A digital creative advertisement design and order management method and system

Abstract

The application relates to the technical field of digital creative advertisement, and discloses a digital creative advertisement design and production and order management method and system, which generates audience attention time sequence trajectory graphs through an audience attention time sequence prediction model, performs space-time alignment analysis on brand element time sequence position constraint ranges and the audience attention time sequence trajectory graphs, generates attention-consistency joint time sequence constraint parameters, executes frame-by-frame video advertisement content generation guided by attention, and generates video advertisement creative files through two-dimensional evaluation and screening of brand element time sequence consistency scores and audience attention time sequence coverage rate scores, so that the directional contradiction problem of brand element interframe consistency constraint and target audience attention guidance in the time sequence dimension is solved.

Description

Technical Field

[0001] This invention relates to the field of digital creative advertising technology, and more specifically, to a method and system for digital creative advertising design, production, and order management. Background Technology

[0002] In the entire process of digital creative advertising design, production and order management, after receiving a client's video advertising order, the platform usually extracts brand characteristics based on the brand's historical materials, uses AI-assisted content generation module to generate video advertising creative content frame by frame, and ensures that brand elements remain stable between frames through brand temporal consistency constraints.

[0003] In existing technologies, brand temporal consistency constraint technology is used to ensure the stability of brand elements' position and state between frames, while universal visual quality assessment technology is used to score and screen the generated video advertising content based on universal aesthetic standards.

[0004] However, in video advertising orders targeting specific audience groups, the attention focus of different audience groups dynamically shifts over time during video playback. While brand elements may remain consistent across frames, they may consistently reside in areas of low attention for the target audience, leading to inefficient brand message delivery. Directly moving brand elements to the audience's attention hotspots could disrupt the frame-to-frame consistency constraint. Existing general visual quality assessments score static images based on universal aesthetic standards, failing to recognize differences in visual attention among different audience groups or handle changes in attention trajectories over video time. This results in ad creatives passing general scoring but failing to achieve the desired actual perception for the target audience, increasing the manual modification costs in subsequent collaborative editing stages and reducing order fulfillment efficiency. Summary of the Invention

[0005] This invention provides a method and system for digital creative advertising design, production, and order management, which solves the technical problems in related technologies such as the mismatch between the presentation position of brand elements and the distribution of target audience attention during the video advertising generation process, the difficulty in ensuring the temporal consistency of brand elements, and the lack of quantitative assessment of audience attention coverage.

[0006] This invention discloses a method for digital creative advertising design, production, and order management, comprising the following steps: acquiring video advertising order data, parsing and extracting brand identifiers, target audience tags, and keyframe instruction sequences of content scripts, and generating structured order requirement specifications;

[0007] Based on the brand identity, historical video advertising materials of the brand are retrieved from the material library. The permissible position range of brand elements in the frame sequence, the permissible offset threshold between frames, and the motion effect transition style parameters are extracted to generate a set of brand temporal gene feature vectors.

[0008] Based on the target audience tags, load the corresponding visual attention distribution pattern parameters and attention temporal decay curve parameters, and inject them into the audience attention temporal prediction model to complete the model configuration;

[0009] The keyframe instruction sequence of the content script is input into the audience attention time-series prediction model to obtain the audience attention heatmap of each frame, and aggregated along the time axis to generate the audience attention time-series trajectory map.

[0010] The temporal position constraint range of brand elements in the brand temporal gene feature vector set is spatiotemporally aligned with the audience attention temporal trajectory map. The spatial overlap between the allowable position area of ​​brand elements and the attention hot zone is calculated frame by frame. Frames with spatial overlap below a preset overlap threshold are marked as attention guidance intervention frames and attention perception position adjustment vectors are calculated to generate attention-consistency joint temporal constraint parameters.

[0011] Based on the brand temporal gene feature vector set and the attention-consistency joint temporal constraint parameters, brand static constraint verification and temporal consistency constraint verification are performed on non-intervention frames, and attention perception position adjustment guidance and relaxed consistency constraint verification are performed on attention-guided intervention frames to generate a temporal associated frame sequence of attention guidance.

[0012] The brand element temporal consistency score and attention temporal coverage score are calculated for the temporally associated frame sequence. Video candidate content with both scores exceeding the corresponding thresholds is selected and a video ad creative file is generated.

[0013] Furthermore, the audience attention temporal prediction model includes a spatial attention encoding unit, a temporal propagation unit, and an audience adaptation unit. The spatial attention encoding unit receives single-frame content description information and outputs the basic spatial attention distribution of a single frame. The temporal propagation unit receives the spatial attention distribution sequence of adjacent frames, propagates the attention state along the time axis, and outputs inter-frame attention transfer features. The audience adaptation unit receives the inter-frame attention transfer features, the visual attention distribution pattern parameters, and the attention temporal decay curve parameters, and outputs attention heatmaps for each frame that are adapted to the target audience. The spatial attention encoding unit uses a convolutional neural network, and the output layer uses a sigmoid activation function to normalize the attention intensity of each pixel position to the range of zero to one. The temporal propagation unit uses an LSTM neural network, which maps the hidden state vector to inter-frame attention transfer features through a linear mapping layer.

[0014] Furthermore, the audience adaptation unit employs a parameter condition injection method, using the visual attention distribution pattern parameter as the spatial dimension modulation coefficient and the attention temporal decay curve parameter as the temporal dimension modulation coefficient to perform spatial dimension modulation and temporal dimension modulation on the inter-frame attention transfer features, respectively. The spatial dimension modulation is achieved by multiplying the spatial distribution of the inter-frame attention transfer features pixel by pixel with the spatial weight map corresponding to the visual attention distribution pattern parameter. The temporal dimension modulation is achieved by applying an exponential decay function to the attention intensity value of the current frame. The output of the exponential decay function is the product of the initial attention intensity and the value of a natural exponential function with the decay rate parameter and the time variable as exponents.

[0015] Furthermore, the aggregation along the time axis to generate the audience attention time-series trajectory map includes: arranging the two-dimensional attention heatmaps of each frame into a three-dimensional data structure of two spatial dimensions plus one temporal dimension according to the frame time sequence; employing an inter-frame attention peak matching algorithm based on a spatial distance threshold, for each attention peak point in the current frame whose attention intensity exceeds a preset intensity threshold, searching for attention peak points in the next frame whose spatial distance is less than a preset spatial distance threshold; if they exist, establishing a trajectory association; if they do not exist, marking the trajectory as terminated in the current frame; and marking unassociated attention peak points in the next frame as the starting point of a new trajectory; the attention intensity change trend is obtained by performing time-series difference calculation on the attention intensity values ​​of corresponding peak points in each frame on the same trajectory; the preset intensity threshold is taken as 60% of the maximum attention intensity value in the attention heatmap of each frame, and the preset spatial distance threshold is taken as 5% of the screen width.

[0016] Further, the calculation of the attention perception position adjustment vector includes: determining the spatial position corresponding to the maximum attention intensity in the audience attention heatmap of the attention guidance intervention frame as the attention center point; calculating the unit direction vector pointing from the current center position of the brand element to the attention center point as the target direction; taking the inter-frame allowable offset threshold as the maximum allowable offset; the attention perception position adjustment vector is the product of the adjustment intensity coefficient, the maximum allowable offset, and the unit vector of the target direction, wherein the value of the adjustment intensity coefficient is greater than zero and less than or equal to one; wherein, the spatial overlap is the ratio of the intersection area of ​​the brand element's allowable position area and the attention heat area to the area of ​​the brand element's allowable position area, and the preset overlap threshold value is 0.3.

[0017] Furthermore, it also includes: calculating the cumulative offset of the attention perception position adjustment vector of each intervention frame within a segment of consecutive frames marked as attention-guided intervention frames along the time axis, wherein the cumulative offset is the magnitude of the vector sum of the adjustment vectors of each intervention frame within the segment; when the cumulative offset exceeds the global position offset upper limit defined in the brand temporal gene feature vector set, applying a scaling factor to the adjustment vectors of subsequent intervention frames in the segment to scale them proportionally, so that the cumulative offset converges to within the global position offset upper limit; wherein the scaling factor is the quotient obtained by dividing the difference between the global position offset upper limit and the magnitude of the vector sum of the adjustment vectors of each intervention frame before the current intervention frame by the magnitude of the adjustment vector of the current intervention frame; the global position offset upper limit is taken as 20% of the diagonal length of the allowable position area of ​​the brand element.

