Product packaging planning-oriented AI-driven full-marketing content design method

Abstract

The invention discloses a product package planning-oriented AI-driven full marketing content design method, and relates to the technical field of marketing content design, and the method comprises the following steps: obtaining a product package design draft, extracting color, composition, font and semantic features based on a deep convolutional neural network, and carrying out the fusion to form a brand tonality mother vector of brand style tonality; calling a multi-modal analysis module, extracting visual and language features consistent with brand tonality vector dimensions from multi-modal marketing materials published by a brand recently frame by frame or sentence by sentence, and generating a marketing material feature vector set; according to the method, the brand tonality mother vector is constructed as the style anchor point, and the multi-modal material analysis, the time-sensitive style evaluation mechanism and the self-adaptive style jump strategy are combined, so that the content style consistency is monitored in real time, the capability of dynamically regulating and controlling the style according to the content fatigue risk is realized, and the real-time monitoring of the content style consistency is realized. And the content generation system is ensured to intelligently sense the change of the propagation environment and make a strategic response.

Description

Technical Field

[0001] The present invention relates to the field of marketing content design technology, and in particular to an AI-driven full marketing content design method for product packaging planning. Background Art

[0002] AI-driven full marketing content design for product packaging planning uses product packaging design as the starting point and core basis for generating marketing ideas. Leveraging generative AI and multimodal content modeling technology, it automates the entire content creation process, from packaging to various marketing materials (such as promotional copy, visual posters, e-commerce detail pages, short video scripts, etc.). This approach extracts the brand's visual elements, language tone, and design style embedded in the packaging, driving the AI system to generate stylistically consistent and semantically relevant multimodal marketing content. Adaptive optimization is then performed based on user profiles and market feedback, resulting in an integrated marketing content design model characterized by unified brand expression, efficient and collaborative content creation, and intelligently evolving communication strategies. This approach effectively breaks down the barriers between packaging design and marketing communications, transforming packaging from a mere physical carrier into a "style anchor" and "creative hub" that drives the entire marketing content planning and production process.

[0003] The existing technology has the following deficiencies: In the process of AI-driven marketing content design for product packaging planning, to achieve a unified brand tone, existing technologies generally adopt fixed style transfer and content generation strategies, using the packaging visual style and language expression tone as a template basis to guide the generation of subsequent multimodal marketing content. However, in the process of pursuing content consistency, such methods often lack dynamic adjustment mechanisms for the diversity of typesetting structure, flexibility of color matching, and semantic level of copywriting. As a result, the generated graphic and text content tends to be homogenized in expression, making it difficult to present emotional tension and creative changes. With the repeated release of such content on multiple channels, users are prone to aesthetic fatigue and perceptual resistance, which in turn weakens the dissemination appeal of the content and the efficiency of interactive conversion on social platforms. At the same time, this problem also limits the system's ability to quickly adapt to changes in the packaging style of new products, affecting the brand's content response speed and continuous innovation capabilities in a dynamic marketing environment, becoming a major technical bottleneck restricting the large-scale application of AI-driven content systems.

[0004] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Summary of the Invention

[0005] The object of the present invention is to provide an AI-driven full marketing content design method for product packaging planning. By constructing a brand tone master vector, combining multimodal analysis, temporal evaluation and style jump mechanism, it realizes the consistency monitoring and dynamic regulation of content style, improves the intelligence and adaptability of content generation, enhances the attractiveness and interaction efficiency of brand communication, and extends the content life cycle to solve the problems in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions: An AI-driven full marketing content design method for product packaging planning, including the following steps: S101. Obtain the product packaging design draft, extract color, composition, font and semantic features based on a deep convolutional neural network, and fuse them to form a brand tone master vector of brand style tone; S102. Call the multimodal analysis module to extract visual and language features consistent with the dimensions of the brand tone master vector frame by frame or sentence by sentence from the multimodal marketing materials recently released by the brand, and generate a set of marketing material feature vectors; S103. Calculate the cosine similarity between each marketing material feature vector and the brand tone master vector to obtain a style consistency score, and construct a style consistency score sequence in chronological order; S104. Based on a preset temporal decay weight function, weight the style consistency score sequence to construct a dynamic consistency weight matrix, where new materials correspond to high weights and old materials correspond to low weights; S105. Merge the dynamic consistency weight matrix through an adaptive aggregation algorithm to obtain an initial style unity intensity value, and normalize and output a style unity coefficient; S106. When the style unity coefficient exceeds a preset risk threshold, construct a multi-dimensional style space mapping graph based on the current brand tone state, calculate the jump path according to the style unity coefficient, and dynamically adjust the jump amplitude in combination with the tone unity intensity.

