidate set are sorted in descending order of optical consistency score, and a preset number of materials with the highest ranking are selected as the final output.

8. A promotional method for marketing ice jelly according to claim 1, characterized in that, Based on the optical consistency score and user interaction popularity data of candidate materials, candidate materials are pushed to target users and associated with marketing incentive strategies, including: Candidate materials are divided into high, medium and low segments based on optical consistency scores, and the segment boundaries are dynamically adjusted according to the statistical distribution characteristics of historical score samples. Extract the target user's material preference characteristics from user interaction heat data, and adjust the push priority order of materials in each segment through heat correction factors; Establish mapping rules between different time zones and marketing incentive strategies: high-level time zones are associated with brand exposure incentive strategies, medium-level time zones are associated with conversion guidance incentive strategies, and low-level time zones are associated with user feedback incentive strategies. Candidate materials are pushed to target users based on the revised push priority, and the associated marketing incentive strategy execution components are activated simultaneously.

9. A promotion system for marketing ice jelly, used to implement the promotion method for marketing ice jelly as described in any one of claims 1-8, characterized in that, include: The material acquisition module is used to collect the basic parameters of material refraction disturbance and ingredient texture distribution of ice powder, and generate multimodal feature templates. The model building and evaluation module is used to construct an optical consistency evaluation model for ice powder UGC materials based on multimodal feature templates. The optical consistency evaluation model includes refractive gradient matching rules and texture clustering rules. The dynamic feature extraction module is used to extract the material refraction perturbation tensor and ingredient texture vortex features of consecutive frames from user-uploaded ice jelly UGC materials. The dual-analysis matching module is used to dynamically match the material refractive perturbation tensor based on the refractive gradient matching rule to generate the refractive deviation coefficient, and to perform vortex core density analysis on the vortex features of the ingredient texture based on the texture clustering rule to generate the texture matching degree. The candidate selection module is used to generate an optical consistency score based on the refractive deviation coefficient and texture matching degree, and to select candidate materials based on the optical consistency score; The strategy execution module is used to push candidate materials to target users and associate them with marketing incentive strategies based on the optical consistency score of candidate materials and user interaction popularity data.

Images