[0018] Furthermore, it also includes: linearly interpolating and extending the attention-aware position adjustment vector of each attention-guided intervention frame along the time axis to the preceding non-intervention frame and the subsequent non-intervention frame to generate a progressive adjustment vector for the transition frame; the number of transition frames is the rounded-up value of the ratio of the magnitude of the attention-aware position adjustment vector to the inter-frame allowable offset threshold; the progressive adjustment vector of the preceding transition frame is the ratio of the transition frame number to the number of transition frames multiplied by the attention-aware position adjustment vector, and the progressive adjustment vector of the subsequent transition frame is a product of the ratio of the transition frame number to the number of transition frames and the attention-aware position adjustment vector; the progressive adjustment vector of the transition frame is then added to the attention-consistency joint temporal constraint parameter.

[0019] Furthermore, in the relaxed consistency constraint verification, the relaxed threshold is the sum of the original inter-frame allowable offset threshold and the relaxed coefficient multiplied by the magnitude of the attention perception position adjustment vector, with the relaxed coefficient being 0.8; the brand element temporal consistency score is calculated by subtracting the weighted average of the normalized values ​​of brand element position offset, size change, and tone deviation between adjacent frames from one, with weighting coefficients of 0.5, 0.3, and 0.2 respectively, and the sum of the three being one; the attention temporal coverage score is the ratio of the sum of the audience attention intensity values ​​at the location of the brand key elements in each frame to the sum of the maximum audience attention intensity values ​​in each frame.

[0020] Furthermore, it also includes: generating an attention coverage visualization overlay for each frame in the video ad creative file, marking the location of the audience's attention hotspots with semi-transparent thermal color marks and marking the location of key brand elements with marker boxes, overlaying the attention coverage visualization overlay with the original frame content to generate a preview version with attention coverage visualization and attaching it to the video ad creative file; pushing the video ad creative file to the online collaborative editing module, recording collaborative modification operations and generating version snapshots and change logs through version control, and outputting the final delivery file; and generating a delivery acceptance report based on the final delivery file and the structured order requirement specification execution requirement compliance verification and technical specification verification, combined with the brand element temporal consistency score and attention temporal coverage rate score.

[0021] This invention provides a digital creative advertising design, production, and order management system, characterized in that it includes:

[0022] The order parsing module is used to acquire video ad order data, parse and extract brand identifiers, target audience tags and content script keyframe instruction sequences, and generate structured order requirement specifications.

[0023] The brand gene extraction module is used to retrieve historical video advertising materials of the brand from the material library based on the brand logo, extract the permissible position range of brand elements in the frame sequence, the permissible offset threshold between frames and the motion effect transition style parameters, and generate a brand temporal gene feature vector set.

[0024] The attention model configuration module is used to load the corresponding visual attention distribution pattern parameters and attention temporal decay curve parameters based on the target audience group labels, and inject them into the audience attention temporal prediction model to complete the model configuration.

[0025] The attention trajectory generation module is used to input the content script keyframe instruction sequence into the audience attention temporal prediction model, obtain the audience attention heatmap of each frame, and aggregate it along the time axis to generate the audience attention temporal trajectory map.

[0026] The spatiotemporal alignment analysis module is used to perform spatiotemporal alignment analysis between the temporal position constraint range of brand elements in the brand temporal gene feature vector set and the audience attention temporal trajectory map, calculate the spatial overlap on a frame-by-frame basis, mark frames with spatial overlap below a preset overlap threshold as attention guidance intervention frames and calculate the attention perception position adjustment vector, and generate attention-consistency joint temporal constraint parameters.

[0027] The content generation module is used to perform brand static constraint verification and temporal consistency constraint verification on non-intervention frames based on the brand temporal gene feature vector set and the attention-consistency joint temporal constraint parameters, and to perform attention perception position adjustment guidance and apply relaxed consistency constraint verification on attention-guided intervention frames to generate a temporal associated frame sequence guided by attention.

[0028] The dual-dimensional evaluation module is used to calculate the brand element temporal consistency score and attention temporal coverage score for the temporally associated frame sequence, filter video candidate content whose dual-dimensional scores both exceed the corresponding thresholds, and generate video ad creative files.

[0029] This invention obtains information on the dynamic migration of the target audience's attention focus along the video timeline by configuring an audience attention temporal prediction model and generating an audience attention temporal trajectory map. It calculates the spatial overlap between the permissible location area of ​​brand elements and the attention hotspot through spatiotemporal alignment analysis, marking frames with insufficient spatial overlap as attention-guiding intervention frames and calculating attention perception position adjustment vectors to generate attention-consistency joint temporal constraint parameters. During the frame-by-frame generation stage, the original static brand constraints and temporal consistency constraints are applied to non-intervention frames, while attention perception position adjustment guidance is performed on attention-guiding intervention frames, and relaxed consistency constraint verification is applied. By using a dual-dimensional evaluation of brand element temporal consistency score and audience attention temporal coverage score to screen video ad creative files, this invention solves the technical problem of directional contradictions between brand element frame consistency constraints and target audience attention guidance in the temporal dimension, and the inability of existing general visual quality assessments to perceive the differentiated attention distribution of the audience group in the temporal dimension. This achieves the technical effect of improving the efficiency of brand information delivery to the target audience while maintaining brand temporal consistency and reducing the cost of manual modification in the collaborative editing process. Attached Figure Description

[0030] Figure 1 This is a flowchart of the digital creative advertising design, production, and order management method provided in an embodiment of the present invention;

[0031] Figure 2 This is a schematic diagram of the keyframe instruction timing distribution and scene types provided in the embodiments of the present invention;

[0032] Figure 3 This is a schematic diagram illustrating the range of time-series genetic feature parameters of the brand logo provided in an embodiment of the present invention;

[0033] Figure 4 This is a schematic diagram of the target audience visual attention distribution weights provided in an embodiment of the present invention;

[0034] Figure 5 This is a schematic diagram of the temporal decay curve of audience attention intensity provided in an embodiment of the present invention;

[0035] Figure 6 This is a schematic diagram of the vertical position migration trajectory of the attention hotspot provided in an embodiment of the present invention;

[0036] Figure 7 This is a schematic diagram of spatiotemporal alignment analysis - spatial overlap variation provided in an embodiment of the present invention;

[0037] Figure 8 This is a schematic diagram comparing the effects of attention-guided intervention provided in an embodiment of the present invention. Detailed Implementation

[0038] In the entire process of digital creative advertising design, production, and order management, after receiving a client's video ad order, the platform typically extracts brand DNA features based on historical brand materials, uses an AI-assisted content generation module to generate video ad creative content frame by frame, and ensures that brand elements (such as logo position and main color areas) remain stable between frames through brand temporal consistency constraints. However, in video ad orders targeting specific audience groups, the attention focus of different audience groups dynamically shifts over time during video playback. Although brand elements remain consistent between frames, they may continuously be in the target audience's low attention zone, resulting in low brand message delivery efficiency. Directly moving brand elements to the audience's attention hotspot may violate the frame consistency constraint. At the same time, existing general visual quality assessments score static images based on universal aesthetic standards, failing to perceive differences in visual attention among different audience groups or handle changes in attention trajectory over time. This leads to ad creatives passing general scoring but poor actual perception by the target audience, thus increasing the manual modification costs in subsequent collaborative editing stages and reducing order fulfillment efficiency.

[0039] Therefore, the technical problem that this implementation method needs to solve is that in the process of AI-assisted generation of video ads, there is a directional contradiction between the constraints of inter-frame consistency of brand elements and the guidance of target audience attention in the temporal dimension, and the existing general visual quality assessment cannot perceive the differentiated attention distribution of the audience in the time dimension, resulting in a disconnect between the efficiency of brand information transmission of the generated content and the actual perception effect of the target audience.

[0040] According to an embodiment of this method, a digital creative advertising design, production, and order management method is provided. It should be understood that the executing entity of this embodiment is a digital creative advertising platform server configured with a demand analysis module, a material retrieval module, an AI-assisted content generation module, an online collaborative editing module, and a delivery and acceptance module. The digital creative advertising platform server possesses video frame sequence processing capabilities and model inference calculation capabilities.

[0041] At least one embodiment of the present invention discloses a method for digital creative advertising design, production, and order management, such as... Figure 1 As shown, it includes the following steps:

[0042] Step 1: Obtain video ad order data and parse it to generate structured order requirement specifications;

[0043] We acquire video ad order data submitted by clients, and extract brand identifiers, target audience tags, video duration specifications, and keyframe instruction sequences from the video ad order data through requirement parsing and processing to generate structured order requirement specifications. Specifically, the brand identifier is used to associate with historical brand materials in the platform's material library, the target audience tags are used to represent the target audience category for the ad campaign, and the keyframe instruction sequences from the content script are used to describe the visual content requirements at each key time point in the video.