[0007] Preferably, the steps for generating the brand tone master vector include: Obtain the product packaging design draft image and input it into a pre-trained deep convolutional neural network model for preprocessing operations to extract color and composition features; Identify the text area in the packaging image through a text detection and recognition model, and extract font features by a font classification network; Input the recognized text into a semantic embedding module to extract language tone and emotional semantic features; Fuse and encode the color, composition, font and semantic features, and weight them through an attention mechanism to generate a brand tone master vector.

[0008] Preferably, the steps for calling the multimodal analysis module to extract marketing material feature vectors include: Collect the image, video, and text marketing materials released by the brand within the preset time range and perform standardization processing; Sample the video materials frame by frame, extract key frame images, and extract visual features through a deep convolutional neural network; Divide the text materials sentence by sentence and extract semantic sentiment and language tonality features through a natural language processing module; Integrate the image and text features in the order of the materials to generate a set of marketing material feature vectors consistent with the dimension of the brand tonality mother vector.

[0009] Preferably, the steps for constructing the style consistency score sequence include: Perform L2 normalization on each feature vector in the marketing material feature vector set and the brand tonality mother vector; Call the cosine similarity function to calculate the similarity score between each feature vector and the brand tonality mother vector; Sort the similarity scores in the order of the original release time of the materials; Generate a style consistency score sequence in chronological order.

[0010] Preferably, the steps for constructing the dynamic consistency weight matrix include: Calculate the relative time offset according to the timestamp information of each item in the style consistency score sequence; Call the time series decay weight function model to calculate the time decay weight for each score; Multiply each style consistency score by the corresponding weight value to generate a weighted score; Map all the weighted scores to a set of dynamic consistency weight matrices arranged in the order of the materials.

[0011] Preferably, the time series decay weight function is an exponential decay function, and its expression is W(t) = e^(-λt), where t is the time offset and λ is the decay coefficient. The exponential decay function is used to ensure that new materials correspond to high weights and old materials correspond to low weights, realizing the priority of content style evaluation in response to time changes.

[0012] Preferably, the steps for generating the style unity coefficient include: Receive and parse the dynamic consistency weight matrix, and extract the time series weighted feature scores of each style sub-dimension; Perform weighted fusion or non-linear merging on the column feature values to obtain the initial style unity intensity value; Use a normalization algorithm to map the initial intensity value to a style unity coefficient between 0 and 1; Store the style unity coefficient in the content monitoring dashboard for real-time viewing and analysis.

[0013] Preferably, the steps of calculating the jump path of the style unity coefficient and dynamically adjusting the jump amplitude in combination with the tonality unity intensity include: When the style unity coefficient exceeds the preset risk threshold, construct a multi-dimensional style space mapping diagram based on the brand tonality state; Determine the current style coordinates in the style space according to the style unity coefficient, and calculate the jump path; Dynamically adjust the jump amplitude in combination with the brand tonality unity intensity to avoid tonality fragmentation; Update the style parameters of the content generation model according to the jump coordinates to achieve differential style control and creative guidance for subsequent content.

[0014] In the above technical solution, the technical effects and advantages provided by the present invention are: By constructing a brand tonality mother vector as a style anchor point, combining multi-modal material analysis, a time-sensitive style evaluation mechanism, and an adaptive style jump strategy, the present invention not only realizes real-time monitoring of content style consistency, but also has the ability to dynamically adjust the style according to the content fatigue risk, ensuring that the content generation system can intelligently perceive changes in the communication environment and make strategic responses. This method significantly improves the flexibility and intelligence level of content creation, enhances the content attractiveness and user interaction efficiency of the brand in multi-channel communication, effectively extends the life cycle and communication value of brand content, and has outstanding practicality and broad promotion prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments recorded in the present invention. For those of ordinary skill in the art, other drawings can also be obtained according to these drawings.

[0016] Figure 1 It is a method flow chart of an AI-driven full-marketing content design method for product packaging planning according to the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0017] Now, example embodiments will be described more fully with reference to the accompanying drawings. However, the example embodiments can be implemented in various forms and should not be construed as limited to the examples set forth herein; rather, these example embodiments are provided so that this disclosure will be more complete and comprehensive, and will fully convey the concept of the example embodiments to those skilled in the art.

[0018] The present invention provides an Figure 1 AI-driven full-marketing content design method for product packaging planning as shown in the following, including the following steps: S101. Obtain the product packaging design draft, extract the color features, composition features, font features, and semantic features from the packaging design draft respectively based on the deep convolutional neural network model, and perform fusion processing on the multiple visual and language features obtained above to construct a brand tonality master vector for representing the brand style tonality; The steps for generating the brand tonality master vector specifically include the following four sub-steps, which are used to extract representative brand visual and language style features from the product packaging design draft to construct a high-dimensional semantic-consistent tonality representation vector, thereby providing basic support for subsequent marketing content generation and style consistency evaluation: Obtain the packaging design draft image file of the target product and input it into the pre-trained deep convolutional neural network model for image preprocessing operations, including size normalization, color standardization, and removal of background interference regions. The preprocessed packaging design image is input into the backbone feature extraction network composed of multiple convolutional layers and pooling layers to automatically learn and extract the low-level color distribution features (such as the main color, saturation, brightness ratio, etc.) and mid-level composition structure features (such as layout, spatial hierarchy, edge structure, etc.) in the image. These features achieve enhanced spatial perception through the combination of multi-scale convolutional kernels, ensuring that the model can identify the composition forms and color matching patterns with style representativeness in the image.