[0044] Furthermore, the aforementioned content script keyframe instruction sequence refers to the ordered set of instructions obtained after timeline parsing of the video advertisement content script. Each instruction corresponds to a keyframe time point and its on-screen content description, scene transition marker, and camera motion parameters. Each instruction in the content script keyframe instruction sequence contains at least three pieces of information: a timestamp, an on-screen content text description, and a scene type label.

[0045] Step 2: Retrieve historical brand materials and extract the brand's temporal gene feature vector set;

[0046] Based on the brand identity in the structured order requirements specification, historical video ad creatives of the target brand are retrieved from the platform's creative library. Keyframe sequences and inter-frame transition features are extracted from these historical video ad creatives. The temporal position distribution patterns and motion effect style characteristics of brand elements in historical video ad creatives are analyzed to generate a brand temporal gene feature vector set. This set includes the permissible position range, size ratio range, permissible inter-frame offset threshold, and motion effect transition style parameters of key brand elements (such as brand logo, main color area, and standard font) in the frame sequence.

[0047] Furthermore, the aforementioned inter-frame allowable offset threshold refers to the maximum change in the position, size, and other attributes of brand elements between adjacent frames. The inter-frame allowable offset threshold is determined based on the statistical distribution of the actual changes in the inter-frame changes of brand elements in the target brand's historical video advertising materials. Specifically, the 95th percentile value of the distribution of the inter-frame position offset of brand elements in historical video advertising materials is taken as the inter-frame allowable offset threshold.

[0048] Step 3: Load audience attention parameters and configure the audience attention time-series prediction model;

[0049] Based on the target audience tags in the structured order requirement specifications, the visual attention distribution pattern parameters and attention time decay curve parameters of the corresponding group are loaded from the pre-stored audience attention parameter library. The visual attention distribution pattern parameters and attention time decay curve parameters are then injected into the corresponding parameter layer of the audience attention time prediction model to complete the configuration of the audience attention time prediction model.

[0050] The audience attention temporal prediction model is a learning model based on a temporal attention prediction network. It comprises three units: a spatial attention encoding unit, a temporal propagation unit, and an audience adaptation unit. The spatial attention encoding unit receives content description information from a single frame and outputs the basic spatial attention distribution of that frame. The temporal propagation unit receives the spatial attention distribution sequence of adjacent frames, propagates the attention state along the time axis, and outputs inter-frame attention transfer features. The audience adaptation unit receives the inter-frame attention transfer features, the currently loaded visual attention distribution pattern parameters, and the attention temporal decay curve parameters, and outputs attention heatmaps for each frame adapted to the target audience.

[0051] Furthermore, the spatial attention encoding unit of the audience attention temporal prediction model employs a convolutional neural network (CNN). The input to the CNN is a feature vector representing the content description information of a single frame, and the output is a two-dimensional heatmap of the spatial attention baseline distribution of that single frame. The CNN contains multiple convolutional and pooling layers, and finally, an upsampling layer restores the spatial resolution to the same level as the input frame. The output layer uses a sigmoid activation function to normalize the attention intensity at each pixel location. Interval.

[0052] Furthermore, the temporal propagation unit of the audience attention temporal prediction model employs an LSTM neural network. The input to the LSTM neural network is the spatial attention basis distribution sequence of adjacent frames, and the output is the hidden state vector of the inter-frame attention transfer feature. This hidden state vector is mapped to the inter-frame attention transfer feature through a linear mapping layer. The weight matrix dimension of the linear mapping layer is... ,in Let the dimension of the LSTM hidden state vector be . is the target dimension for inter-frame attention transfer features.

[0053] Furthermore, the audience adaptation unit employs a fully connected neural network. The input to the fully connected neural network is a concatenated vector of inter-frame attention transfer features, visual attention distribution pattern parameters, and attention temporal decay curve parameters. The output is an attention heatmap for each frame tailored to the target audience. The fully connected neural network contains two hidden layers, with ReLU activation. The output layer is mapped to a two-dimensional heatmap corresponding to the image spatial resolution via the fully connected layer. The output is normalized so that the sum of attention intensities at each position is 1. The audience attention temporal prediction model is trained using supervised learning. The training data consists of video frame sequences labeled with real audience eye-tracking heatmaps. The loss function is the mean squared error between the predicted attention heatmap and the real eye-tracking heatmap, and the optimization strategy uses the Adam optimization algorithm.

[0054] Furthermore, the aforementioned visual attention distribution pattern parameters refer to spatial attention preference parameters obtained from eye-tracking experimental data or behavioral data of a specific audience group. Different audience groups have different weights in allocating attention to different areas of the screen (such as the central area, text area, and facial area). The aforementioned attention temporal decay curve parameters refer to the decay rate and decay pattern parameters of the target audience's attention intensity at the same spatial location during video playback as the duration decreases.

[0055] Furthermore, to adapt to the varying granularities of attention patterns among different audience groups, the audience adaptation unit employs a parameter conditional injection approach. Visual attention distribution pattern parameters are used as modulation coefficients in the spatial dimension, and attention temporal decay curve parameters are used as modulation coefficients in the temporal dimension. These modulate the inter-frame attention transfer features output by the temporal propagation unit in both spatial and temporal dimensions, thus adapting to different audience groups without retraining the model. Spatial dimension modulation is achieved through element-wise multiplication, multiplying the spatial distribution of the inter-frame attention transfer features pixel-by-pixel with the spatial weight map corresponding to the visual attention distribution pattern parameters. Temporal dimension modulation is achieved by applying a decay function defined by the attention temporal decay curve parameters to the attention intensity value of the current frame. This decay function is an exponential decay function. ,in For the time elapsed The intensity of attention afterward Initial attention intensity, For decay rate parameters, For time variables, This represents an exponential function.

[0056] Step 4: Generate a temporal trajectory map of audience attention;

[0057] The keyframe instruction sequence of the content script is input into the configured audience attention temporal prediction model. Attention heatmap prediction from the target audience's perspective is performed on each frame to obtain the corresponding audience attention heatmap. The audience attention heatmaps of each frame are aggregated along the time axis to generate an audience attention temporal trajectory map. The audience attention temporal trajectory map marks the locations of attention hotspots and the trends in attention intensity changes for each time period. The locations of attention hotspots represent the spatial areas in each frame where the audience's attention concentration exceeds a preset intensity threshold, and the trends in attention intensity change represent the direction and rate of change of the intensity value of each attention hotspot along the time axis.

[0058] Furthermore, the aforementioned aggregation of audience attention heatmaps along the time axis refers to arranging the two-dimensional attention heatmaps of each frame into a three-dimensional data structure (two-dimensional space plus one-dimensional time) according to the frame time order. In the time dimension, it correlates and tracks the spatially proximate high-value attention regions in adjacent frames, forming a continuous motion trajectory of the attention heatmaps along the time axis. An inter-frame attention peak matching algorithm based on a spatial distance threshold is employed: for the first... Attention peaks in a frame whose attention intensity exceeds a preset intensity threshold are, in the th... The system searches for attention peaks within a frame whose spatial distance is less than a preset spatial distance threshold. If such peaks exist, a trajectory association is established; otherwise, the trajectory is marked as being in the [frame name missing]. Frame termination; for the first Unassociated attention peaks in a frame are marked as the starting points of new trajectories. The trend of attention intensity change is obtained by time-series differencing of the attention intensity values ​​at corresponding peaks in each frame along the same trajectory; a positive difference indicates an increase in attention intensity, while a negative difference indicates a decrease. This is the frame number.

[0059] Furthermore, the aforementioned preset intensity threshold is taken as 60% of the maximum attention intensity in the attention heatmap of each frame, and the aforementioned preset spatial distance threshold is taken as 5% of the screen width.

[0060] Step 5: Perform spatiotemporal alignment analysis to generate attention-consistency joint temporal constraint parameters;

[0061] The spatiotemporal alignment analysis is performed between the temporal position constraints of key brand elements in the brand's temporal gene feature vector set and the audience attention temporal trajectory map. The spatiotemporal alignment analysis includes the following sub-steps:

[0062] Step 501: For each frame, extract the allowable location area of ​​the brand key elements in the frame from the brand temporal gene feature vector set, and at the same time extract the attention hot zone location of the frame from the audience attention temporal trajectory map, and calculate the spatial overlap between the allowable location area of ​​the brand elements and the attention hot zone in each frame.

[0063] Step 502: For frames with spatial overlap below a preset overlap threshold, mark them as attention-guided intervention frames.

[0064] Step 503: For each attention guidance intervention frame, based on the trend of attention intensity change in the corresponding time period of the attention guidance intervention frame in the audience attention time sequence trajectory map, determine the target direction in which the brand element should be offset; based on the inter-frame allowable offset threshold in the brand time sequence gene feature vector set, determine the maximum allowable offset of the brand element in the attention guidance intervention frame; combine the target direction and the maximum allowable offset to calculate the attention perception position adjustment vector of the attention guidance intervention frame.

[0065] Furthermore, the method for determining the target direction is as follows: Find the spatial location corresponding to the maximum attention intensity in the audience attention heatmap of the attention guidance intervention frame as the attention center point, and calculate the unit direction vector pointing from the current center position of the brand element to the attention center point as the target direction. The maximum permissible offset is directly taken as the inter-frame permissible offset threshold defined in the brand temporal gene feature vector set. Attention perception position adjustment vector. The calculation formula is:

[0066] in, To adjust the strength coefficient, the range of values ​​is: This is used to control the ratio of the actual adjustment range to the maximum allowable offset. The original inter-frame allowable offset threshold defined for the brand temporal gene feature vector set. is the unit vector in the direction of the target.

[0067] Step 504: Combine the intervention frame label of each frame, the attention perception position adjustment vector of each attention-guided intervention frame, and the original temporal position constraint range used by the non-intervention frames to generate attention-consistency joint temporal constraint parameters.

[0068] Furthermore, the aforementioned spatial overlap refers to the ratio of the area of ​​intersection between the permissible location of the brand element and the attention hotspot within the same frame to the area of ​​the permissible location of the brand element. When this ratio is lower than a preset overlap threshold, it indicates that even if the brand element is within its permissible location, it still lacks sufficient spatial overlap with the target audience's attention hotspot, resulting in insufficient coverage of the brand information's audience attention. Spatial Overlap The calculation formula is:

[0069] in, Permitted areas for brand elements As a hotspot for attention, Indicates the area of ​​the region. This indicates the operation of region intersection.

[0070] Furthermore, the aforementioned preset overlap threshold is set to 0.3, meaning that when the intersection area of ​​the brand element's permissible location area and the attention hotspot is less than 30% of the area of ​​the brand element's permissible location area, the frame is marked as an attention guidance intervention frame.

[0071] Furthermore, to avoid the cumulative offset caused by intervention adjustments across multiple consecutive frames leading to deviations of brand element positions from the global constraints of brand genetic features over a longer time span, in addition to step 503, the following steps are included: For consecutive frames marked as attention-guided intervention frames along the time axis, calculate the cumulative offset of the attention-perceived position adjustment vector for each intervention frame within the frame segment. When the cumulative offset exceeds the global position offset upper limit defined in the brand temporal genetic feature vector set, scale the adjustment vectors of subsequent intervention frames in the frame segment proportionally to bring the cumulative offset within the global position offset upper limit. (Cumulative offset) The magnitude of the vector sum of the adjustment vectors of each intervention frame within the frame segment is calculated as follows: ,in The number of intervention frames within a frame segment. For the first Attention-aware position adjustment vector for each intervention frame To intervene in the sequence number of the frame within the frame segment, Represents the magnitude of the vector. When At that time, for the first frame segment frame( The adjustment vector applies a scaling factor. The scaled adjustment vector is ,in This is the upper limit of the global position offset. This is the sequence number of the current intervention frame within the frame segment.

[0072] Furthermore, the aforementioned global position offset upper limit The value is 20% of the diagonal length of the permissible position area of ​​the brand element.

[0073] Furthermore, to create a smooth transition in brand element positions before and after the attention-guided intervention frames, rather than abrupt jumps, the following steps are added to step 504: The attention-aware position adjustment vector of each attention-guided intervention frame is linearly interpolated and extended along the time axis to the preceding and following non-intervention frames, generating a progressive adjustment vector for the transition frames. This progressive adjustment vector is then added to the attention-consistency joint temporal constraint parameters. The number of transition frames is determined based on the ratio of the inter-frame allowable offset threshold to the magnitude of the attention-aware position adjustment vector. Calculated as ,in This indicates rounding up. (This refers to rounding up the integer part of the data before and after the intervention frame.) Frame, number The progressive adjustment vector for each transition frame Calculated via linear interpolation:

[0074] The preceding transition frame is The subsequent transition frame is ,in This is the transition frame number, and its value range is... arrive .

[0075] Step 6: Perform attention-guided frame-by-frame video ad content generation;

[0076] Based on the brand's temporal gene feature vector set and attention-consistency joint temporal constraint parameters, an AI-assisted content generation module performs frame-by-frame video ad content generation, generating attention-guided temporally related frame sequences. The frame-by-frame video ad content generation process includes the following sub-steps:

[0077] Step 601: For the first frame, perform brand gene constraint generation based on the permissible position range, size ratio range, and animation style parameters of the brand elements corresponding to the first frame in the brand temporal gene feature vector set, and output the anchor frame content. The position and state of the brand elements in the anchor frame content serve as the initial reference benchmark for the generation of subsequent frames.

[0078] Step 602: For each subsequent frame, the position and state of the brand elements in the previous frame, as well as the constraint information corresponding to each subsequent frame in the attention-consistency joint temporal constraint parameters, are simultaneously injected as conditions into the generation process of the AI-assisted content generation module.

[0079] Step 603: For frames marked as non-intervention frames in the attention-consistency joint temporal constraint parameters, perform brand static constraint verification and temporal consistency constraint verification. The brand static constraint verification determines whether the position and size of the brand elements in the non-intervention frame are within the allowable range defined by the brand temporal gene feature vector set; the temporal consistency constraint verification determines whether the position offset and state change of the brand elements in the non-intervention frame relative to the preceding frame are within the allowable inter-frame offset threshold. If both verifications pass, the non-intervention frame content passes the generation verification.

[0080] Furthermore, the static brand constraint verification is achieved by determining whether the center coordinates of the brand element fall within the allowable position area and whether the size of the brand element is within the size ratio range. The temporal consistency constraint verification is achieved by calculating the Euclidean distance between the center coordinates of the brand element in the current frame and the center coordinates of the brand element in the previous frame, and determining whether the Euclidean distance is less than or equal to the inter-frame allowable offset threshold.

[0081] Step 604: For frames marked as attention-guided intervention frames in the attention-consistency joint temporal constraint parameters, guide the offset of the brand element position based on the attention-perceived position adjustment vector of the attention-guided intervention frame, and perform attention-perceived position adjustment guidance. At the same time, apply a relaxed consistency constraint check to the attention-guided intervention frame, that is, replace the inter-frame allowable offset threshold in the temporal consistency constraint check with a relaxed threshold determined by the sum of the magnitude of the attention-perceived position adjustment vector and the original inter-frame allowable offset threshold.

[0082] Step 605: Concatenate the generated content of each frame according to the frame time order, and output the attention-guided temporally related frame sequence.

[0083] Furthermore, the aforementioned relaxed consistency constraint check means that brand elements are allowed a larger positional shift between frames compared to non-interventional frames for attention-guided intervention frames. However, this relaxation is constrained by the attention-perceived position adjustment vector and is not unlimited. The formula for calculating the relaxation threshold is:

[0084] in, This is the relaxed inter-frame allowable offset threshold. The original inter-frame allowable offset threshold defined for the brand temporal gene feature vector set. Adjust the attention-aware position vector for the attention-guided intervention frame. Adjust the magnitude of the vector for attention-aware location. This is a relaxation factor, and its value range is... This is used to control the ratio of the relaxation magnitude to the adjustment vector magnitude.

[0085] Furthermore, the aforementioned relaxation coefficients The value is 0.8.

[0086] Step 7: Calculate the two-dimensional score and filter to generate video ad creative files;

[0087] A two-dimensional scoring and filtering process is performed on the temporally correlated frame sequence guided by attention, generating video ad creative files with annotations on the attention-guided effect. The two-dimensional scoring and filtering process includes the following sub-steps:

[0088] Step 701: Calculate the temporal consistency score of brand elements in the attention-guided temporally related frame sequence along the time axis. The temporal consistency score is determined based on the weighted average of the brand element position offset, size change, and tone deviation between adjacent frames. The smaller the position offset and size change, the higher the temporal consistency score.