[0019] To obtain the font features in the packaging image, the present invention introduces an auxiliary path based on the joint model of text detection and recognition to locate, segment, and recognize the text regions in the packaging image. The recognized text content further extracts its font style features, such as font type (serif / sans-serif), thickness, inclination, stroke structure, etc., through the pre-trained font classification network, and is associated and encoded with the original image features. The font features obtained through this path not only reflect the visual communication language in the packaging design but also supplement the unified representation of the brand style at the font expression level.

[0020] To extract semantic features, the present invention uses the joint language-visual multi-modal alignment model to input the packaging text content recognized in the second step into the semantic embedding module based on the Transformer structure to extract the language tonality and semantic emotion features it expresses. This semantic embedding module supports the recognition of subjective style features such as "young and energetic", "sense of technology", "warm and friendly", etc., and realizes the collaborative coding from visual language to semantic style through fusion mapping with the image features extracted from the visual channel. The introduction of semantic features enables the extracted brand tonality representation to not only have the visible style characteristics at the sensory level but also have the language consistency description ability at the content level.

[0021] The color feature vector, composition feature vector, font feature vector, and semantic feature vector extracted from the above three channels are jointly encoded in high dimensions through a feature-level fusion network, and an attention mechanism is introduced to dynamically allocate the weights of the features in each dimension to form the final brand tonality master vector. As a high-dimensional representation that fuses visual and language style features, this brand tonality master vector has good style abstraction ability and expression consistency.

[0022] This step constructs a high-dimensional vector representation, namely the brand tonality master vector, which can comprehensively represent the brand's visual and language style tonality, through multi-dimensional deep feature extraction and fusion of the product packaging design draft. Product packaging is usually one of the most representative visual symbols in brand communication, which centrally reflects the brand's core style, aesthetic tendency, and communication context. By extracting visual features such as color, composition, and font in the packaging image based on a deep convolutional neural network model, and combining the semantic tonality features contained in the text content in the picture, a structured modeling of the brand tonality can be achieved. Fusing these visual and language features can not only enhance the model's ability to express the brand style consistently but also provide a standardized and quantifiable "style anchor point" for subsequent marketing content generation.

[0023] S102. Invoke the multimodal parsing module to process the multimodal marketing materials recently released by the brand, and extract visual and language feature vectors consistent with the dimensions of the brand tonality master vector frame by frame (for video materials) and sentence by sentence (for text materials), and generate a set of marketing material feature vectors in the order of the materials. In this embodiment, the step of invoking the multimodal parsing module to extract the brand marketing material feature vectors aims to uniformly extract the unified features at the semantic and style levels from various marketing materials (including images, videos, texts, etc.) released by the brand in different media channels to match and compare with the brand tonality master vector. This step includes the following specific operation processes: Collect all the marketing material data released by the brand within a preset time window (such as the past 30 days), covering video advertisement clips, social platform graphic and text pushes, e-commerce detail page images, product short videos, and various copywriting texts. For video materials, the system uniformly samples them frame by frame and extracts key frame images at a certain time interval; for text materials, they are structurally split sentence by sentence, and each sentence is used as an independent analysis unit. After the above materials are standardized (including size unification, frame rate normalization, language cleaning, etc.), they are uniformly sent to the multimodal parsing module for feature extraction preparation.

[0024] For the key frame image parts in images and videos, a deep convolutional neural network with the same structure as the packaging design draft feature extraction model is called to extract the visual features of each frame of the image respectively. The extraction content includes the main color distribution, composition rules, background complexity, visual center positioning, and font layout features of the image. The dimensions of the above visual features are strictly aligned with the visual part dimensions in the brand tone mother vector to ensure that one-to-one style similarity calculation can be achieved subsequently. For the image content containing text, the text is further recognized through OCR technology and semantically jointly encoded with the image features.

[0025] For text-based materials (including short copywriting, e-commerce titles, detailed paragraphs, etc.), the system uses a natural language processing module to perform semantic embedding on each sentence and extract feature information such as the semantic sentiment tendency, language tone, keyword distribution, and language structure complexity of the sentence. This natural language processing module is built based on the Transformer architecture and matches the dimension of the language feature vector in the brand tone mother vector to ensure that the multi-level expression of the text tone can be accurately embedded into the unified feature space. The semantic feature extraction not only includes the theme information at the content level but also integrates subjective style factors such as sentence style and emotional tension, thereby realizing the comprehensive mapping of semantic tone and visual style.