[0089] Furthermore, time-series consistency scoring The calculation formula is:

[0090] in, This represents the total number of frames in the video. The frame number, For the first Frame and the The normalized Euclidean distance between the center positions of brand elements in frames. For the first Frame and the The normalized value of the change in brand element size between frames. For the first Frame and the The normalized value of the color difference of the main color tone of the brand elements between frames. , , These are the weighting coefficients for positional offset, size change, and hue deviation, respectively. .

[0091] Furthermore, the above weighting coefficients are taken as follows: , , .

[0092] Step 702: Calculate the attention time-series coverage score of key brand information based on the audience attention time-series trajectory map. The attention time-series coverage score is determined by the ratio of the integral of the audience attention intensity value at the location of the key brand element in each frame along the time axis to the integral of the attention intensity throughout the entire video. The higher the proportion of time that the brand element is in the attention hot zone, the higher the attention time-series coverage score.

[0093] Furthermore, the formula for calculating the attention temporal coverage score is as follows:

[0094] in, Score attention temporal coverage. This represents the total number of frames in the video. The frame number, For the first Location of key brand elements in the frame The audience attention intensity value at that location For the first The spatial placement of key brand elements within the frame. For the first The maximum intensity of audience attention within a frame. Attention temporal coverage score reflects the actual extent to which brand elements capture the target audience's attention resources throughout the entire video.

[0095] Step 703: Filter video candidates whose brand element temporal consistency score and attention temporal coverage score both exceed the corresponding thresholds using template matching rules. Add attention guidance effect annotation information to the selected video candidates and generate a video ad creative file. The attention guidance effect annotation information includes the position adjustment amount of each attention guidance intervention frame, the change in attention coverage before and after the adjustment, and the change in consistency score.

[0096] Furthermore, among the aforementioned thresholds, the brand element temporal consistency score threshold is 0.75, and the attention temporal coverage score threshold is 0.60.

[0097] Furthermore, to provide more intuitive information on attention guidance effects in the video ad creative file to assist decision-making in subsequent collaborative editing, the following steps are included in addition to step 703: An attention overlay visualization layer is generated for each frame in the video ad creative file; the audience's attention hotspots are marked with semi-transparent heatmaps in the attention overlay visualization layer, and key brand elements are marked with marker boxes; the attention overlay visualization layer is superimposed on the original frame content to generate a preview version with attention overlay visualization and is attached to the video ad creative file. The heatmaps use a gradient color mapping from blue (low attention intensity) to red (high attention intensity), and the transparency of the attention overlay visualization layer is set to 50% to maintain the visibility of the original frame content.

[0098] Step 8: Push to the collaborative editing module and output the final delivery file;

[0099] Push video ad creative files to the online collaborative editing module to record collaborative modification operations initiated by clients and designers through multiple devices. Version control generates version snapshots and change logs for each collaborative modification operation. When the collaborative modification operation is completed, the final version delivery file is output.

[0100] Step 9: Perform delivery acceptance verification and generate a delivery acceptance report;

[0101] Based on the final delivery document and the structured order requirements specification, requirements compliance verification and technical specification verification are performed. Requirements compliance verification determines the degree of match between the video duration, screen resolution, and content theme of the final delivery document and the corresponding fields in the structured order requirements specification. Technical specification verification determines whether the encoding format, color space, and file size of the final delivery document comply with the platform's delivery technical standards. Combining brand element temporal consistency scoring and audience attention temporal coverage scoring, a delivery acceptance report is generated, triggering an order status update to the acceptance and settlement stage.

[0102] This implementation addresses the directional contradiction between brand inter-frame consistency constraints and audience attention guidance in the temporal dimension of video advertising orders targeting specific audiences by employing technical means of audience attention temporal prediction and attention-consistency joint temporal constraints.

[0103] Because in steps 3 and 4, by loading the differentiated visual attention distribution pattern parameters and attention temporal decay curve parameters of the target audience, the audience attention temporal prediction model is configured, and frame-by-frame attention heatmap prediction and time axis aggregation are performed on the key frame instruction sequence of the content script, an audience attention temporal trajectory map is generated. This enables the AI-assisted content generation process to obtain information on the dynamic migration of the target audience's attention focus along the video time axis, thereby overcoming the limitations of existing general visual quality assessments that cannot perceive the differentiated attention distribution of the audience and cannot handle the temporal dimension of attention changes.

[0104] In step 5, by performing spatiotemporal alignment analysis between the brand element temporal position constraint range and the audience attention temporal trajectory map, the spatial overlap between the brand element's permissible position area and the attention hot zone is calculated frame by frame. Frames with insufficient spatial overlap are marked as attention guidance intervention frames. An attention perception position adjustment vector that takes into account both the attention guidance direction and the permissible offset limit between frames is calculated. This locates the specific frame position where the brand consistency constraint and the audience attention hot zone do not match, and generates quantifiable joint constraint parameters, thus overcoming the problem of the lack of coordination mechanism between brand consistency constraints and attention guidance.

[0105] Because in step 6, the original static brand constraints and temporal consistency constraints are used for non-intervention frames, while attention perception position adjustment guidance and relaxed consistency constraint verification are applied to attention-guided intervention frames. This allows brand elements to be shifted to the target audience's attention hotspot in the attention-guided intervention frames to a limited extent without significantly disrupting inter-frame consistency. Thus, both brand temporal consistency and audience attention coverage are taken into account in the frame-by-frame generation stage.

[0106] Because in step 7, the dual-dimensional evaluation and joint screening of brand element temporal consistency score and audience attention temporal coverage score replaces the general visual quality assessment based solely on universal aesthetic standards, ensuring that the screened video ad creative files simultaneously meet the requirements of brand consistency and target audience attention coverage.

[0107] Therefore, this implementation method improves the efficiency of brand information delivery to the target audience during the AI-assisted content generation stage, reduces repeated modifications caused by poor perception by the target audience in the collaborative editing process, and accelerates the entire process of order fulfillment from creative generation to delivery and acceptance.

[0108] In March 20XX, an international sportswear brand submitted a video ad order to a digital creative advertising platform for its new youth basketball shoe collection. The order required the production of a 15-second vertical video ad for short video platforms, targeting basketball enthusiasts aged 13-18. The brand provided a package of materials including the brand logo, standard color scheme, and footage from three previous youth-oriented video ads, requiring consistency in brand visual identity while ensuring brand elements effectively attract the attention of the target youth audience.

[0109] like Figure 2-8 As shown, it includes the following steps:

[0110] The platform server receives order data packets submitted by customers and extracts brand identification codes, target audience tags, video technical specifications, and content script text from the order through the requirements parsing module. The content script describes the three key stages of the video: the opening basketball scene, the product close-up in the middle, and the brand logo display at the end. The system parses the script text into an ordered sequence containing six keyframe instructions, with each keyframe corresponding to a specific time point on the video timeline and its required visual content.

[0111] Table 1: Structured Order Requirements Specifications

[0112] Table 2 Content Script Keyframe Instruction Sequence:

[0113] The system retrieved three past youth-oriented video advertisements for the brand from the material library based on the brand identification code BRAND-SPT-A07. Each video was 12-18 seconds long, and a total of 156 keyframe sequences were extracted. The analysis module statistically analyzed the position, size, and animation of the brand logo in these historical materials. It was found that the brand logo was usually located in the upper part of the screen at the end of the historical materials (the last 3-5 seconds), occupying 8%-12% of the screen area. The 95th percentile of the logo's center position offset between adjacent frames was 18 pixels. At the same time, the distribution area and color stability parameters of the brand's standard color (orange-red, RGB value 255-87-51) in each frame were extracted.

[0114] Table 3 Brand Time-Series Gene Feature Vector Set:

[0115] Calculation of the diagonal length of the permissible area for the brand logo: With a screen width of 1080 pixels and a height of 1920 pixels, the horizontal span of the permissible area for the logo is... Pixels, vertical span is pixels, diagonal length:

[0116] The global position offset limit is set to 20% of the diagonal length, i.e. Pixel.

[0117] The system loads corresponding visual attention distribution pattern parameters and attention decay curve parameters from the audience attention parameter database based on the target audience tag "teenager basketball enthusiasts (13-18 years old)". The attention parameters for this audience group are statistically derived from 2300 sets of eye-tracking experiment data accumulated by the platform. This shows that the group's attention weight for areas with human movement in the image is 0.45, for product details is 0.35, and for text information is 0.20. The attention decay curve parameters show that the group's attention intensity decays at the same spatial location at a rate that is the reciprocal of 0.32 seconds, conforming to an exponential decay pattern.