[0026] The feature vectors of each material unit (frame image or sentence) extracted above are integrated in the order of the material source to generate a complete set of marketing material feature vectors. Each feature vector maintains consistency with the brand tone mother vector in terms of dimension and structure, making subsequent similarity comparison, style consistency scoring, and dynamic style regulation possible. This feature set not only retains the expression features of the content at the visual and language levels but also embeds the time dimension information of the material, which can be used as the basic input for the entire content system to monitor the style and optimize the content, realizing cross-modal, cross-platform, and cross-time brand style consistency modeling and intervention control.

[0027] This step calls the multi-modal parsing module to perform refined feature extraction and standardized expression on various marketing contents recently released by the brand, and constructs a "marketing material feature vector set" that is strictly aligned with the dimensions of the brand tone master vector, so as to achieve the basic support for style consistency evaluation and dynamic content control. In practical applications, the brand's marketing contents usually exist in multi-modal forms, such as videos, graphics and texts, short copywriting, etc., which respectively carry the external expressions of visual style and language tone. This step first processes video materials frame by frame, extracts the visual features of each key frame therein (including color, composition, font layout, etc.), and identifies and encodes the text in the image; secondly, analyzes text materials sentence by sentence, and extracts features such as semantic emotions, sentence styles and tone tendencies in the language expression. Through this "cross-modal + unified dimension" processing method, the system can perform structured unified modeling on a large amount of heterogeneous contents, providing conditions for subsequent similarity calculation with the brand tone master vector. This step not only ensures the accuracy and consistency of content style analysis, but also provides a quantifiable and traceable data basis for subsequent style consistency scoring, style fatigue monitoring, creative style jumping, etc., so as to achieve style perception, intelligent decision-making and dynamic control of the entire process of marketing contents.

[0028] S103. Calculate the cosine similarity between each feature vector in the marketing material feature vector set and the brand tone master vector respectively to obtain the style consistency scores of multiple single materials, and construct a style consistency score sequence in the order of material time; In this embodiment, for the step of calculating the cosine similarity between each feature vector in the marketing material feature vector set and the brand tone master vector respectively to obtain the style consistency score and constructing the score sequence in time order, it specifically includes the following sub-steps, aiming to achieve the quantitative evaluation of the content style fit and time series modeling, which is convenient for subsequent trend monitoring and style dynamic control.

[0029] For the constructed marketing material feature vector set, traverse each feature vector unit in turn, and call the cosine similarity calculation function to calculate the similarity between this feature vector and the pre-generated brand tone master vector. The cosine similarity function outputs a similarity value ranging from 0 to 1 by calculating the cosine value of the angle between two vectors in the vector space, so as to measure the consistency degree between the style features of the current material and the overall brand tone. The closer the value is to 1, the more highly matched the material is with the brand tone in terms of visual and language styles. This calculation method has direction sensitivity and can effectively capture the structural differences between style features, so it is particularly suitable for brand style consistency analysis.

[0030] To ensure the accuracy of calculations and the unity of dimensions, the present invention performs L2 normalization on all feature vectors before they enter the similarity calculation process to eliminate the influence caused by different vector norms. This normalization step not only improves the calculation stability but also provides a mathematical basis for fair comparison between feature vectors of different modalities. The normalized vectors are projected onto the unit sphere, and the similarity calculation results are more accurate, avoiding deviations caused by content complexity or modality differences.

[0031] After calculating the similarity of each material unit, the obtained style consistency scores are sorted in the order of the original release time of the materials and stored one by one in the style consistency score sequence according to this time order. The time sorting mechanism adopted here can ensure that the subsequent analysis of the style evolution trend has a time dimension reference value, which helps to identify the fluctuation rules of style consistency in different release cycles. Each score carries timestamp information, which is used to construct the subsequent dynamic weighting matrix and decay function model.

[0032] After completing the traversal and calculation of the entire material set, the system outputs a complete style consistency score sequence, which can be used as the input basis for the brand tone monitoring system. Through this sequence, the operator can track the degree of fit between each piece of content and the brand tone in a data-driven manner, and it can be further used to construct a style monitoring dashboard, trigger a creative control mechanism, or generate a content fatigue risk warning. This step realizes the quantitative evaluation of the consistency between a single marketing material and the overall brand tone at the style level by calculating the cosine similarity between each vector in the marketing material feature vector set and the brand tone master vector, and constructs a style consistency score sequence in time order, providing a data basis for subsequent style trend analysis, dynamic monitoring, and content strategy adjustment. In the marketing process of a brand with high-frequency placements on multiple platforms, the consistency of the styles of its various materials directly affects users' brand recognition and content acceptance. Traditional methods are difficult to systematically measure this style consistency, while this step can accurately measure the matching degree between the material and the brand tone by calculating the cosine value of the angle between the feature vectors. The output similarity score ranges from 0 to 1, enabling the system to capture trends of style deviation, change, or convergence in real time. At the same time, by sorting according to the material release time and constructing a style consistency score sequence, the system can continuously monitor the evolution trajectory of the brand content style in the time dimension, providing a timeliness basis for subsequent introduction of time-series decay weighting, jump trigger mechanisms, etc. This step not only improves the objectivity and operability of content consistency evaluation but also lays a solid foundation for achieving efficient and flexible content style control.

[0033] S104. Based on a preset time-series decay weight function, perform weighted processing on each score value in the style consistency score sequence to construct a dynamic consistency weight matrix, where newer materials correspond to higher weights and older materials correspond to lower weights. In this embodiment, for the step of performing weighted processing on each score value in the style consistency score sequence based on a preset time-series decay weight function to construct a dynamic consistency weight matrix, the aim is to establish a time-sensitive style perception mechanism, enabling the system to more reasonably measure and respond to the evolution state of the brand content style in different time dimensions, thereby achieving the dynamic nature and precision control of style monitoring and regulation. This step includes the following key operation processes: Based on the style consistency score sequence constructed in the previous step, the system sorts each score according to the timestamp information of the materials and binds a corresponding time offset to each score. This time offset is a relative time difference calculated by backtracking from the current system time, and the unit can be hours, days, or weeks. This time offset information will be used as an input variable for the subsequent time-series decay function to apply decay processing to each score value. By establishing the mapping relationship between each score item and its corresponding time difference, the system can achieve precise time perception characteristics, providing a basis for dynamic weight allocation.

[0034] Call the preset time-series decay weight function model to calculate the weight for each style consistency score item. This weight function can take forms such as an exponential decay function, a hyperbolic decay function, or an adjustable power decay function. The specific function expression can be set as W(t) = e^(-λt), where t is the relative time offset value and λ is the decay coefficient, which controls the weight decline speed. For newer materials, due to the smaller t value, the corresponding weight value is close to 1, indicating a higher influence in style judgment; while for older materials, because the t value is larger, their weight values approach 0, and the influence gradually weakens. The design of this function ensures the response priority of content evaluation to the "current style state", enabling the system to have time awareness ability and rapid adaptation ability.

[0035] According to the time decay weights calculated above, apply them to each score item in the style consistency score sequence respectively, and generate a new weighted score sequence in an item-by-item weighted manner. Specifically, the system performs a vector-level multiplication operation, multiplying each score by its corresponding time-series decay weight to output a set of weighted score values. Subsequently, map this weighted score sequence to a multi-dimensional weight matrix. In this matrix structure, each row represents a material sample, and each column corresponds to a feature dimension or tonality factor, which can be extended for data alignment in multiple dimensions such as image style, language intonation, and visual composition. In the matrix structure, the row vectors of newer materials have a higher overall weight contribution, while the influence of older materials is automatically weakened due to weight decay.

[0036] The constructed dynamic consistency weight matrix is used as the input basis for subsequent aggregation operations and style jump judgment. This matrix not only retains the semantic information and structural distribution characteristics of the style consistency scores, but also introduces a dynamic weight adjustment mechanism that changes over time, thus significantly improving the real-time performance and strategic flexibility of style judgment. At the same time, the weight patterns in the matrix can also be used as the basis for content life cycle analysis to identify active areas and risk intervals of style fluctuations within a certain period, providing strategic references for content creative planning.

[0037] In this step, by introducing a preset time-series decay weight function, a time-sensitive weighting process is performed on the style consistency score sequence, thereby constructing a dynamic consistency weight matrix to achieve dynamic modeling and real-time monitoring of the brand content style state. In actual marketing content management, the materials released at different time nodes do not have the same impact on the brand tone. Especially in the context of high-frequency and multi-platform content delivery, recent materials are more representative of shaping user perception and brand image. By assigning higher weights to newer materials and lower weights to older materials, this step enables the system to prioritize responding to the content performance within the current time window when evaluating style consistency, and has a stronger sense of timeliness. At the same time, the constructed dynamic consistency weight matrix not only retains the structural information of the original content style scores, but also reflects the trend direction and risk concentration areas of style changes through the time-weighting mechanism. This provides basic data support with a time dimension for subsequent modules such as style aggregation, risk monitoring, and jump triggering, effectively improving the accuracy and intelligent level of content style management.

[0038] S105. Perform a merging operation on multiple feature dimensions of the dynamic consistency weight matrix through an adaptive aggregation algorithm to obtain an initial style unity strength value, normalize the initial style unity strength value, output a style unity coefficient with a value range between 0 and 1, and store the style unity coefficient in the content monitoring dashboard for real-time viewing; In this embodiment, the step of using an adaptive aggregation algorithm to perform a merging operation on multiple feature dimensions of the dynamic consistency weight matrix to generate a style unity strength value and normalize it into a style unity coefficient specifically includes the following operation process. This step aims to integrate the style consistency scoring results distributed in different modalities and dimensions into a quantifiable, monitorable, and traceable core indicator to comprehensively reflect the degree of unity of the current brand content at the visual and language style levels, providing a real-time reference basis for content operation and style control.