[0118] Table 4 Attention parameters of the target audience:

[0119] The system injects these parameters into the audience adaptation unit of the audience attention temporal prediction model. The spatial attention encoding unit's convolutional neural network input feature vector has a dimension of 512, and the output 2D heatmap has the same resolution as the screen resolution of 1080×1920. The temporal propagation unit's LSTM hidden state vector has a dimension of 256, and the inter-frame attention transfer feature target has a dimension of 128. The audience adaptation unit's fully connected neural network concatenates the 128-dimensional inter-frame attention transfer feature with the 5-dimensional audience parameter vector to form a 133-dimensional input, which, after passing through two hidden layers (64 and 32 nodes respectively), outputs an attention heatmap corresponding to the screen resolution.

[0120] The system inputs the six keyframe instruction sequences from Table 2 into the configured audience attention temporal prediction model. The model performs attention heatmap prediction on each keyframe and its interpolated intermediate frames (a total of 450 frames, corresponding to 15 seconds × 30 fps). In the frame corresponding to instruction 001 at timestamp 0.0, the audience attention hotspot is concentrated in the area of ​​the player dribbling action in the lower part of the screen, with a peak attention intensity of 0.82. In the frame corresponding to instruction 002 at timestamp 3.0, the attention hotspot shifts to the close-up area of ​​the side of the sneakers in the lower part of the screen, with a peak attention intensity of 0.78. In the frame corresponding to instruction 005 at timestamp 12.0, the attention hotspot is dispersed in the area of ​​the product model text and brand logo combination in the middle of the screen, with the peak attention intensity dropping to 0.61.

[0121] The system aggregates attention heatmaps from each frame along the timeline and uses a peak matching algorithm based on a spatial distance threshold to track the trajectory of the attention heatmap. The preset intensity threshold is set to 60% of the maximum attention intensity in each frame, and the preset spatial distance threshold is set to 5% of the 1080-pixel screen width, i.e., 54 pixels. Tracking results show that from timestamp 0.0 to 6.0 seconds, the attention heatmap mainly moves from the lower center of the screen to the lower part, with a vertical position ranging from 50% to 75% of the screen height; from timestamp 9.0 to 15.0 seconds, the attention heatmap gradually migrates towards the upper center of the screen, with a vertical position ranging from 30% to 50% of the screen height.

[0122] Table 5: Location and intensity of audience attention hotspots at key time points:

[0123] From timestamp 0.0 to 14.0, the target audience's attention intensity at different spatial locations exhibits a continuously decreasing trend, conforming to the exponential decay pattern of the attention temporal decay curve parameters. Taking the initial attention intensity at timestamp 0.0 as an example, the attention intensity decay after 3.0 seconds is calculated as follows:

[0124] Actual tracking data shows that the attention intensity in this area was 0.78 at 3.0 seconds. However, this value is the initial intensity of the new attention hotspot (close-up area on the side of the shoe), not the sustained attention intensity of the original area (the area where the player dribbles). The attention intensity of the original area has dropped to about 0.35, which is basically consistent with the theoretical decay value.

[0125] The system performs spatiotemporal alignment analysis between the brand's temporal gene feature vector, the permissible position range of the brand logo at the end stage (timestamp 12.0-15.0 seconds), and the audience's attention temporal trajectory map.

[0126] For instruction frame 005 corresponding to timestamp 12.0 seconds, the center point of the permissible position range for the brand logo is located at screen coordinates (540, 576), corresponding to a vertical position of 30% of the screen height of 1920 pixels. The permissible logo position area is a rectangular area centered on this center point, with a horizontal span of 216 pixels and a vertical span of 192 pixels, and the area is [area missing]. Square pixel.

[0127] The center of the audience's attention hotspot in this frame is located at coordinates (540, 620). The area where the attention intensity exceeds the threshold (0.61 × 0.6 = 0.366) is an elliptical region with a radius of approximately 210 pixels centered at this center point, and an area of ​​approximately 142,000 square pixels. The intersection of the brand logo's permissible position area and the attention hotspot is the vertical overlap between the two areas, with an overlap area of ​​approximately... Square pixel.

[0128] Spatial overlap is calculated as follows:

[0129] Because the spatial overlap of the frame at timestamp 12.0 seconds (0.458) exceeds the preset overlap threshold of 0.3, this frame was not marked as an attention-guided intervention frame. However, in instruction frame 006 corresponding to timestamp 14.0 seconds, the center point of the brand logo's permissible position is still located at coordinates (540, 576), while the center of the audience's attention hotspot has moved to coordinates (540, 580), the area of ​​the attention hotspot has shrunk to 128,000 square pixels, and the area of ​​the intersection region has decreased to... Square pixels, spatial overlap reduced to:

[0130] Although the value is still higher than the 0.3 threshold, considering that the peak attention intensity has dropped to 0.58, the system detected that the spatial overlap of multiple intermediate frames in the time stamp range of 13.5 seconds to 14.5 seconds was less than 0.3, and marked these frames as attention-guided intervention frames.

[0131] Table 6. Spatiotemporal alignment analysis results (timestamp interval 12.0-15.0 seconds):

[0132]

[0133] For the attention guidance intervention frame at timestamp 14.5 seconds, the system calculates the attention perception position adjustment vector. The attention center point of this frame is located at coordinates (540, 578), the current center position of the brand logo is (540, 576), and the target direction unit vector is:

[0134] That is, a purely vertical downward direction. The inter-frame allowable offset threshold defined in the brand temporal gene feature vector set is 18 pixels, and the adjustment intensity coefficient is set to 0.6. The attention-aware position adjustment vector is calculated as follows:

[0135] Generate attention-consistency joint temporal constraint parameters;

[0136] The system marks the frame with timestamp 14.5 seconds as the attention-guided intervention frame, recording its attention perception position adjustment vector as (0, 10.8) pixels, indicating that the brand logo center point should be shifted downwards by 10.8 pixels to coordinates (540, 586.8). For non-intervention frames (frames with timestamps 12.0-14.0 seconds and after 14.5 seconds), the permissible position range and inter-frame permissible offset threshold of 18 pixels from the original brand temporal gene feature vector set are used.

[0137] To achieve a smooth transition, the system calculates the number of transition frames:

[0138] Apply the preceding transition adjustment vector to the frame preceding the 14.5-second frame (timestamp 14.467 seconds). The pixel will have its subsequent transition adjustment vector applied in the next frame (timestamp 14.533 seconds). Pixel.

[0139] Table 7 Attention-Consistency Joint Timing Constraint Parameters (Partial Frames):

[0140]

[0141] The relaxation threshold for attention-guided intervention frames is calculated as follows:

[0142] The system generates brand gene constraints for the first frame (timestamp 0.0 seconds, frame corresponding to instruction 001) based on the brand temporal gene feature vector set. This frame is a top-down view of a basketball court, with the brand logo appearing as a watermark in the upper right corner of the image, centered at (972, 150), occupying 2.8% of the image area, which is within the permissible range for brands in sports scene frames. This frame serves as the anchor frame, and the position and state of its brand elements serve as the initial reference benchmark for the generation of subsequent frames.

[0143] For each frame between timestamps 0.0 and 14.0, the system injects the brand element positions and states from the preceding frames, as well as the constraint information corresponding to each frame in the attention-consistency joint temporal constraint parameters, into the AI-assisted content generation module. During the generation of instruction frame 005 at timestamp 12.0, the brand logo transitions from the watermark position in the upper right corner of the screen to its display position in the upper middle of the screen, with center coordinates of (540, 576), and its size expands to occupy 10.2% of the screen area.

[0144] The system performs a static brand constraint check on this frame: the logo center coordinates (540, 576) correspond to 50% of the screen width horizontally and 30% of the screen height vertically, both within the allowable range (40%-60% horizontally, 25%-35% vertically); the logo size of 10.2% is within the allowable range of 8%-12%, and the static constraint check passes. The timing consistency constraint check calculates the Euclidean distance between the logo center coordinates (540, 576) in this frame and the logo center coordinates (540, 588) in the previous frame (timestamp 11.967 seconds).

[0145] The distance of 12 pixels is less than the inter-frame allowable offset threshold of 18 pixels, and the timing consistency constraint check is passed.

[0146] For the attention-guided intervention frame at timestamp 14.5 seconds, the system guides the brand logo center point downwards from its original permissible position (540, 576) to (540, 586.8) based on the attention-aware position adjustment vector (0, 10.8) pixels. The system applies relaxed consistency constraints to this frame and calculates the Euclidean distance between the logo center coordinates (540, 586.8) in this frame and the logo center coordinates (540, 581.4) in the preceding transition frame (timestamp 14.467 seconds).