[0039] The system receives and parses the dynamic consistency weight matrix generated in the previous processing step. Each row in the matrix structure corresponds to a single material sample feature vector with temporal weight correction, and each column represents a specific style sub-dimension (such as color coordination, composition similarity, font style matching degree, semantic tone fitting degree, etc.). Each value in the matrix has combined the cosine similarity and the time decay factor, and can truly reflect the contribution differences of each dimension in different time segments. To unify the style evaluation criteria, this embodiment introduces an adaptive aggregation algorithm, which has the ability of multi-dimensional feature fusion and can dynamically adjust the fusion weights of different dimension features according to the input data distribution.

[0040] Perform feature merging operations on the dynamic consistency weight matrix. In this process, first perform weighted average or non-linear weighting (such as weighting using the attention mechanism, information entropy adjustment factor, or principal component weight extraction based on the covariance matrix) on each column of features to obtain the representative scores of each style sub-dimension; then perform combined merging on the scores of multiple dimensions through the adaptive aggregation algorithm. The aggregation algorithm can be selected from weighted average, fuzzy logic fusion, neural network regression, or a linear regression model based on multi-factor evaluation. The core feature of this algorithm is that it can automatically adjust the fusion strategy of each dimension according to the content distribution state. For example, when the current semantic tone difference is large but the visual style is highly consistent, the system can appropriately increase the weight coefficient of the semantic dimension to more accurately reflect the "global unity" of the style.

[0041] After completing the aggregation of all dimensions, the system outputs an initial style unity strength value. This value is usually an unnormalized floating-point value, indicating the degree of style fit between the current content set as a whole and the brand tone. To improve the operability of cross-project comparison and the convenience of monitoring, the present invention further designs a set of normalization mapping mechanisms. This mechanism is based on the maximum-minimum normalization method or the Z-Score standardization strategy of sample data, and converts the initial strength value into a normalized index between 0 and 1. The normalization process not only eliminates the influence of data scale differences on the final judgment, but also facilitates horizontal comparison with the unity data of other brands or time periods on the dashboard.

[0042] Write the normalized style unity coefficient into the visual monitoring dashboard through the content interface and display it in real time as the core index of the brand content style consistency. The dashboard supports various visualization modes such as trend charts, heat maps, and interval distribution charts, which are convenient for brand parties, content operation personnel, and creative teams to track style fluctuations in different time periods and timely identify risk signals of overly concentrated or discrete styles. If the style unity coefficient is too high for a long time, the dashboard will issue a style fatigue risk warning, prompting the content system to trigger the style jump mechanism; if it is too low, it will prompt that the style lacks unity and affects the brand recognition.

[0043] This step intelligently merges the style feature scores of multiple dimensions in the dynamic consistency weight matrix to calculate a quantitative index that can comprehensively reflect the degree of unity of the brand's current content style, namely the style unity coefficient. Since marketing content often includes multiple modalities and style feature dimensions such as vision, language, and layout, relying solely on a single dimension to judge cannot accurately evaluate the overall unity of the brand tone. Therefore, in this step, by introducing an adaptive aggregation algorithm, the fusion weights are dynamically allocated according to the importance and difference of different feature dimensions to achieve a comprehensive analysis of multi-dimensional style information. The initial style unity intensity value obtained after aggregation is normalized and mapped to a standardized coefficient between 0 and 1, making it have good horizontal comparability and time series visibility. Finally, this style unity coefficient is written into the content monitoring dashboard in real time for the brand operator to view and analyze the style consistency level of the brand's current content at any time. This not only enhances the scientificity and systematicness of style management but also provides a key decision-making basis for subsequent triggering of style jumps and optimizing creative strategies, significantly improving the intelligent control ability of brand content.

[0044] S106. When the style unity coefficient exceeds the preset risk threshold, construct a multi-dimensional style space mapping graph based on the current brand tone state, calculate the jump path of the style coordinates in the style space according to the style unity coefficient value, and dynamically adjust the jump amplitude of the style coordinates in combination with the current brand tone unity intensity, so as to achieve differential control and creative guidance of the subsequent content style; In this embodiment, for the step of when the style unity coefficient exceeds the preset risk threshold, the system constructs a multi-dimensional style space mapping graph and calculates the style jump path to achieve differential control and creative guidance of the subsequent content style, which specifically includes the following steps. This step aims to solve the problems of aesthetic fatigue and decreased communication efficiency caused by the long-term excessive consistency of the brand content style, and through dynamic jumps in the style space, reasonably guide the subsequent content to produce appropriate changes, enhancing the vitality of brand communication and user attraction.

[0045] When the system detects that the style consistency coefficient exceeds a preset risk threshold (for example, greater than 0.85), it indicates that the brand's recently released content exhibits excessive convergence in visuals, language, and typography. This triggers a style jump warning mechanism, constructing an interpretable and controllable multidimensional style space mapping based on the brand's historical tonality data, target audience profile, product attributes, and distribution channels. This style space uses various stylistic feature dimensions as axes, such as color tendency (warm / cold, neutral), visual tension (contrast, dynamic), composition (symmetrical / asymmetrical), language style (rational / emotional), and semantic emotion (positive / neutral / suggestive). This style space is constructed using an embedding model trained with historical marketing materials and style label data to achieve continuous expression of brand tonality in this multidimensional space.

[0046] Based on the current value range of the style unity coefficient, the style center coordinates of the brand's current content are determined in the style space. The direction and path of the style coordinate jump are calculated based on the extent of the risk threshold exceeded. The jump path calculation mechanism is based on the principles of maximizing differences and controlling tonal boundaries. That is, while ensuring that the brand's core tonality is not deviated from, the target area with a certain "stylistic distance" from the current style state is selected as the jump direction. For example, if the current style appears overly calm and conservative (such as low-saturation color schemes and overly rational language), the system will choose to jump to a more lively or emotional area in the style space, thereby achieving a localized shift in the tonal dimension and increasing the expressive tension.

[0047] To prevent tonal fragmentation caused by excessive style shifts, the system introduces the strength of brand tonal unity as a shift adjustment factor, dynamically adjusting the magnitude of style shifts based on this metric. Tonal unity strength can be determined through historical stability analysis and user cognitive feedback modeling, reflecting the brand's ability to maintain consistent tonal consistency. If the brand's tonal stability is high, the system allows for medium or large shifts within the style space to increase creative tension. Conversely, if the brand's tonal sensitivity is high, the system automatically limits the magnitude of the shift, making only minor adjustments within adjacent style areas to prevent brand tonal fragmentation or user cognitive misalignment. This dynamic control mechanism ensures that style shifts are not only directional and purposeful, but also controllable and flexible, ensuring the sustainable evolution of creative change within the framework of brand strategy.

[0048] Update the style parameter configuration of the subsequent content generation model based on the calculated jump coordinates, including but not limited to content template selection, image generation style, text generation corpus bias, color scheme change, etc., so that the next marketing materials such as graphics, copywriting, and short video scripts are automatically generated towards the new style coordinates. Each style change during the jump process will be recorded and visually displayed on the brand content style monitoring dashboard, facilitating the operation team to evaluate adjustment effects, user feedback, and content click-through rate changes and other indicators, and realizing data closed-loop feedback optimization.

[0049] When the system detects that the style unity coefficient exceeds the preset risk threshold, that is, when the brand content shows an excessive consistency and homogenization trend in terms of style levels such as vision, language, and layout, it can timely trigger the style jump mechanism. By constructing a multi-dimensional style space mapping graph and performing dynamic coordinate adjustment, it guides the subsequent content to achieve a moderate deviation and creative update in style. Specifically, this step first constructs a visual and computable style space in multiple style dimensions (such as color temperature, semantic emotion, composition style, language style, etc.) based on the historical evolution trajectory of the brand tone, the current content state, and the user acceptance threshold. Then, according to the value and change trend of the style unity coefficient, it judges whether the brand content has fallen into the style fatigue area. If so, it calculates a reasonable style coordinate jump path. The system then combines the dynamic factor of the brand tone unity intensity to adjust the amplitude and direction of the jump, ensuring that while achieving a style-differentiated expression, it does not damage the core consistency of the brand tone. Finally, this step drives the content generation system to moderately introduce new style factors, break the cycle of homogenized content, enhance the freshness, emotional tension, and communication attraction of the content, while retaining the brand recognition, and realizing the dynamic balance between "unity" and "diversity". This intelligent style jump mechanism effectively solves the problem that traditional AI-generated content lacks changes and is difficult to continuously attract users after multiple rounds of delivery, and is an important technical means to enhance the brand communication power and content life cycle.

[0050] Through the above AI-driven full-marketing content design method for product packaging planning, it is possible to effectively solve key problems such as content style homogenization, user aesthetic fatigue, and slow system response in the existing technology, and achieve the introduction of dynamic changes and creative diversity in brand content creation while maintaining the unity of tone. By constructing a brand tone mother vector as a style anchor point, combining multi-modal material analysis, a time-sensitive style evaluation mechanism, and an adaptive style jump strategy, this method not only realizes the real-time monitoring of the content style consistency state, but also has the ability to dynamically adjust the style according to the content fatigue risk, ensuring that the content generation system can intelligently perceive changes in the communication environment and make strategic responses. Ultimately, this method improves the flexibility and intelligence level of content creation, enhances the content attractiveness and user interaction efficiency of the brand in the multi-channel communication process, significantly extends the life cycle and communication value of brand content, and has strong practicality and promotion prospects.