[0147] The distance of 5.4 pixels is less than the relaxed threshold of 26.64 pixels, so the relaxed consistency constraint verification passes.

[0148] The system concatenates the generated content of each frame in chronological order, and outputs a temporally related frame sequence containing 450 frames of attention-guided data, at a frame rate of 30fps and a total duration of 15 seconds.

[0149] The system calculates the temporal consistency score of the brand logo in the attention-guided temporally related frame sequence along the time axis. For 449 pairs of adjacent frames out of 450 adjacent frames, the Euclidean distance between the logo center position, the change in logo size, and the color difference value of the logo's main color tone are calculated between each pair of frames.

[0150] Within the time stamp range of 12.0-15.0 seconds (corresponding to 90 frames), the average Euclidean distance of the logo center position is 8.6 pixels, and the maximum value is 12 pixels (occurring between the 12.0 second frame and the preceding frame); the average change in logo size is 0.08%, and the maximum value is 0.15%; the average color difference of the logo's main color is ΔE=1.2, and the maximum value is ΔE=2.3.

[0151] The following metrics were normalized: positional offset was normalized based on an inter-frame allowable offset threshold of 18 pixels; size variation was normalized based on an allowable size range of 4%; and color difference was normalized based on an allowable color difference threshold ΔE = 5.2. Weighting coefficients of 0.5, 0.3, and 0.2 were used to calculate the temporal consistency score.

[0152] Substituting the interval average into a simplified estimate:

[0153] However, considering that the logo position and size change more smoothly in the 0.0-12.0 second time stamp range, the actual temporal consistency score of the entire sequence is 0.782.

[0154] Based on the audience attention time-series trajectory map, the system extracts the audience attention intensity value at the location of the brand logo in each frame. At frame 12.0, the attention intensity value at the center position of the logo (540, 576) is 0.52; at the attention guidance intervention frame at time 14.5, the attention intensity value at the center position (540, 586.8) of the logo increases to 0.56 after the logo is adjusted; at frame 15.0, the attention intensity value at the center position of the logo (540, 576) is 0.48.

[0155] For the entire sequence of 450 frames, the sum of the attention intensity values ​​at the logo position in each frame is calculated to be 238.7, and the sum of the maximum attention intensity values ​​in each frame is 362.5. The attention temporal coverage score is calculated as follows:

[0156] Table 8. Results of the two-dimensional scoring:

[0157] Since both scores exceeded the corresponding thresholds, the system approved the video as a candidate. The system added attention guidance effect annotation information to the video content: the frame with the timestamp of 14.5 seconds was the attention guidance intervention frame, and the position adjustment was 10.8 pixels vertically downwards. Before the adjustment, the attention intensity value of the logo position in this frame was 0.52, and after the adjustment, it increased to 0.56, with a change in attention coverage of +7.7%; the change in temporal consistency score before and after the adjustment was -0.018 (relative to the theoretical score of 0.800 in the unadjusted case).

[0158] The system generates an attention coverage visualization overlay. Within the overlay at the 14.5-second frame, audience attention hotspots are marked with a gradient from blue (attention intensity 0.2-0.4 area) to yellow (attention intensity 0.4-0.6 area) to red (attention intensity 0.6-0.8 area), and the brand logo is marked with a green box. The overlay's opacity is set to 50%, and it is overlaid with the original frame content to generate a preview version with attention coverage visualization, which is then attached to the video ad creative file.

[0159] Table 9: Information on the effectiveness of attention guidance:

[0160] The system pushed the video ad creative file to the online collaborative editing module. On March 18, 20XX, the brand's design team logged into the collaborative editing interface via the web to view the preview version with attention coverage visualization. They requested a minor color adjustment for the product close-up shot at 9.0 seconds, increasing the saturation of the orange-red on the sneaker upper by 8% to enhance the visual impact of the brand's main color. The platform designer completed this modification via mobile on March 19, 20XX, and the system generated version snapshot V1.1 and recorded the change log. On March 20, 20XX, the brand confirmed the final version, and the system output the final delivery file. The file encoding format was H.264, the color space was BT.709, and the file size was 8.73MB.

[0161] The system performs delivery acceptance verification based on the final delivery document and structured order requirements specifications. Requirements compliance verification results: Video length 15.02 seconds, deviation from the order requirement of 15 seconds is +0.02 seconds, within the allowable error range of ±0.1 seconds; screen resolution 1080×1920, completely consistent with the order requirements; content theme is a product display and brand information communication for youth basketball shoes, with a matching score of 0.94 for the order script theme. Technical specification verification results: Encoding format H.264 conforms to platform delivery standards; color space BT.709 conforms to vertical video standards; file size 8.73MB is lower than the platform's 10MB limit for a single 15-second video.

[0162] The system combined the brand element temporal consistency score (0.782) and the audience attention temporal coverage score (0.658) to generate a delivery acceptance report, with an overall rating of "Excellent". The report noted that the application of attention-guiding technology at the 14.5-second frame of the timestamp increased the brand logo's attention coverage among the target teenage audience by 7.7% during that time period, while the brand temporal consistency score only decreased by 0.018, achieving an effective balance between brand consistency and audience attention guidance. The system triggered an order status update to the acceptance and settlement stage, and the order was settled on March 21, 20XX.

[0163] In this application example, the data starts from the initial order requirements specifications (Tables 1-2), extracts the brand temporal gene feature vector set through brand historical material retrieval (Table 3), then loads the target audience attention parameters (Table 4) to configure the audience attention temporal prediction model, generating an audience attention temporal trajectory map (Table 5). Through spatiotemporal alignment analysis (Table 6), intervention frames with insufficient spatial overlap between the permissible position of brand elements and the audience attention hotspot are identified. Attention perception position adjustment vectors are calculated, and attention-consistency joint temporal constraint parameters are generated (Table 7). During frame-by-frame generation, the original brand constraints are retained for non-intervention frames, while adjustment vectors are applied to intervention frames to guide brand elements to shift towards the attention hotspot. Finally, a dual-dimensional scoring (Table 8) is used to filter and generate video ad creative files with attached attention guidance effect annotation information (Table 9). The entire data flow process demonstrates a complete logical chain from order requirement analysis, brand gene extraction, audience attention prediction, spatiotemporal alignment analysis, joint constraint generation, attention guidance generation to dual-dimensional evaluation and filtering, ensuring the collaborative optimization of brand temporal consistency and target audience attention guidance in the AI-assisted generation process of video ads.

[0164] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for digital creative advertising design, production, and order management, characterized in that, Includes the following steps: Acquire video ad order data, parse and extract brand logos, target audience tags, and keyframe instruction sequences from the content script, and generate structured order requirement specifications; Based on the brand identity, historical video advertising materials of the brand are retrieved from the material library. The permissible position range of brand elements in the frame sequence, the permissible offset threshold between frames, and the motion effect transition style parameters are extracted to generate a set of brand temporal gene feature vectors. Based on the target audience tags, load the corresponding visual attention distribution pattern parameters and attention temporal decay curve parameters, and inject them into the audience attention temporal prediction model to complete the model configuration; The keyframe instruction sequence of the content script is input into the audience attention time-series prediction model to obtain the audience attention heatmap of each frame, and aggregated along the time axis to generate the audience attention time-series trajectory map. The temporal position constraint range of brand elements in the brand temporal gene feature vector set is spatiotemporally aligned with the audience attention temporal trajectory map. The spatial overlap between the allowable position area of ​​brand elements and the attention hot zone is calculated frame by frame. Frames with spatial overlap below a preset overlap threshold are marked as attention guidance intervention frames and attention perception position adjustment vectors are calculated to generate attention-consistency joint temporal constraint parameters. Based on the brand temporal gene feature vector set and the attention-consistency joint temporal constraint parameters, brand static constraint verification and temporal consistency constraint verification are performed on non-intervention frames, and attention perception position adjustment guidance and relaxed consistency constraint verification are performed on attention-guided intervention frames to generate a temporal associated frame sequence of attention guidance. The brand element temporal consistency score and attention temporal coverage score are calculated for the temporally associated frame sequence. Video candidate content with both scores exceeding the corresponding thresholds is selected and a video ad creative file is generated.