[0051] Only some exemplary embodiments of the present invention have been described above by way of illustration. Undoubtedly, for those of ordinary skill in the art, the described embodiments can be modified in various different ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. An AI-driven full marketing content design method for product packaging planning, characterized in that, It includes the following steps: S101. Obtain the product packaging design draft, extract color, composition, font, and semantic features based on a deep convolutional neural network, and fuse them to form a brand tone mother vector of the brand style tone; S102. Call the multimodal parsing module to extract visual and language features consistent with the dimensions of the brand tone mother vector frame by frame or sentence by sentence from the multimodal marketing materials recently released by the brand, and generate a set of marketing material feature vectors; S103. Calculate the cosine similarity between each marketing material feature vector and the brand tone mother vector to obtain a style consistency score, and construct a style consistency score sequence in chronological order; S104. Based on a preset time-series decay weight function, weight the style consistency score sequence to construct a dynamic consistency weight matrix; S105. Merge the dynamic consistency weight matrix through an adaptive aggregation algorithm to obtain an initial style unity intensity value, and normalize and output a style unity coefficient; S106. When the style unity coefficient exceeds the preset risk threshold, construct a multi-dimensional style space mapping graph based on the current brand tone state, calculate the jump path according to the style unity coefficient, and dynamically adjust the jump amplitude in combination with the tone unity intensity.

2. The AI-driven full marketing content design method for product packaging planning according to claim 1, characterized in that, The generation steps of the brand tone mother vector include: Obtain the product packaging design draft image and input it into a pre-trained deep convolutional neural network model for preprocessing operations to extract color and composition features; Identify the text area in the packaging image through a text detection and recognition model, and extract font features by a font classification network; Input the recognized text into a semantic embedding module to extract language tone and emotional semantic features; Fuse and encode color, composition, font, and semantic features, and weight them through an attention mechanism to generate a brand tone mother vector.

3. The AI-driven full marketing content design method for product packaging planning according to claim 1, wherein The steps of calling the multimodal parsing module to extract marketing material feature vectors include: Collect image, video, and text-based marketing materials released by the brand within a preset time range, and perform standardization processing; Sample video materials frame by frame, extract key frame images, and extract visual features through a deep convolutional neural network; Divide text-based materials sentence by sentence, and extract semantic emotion and language tone features through a natural language processing module; Integrate image and text features in the order of materials to generate a set of marketing material feature vectors consistent with the dimensions of the brand tone mother vector.

4. The AI-driven full marketing content design method for product packaging planning according to claim 1, characterized in that, The steps of constructing a style consistency score sequence include: Perform L2 normalization on each feature vector in the set of marketing material feature vectors and the brand tone mother vector; Call the cosine similarity function to calculate the similarity score between each feature vector and the brand tone mother vector; Sort the similarity scores in the original release time order of the materials; Generate a style consistency score sequence in chronological order.

5. The AI-driven full marketing content design method for product packaging planning according to claim 1, wherein, The steps of constructing a dynamic consistency weight matrix include: Calculate the relative time offset according to the timestamp information of each item in the style consistency score sequence; Call the time-series decay weight function model to calculate the time decay weight for each score; Multiply each style consistency score by the corresponding weight value to generate a weighted score; Map all weighted scores to a set of dynamic consistency weight matrices arranged in the order of materials.

6. The AI-driven full marketing content design method for product packaging planning according to claim 5, wherein The time-series decay weight function is an exponential decay function, and its expression is W(t) = e^(-λt), where t is the time offset and λ is the decay coefficient. The exponential decay function is used to ensure that new materials correspond to high weights and old materials correspond to low weights.

7. The AI-driven full marketing content design method for product packaging planning according to claim 1, characterized in that, The steps for generating the style unity coefficient include: Receiving and parsing the dynamic consistency weight matrix, and extracting the time-series weighted feature scores of each style sub-dimension; Performing weighted fusion or non-linear merging on the column feature values to obtain the initial style unity intensity value; Using a normalization algorithm to map the initial intensity value to the style unity coefficient between 0 and 1.

8. The AI-driven full marketing content design method for product packaging planning according to claim 1, characterized in that The steps for calculating the jump path of the style unity coefficient and dynamically adjusting the jump amplitude in combination with the tonality unity intensity include: When the style unity coefficient exceeds the preset risk threshold, constructing a multi-dimensional style space mapping diagram based on the brand tonality state; Determining the current style coordinates in the style space according to the style unity coefficient and calculating the jump path; Dynamically adjusting the jump amplitude in combination with the brand tonality unity intensity; Updating the style parameters of the content generation model according to the jump coordinates.

Images