2. The digital creative advertising design, production, and order management method according to claim 1, characterized in that, The audience attention temporal prediction model includes a spatial attention encoding unit, a temporal propagation unit, and an audience adaptation unit. The spatial attention encoding unit receives the content description information of a single frame and outputs the basic spatial attention distribution of a single frame. The temporal propagation unit receives the spatial attention basic distribution sequence of adjacent frames, propagates the attention state along the time axis, and outputs the inter-frame attention transfer features. The audience adaptation unit receives the inter-frame attention transfer features, the visual attention distribution pattern parameters, and the attention temporal decay curve parameters, and outputs attention heatmaps for each frame that are adapted to the target audience. The spatial attention encoding unit uses a convolutional neural network, and the output layer uses a sigmoid activation function to normalize the attention intensity at each pixel position to the range of zero to one. The temporal propagation unit uses an LSTM neural network, and maps the hidden state vector to inter-frame attention transfer features through a linear mapping layer.

3. The digital creative advertising design, production, and order management method according to claim 2, characterized in that, The audience adaptation unit employs a parameter condition injection method, using the visual attention distribution pattern parameter as the spatial dimension modulation coefficient and the attention temporal decay curve parameter as the temporal dimension modulation coefficient to perform spatial and temporal dimension modulation on the inter-frame attention transfer features, respectively. The spatial dimension modulation is achieved by multiplying the spatial distribution of the inter-frame attention transfer features pixel by pixel with the spatial weight map corresponding to the visual attention distribution pattern parameter. The temporal dimension modulation is achieved by applying an exponential decay function to the attention intensity value of the current frame. The output of the exponential decay function is the product of the initial attention intensity and the value of a natural exponential function with the decay rate parameter and the time variable as exponents.

4. The digital creative advertising design, production, and order management method according to claim 1, characterized in that, The aggregation along the time axis to generate the audience attention time-series trajectory map includes: arranging the two-dimensional attention heatmaps of each frame into a three-dimensional data structure of two-dimensional space plus one-dimensional time according to the frame time sequence; using an inter-frame attention peak matching algorithm based on a spatial distance threshold, for each attention peak point in the current frame whose attention intensity exceeds a preset intensity threshold, searching for attention peak points in the next frame whose spatial distance is less than a preset spatial distance threshold, if they exist, establishing a trajectory association, otherwise marking the trajectory as terminated in the current frame, and marking the unassociated attention peak points in the next frame as the starting point of a new trajectory; the attention intensity change trend is obtained by performing time-series difference calculation on the attention intensity values ​​of the corresponding peak points in each frame on the same trajectory; the preset intensity threshold is taken as 60% of the maximum attention intensity in the attention heatmap of each frame, and the preset spatial distance threshold is taken as 5% of the screen width.

5. The digital creative advertising design, production, and order management method according to claim 1, characterized in that, The calculation of the attention perception position adjustment vector includes: determining the spatial position corresponding to the maximum attention intensity in the audience attention heatmap of the attention guidance intervention frame as the attention center point; calculating the unit direction vector pointing from the current center position of the brand element to the attention center point as the target direction; taking the inter-frame allowable offset threshold as the maximum allowable offset; the attention perception position adjustment vector is the product of the adjustment intensity coefficient, the maximum allowable offset, and the unit vector of the target direction, wherein the value of the adjustment intensity coefficient is greater than zero and less than or equal to one; wherein the spatial overlap is the ratio of the intersection area of ​​the brand element's allowable position area and the attention heat area to the area of ​​the brand element's allowable position area, and the preset overlap threshold value is 0.

3.

6. The digital creative advertising design, production, and order management method according to claim 5, characterized in that, Also includes: Along the time axis, for consecutive frames labeled as attention-guided intervention frames, the cumulative offset of the attention perception position adjustment vector of each intervention frame within the frame segment is calculated. The cumulative offset is the magnitude of the vector sum of the adjustment vectors of each intervention frame within the frame segment. When the cumulative offset exceeds the global position offset upper limit defined in the brand temporal gene feature vector set, the adjustment vectors of subsequent intervention frames in the frame segment are scaled proportionally by a scaling factor to make the cumulative offset converge to within the global position offset upper limit. The scaling factor is the quotient obtained by dividing the difference between the global position offset upper limit and the magnitude of the vector sum of the adjustment vectors of each intervention frame before the current intervention frame by the magnitude of the adjustment vector of the current intervention frame. The global position offset upper limit is 20% of the diagonal length of the allowable position area of ​​the brand element.

7. The digital creative advertising design, production, and order management method according to claim 5, characterized in that, Also includes: For each attention-guided intervention frame, the attention-aware position adjustment vector is linearly interpolated and extended along the time axis to the preceding and subsequent non-intervention frames to generate a progressive adjustment vector for the transition frames. The number of transition frames is the rounded-up value of the ratio of the magnitude of the attention-aware position adjustment vector to the inter-frame allowable offset threshold. The progressive adjustment vector of the preceding transition frame is the ratio of the transition frame number to the number of transition frames multiplied by the attention-aware position adjustment vector, and the progressive adjustment vector of the subsequent transition frame is a product of the ratio of the transition frame number to the number of transition frames and the attention-aware position adjustment vector. The progressive adjustment vector of the transition frames is then added to the attention-consistency joint temporal constraint parameters.

8. The digital creative advertising design, production, and order management method according to claim 1, characterized in that, In the relaxed consistency constraint verification, the relaxed threshold is the sum of the original inter-frame allowable offset threshold and the relaxed coefficient multiplied by the magnitude of the attention perception position adjustment vector, with the relaxed coefficient set to 0.

8. The brand element temporal consistency score is calculated by subtracting the weighted average of the normalized values ​​of brand element position offset, size change, and tone deviation between adjacent frames from one, with weighting coefficients of 0.5, 0.3, and 0.2, respectively, and the sum of the three is one. The attention temporal coverage score is the ratio of the sum of the audience attention intensity values ​​at the locations of key brand elements in each frame to the sum of the maximum audience attention intensity values ​​in each frame.

9. The digital creative advertising design, production, and order management method according to claim 1, characterized in that, Also includes: An attention overlay visual layer is generated for each frame in the video ad creative file. The location of the audience's attention hot zone is marked with a semi-transparent thermal color mark and the location of the brand's key elements is marked with a marker box. The attention overlay visual layer is superimposed with the original frame content to generate a preview version with attention overlay visualization and is attached to the video ad creative file. The video ad creative file is pushed to the online collaborative editing module, collaborative modification operations are recorded, version snapshots and change logs are generated through version control, and the final version delivery file is output. Based on the final delivery document and the verification of compliance with the structured order requirements and technical specifications, a delivery acceptance report is generated by combining the brand element temporal consistency score and the attention temporal coverage score.

10. A digital creative advertising design, production, and order management system, used to execute the digital creative advertising design, production, and order management method according to any one of claims 1-9, characterized in that, include: The order parsing module is used to acquire video ad order data, parse and extract brand identifiers, target audience tags and content script keyframe instruction sequences, and generate structured order requirement specifications. The brand gene extraction module is used to retrieve historical video advertising materials of the brand from the material library based on the brand logo, extract the permissible position range of brand elements in the frame sequence, the permissible offset threshold between frames and the motion effect transition style parameters, and generate a brand temporal gene feature vector set. The attention model configuration module is used to load the corresponding visual attention distribution pattern parameters and attention temporal decay curve parameters based on the target audience group labels, and inject them into the audience attention temporal prediction model to complete the model configuration. The attention trajectory generation module is used to input the content script keyframe instruction sequence into the audience attention temporal prediction model, obtain the audience attention heatmap of each frame, and aggregate it along the time axis to generate the audience attention temporal trajectory map. The spatiotemporal alignment analysis module is used to perform spatiotemporal alignment analysis between the temporal position constraint range of brand elements in the brand temporal gene feature vector set and the audience attention temporal trajectory map, calculate the spatial overlap on a frame-by-frame basis, mark frames with spatial overlap below a preset overlap threshold as attention guidance intervention frames and calculate the attention perception position adjustment vector, and generate attention-consistency joint temporal constraint parameters. The content generation module is used to perform brand static constraint verification and temporal consistency constraint verification on non-intervention frames based on the brand temporal gene feature vector set and the attention-consistency joint temporal constraint parameters, and to perform attention perception position adjustment guidance and apply relaxed consistency constraint verification on attention-guided intervention frames to generate a temporal associated frame sequence guided by attention. The dual-dimensional evaluation module is used to calculate the brand element temporal consistency score and attention temporal coverage score for the temporally associated frame sequence, filter video candidate content whose dual-dimensional scores both exceed the corresponding thresholds, and generate video ad creative files.

Images