Dynamic tweet intelligent generation method and system based on visual feature recognition

Abstract

The invention relates to the technical field of personalized recommendation, in particular to a dynamic tweet intelligent generation method and system based on visual feature recognition, and the method comprises the steps: obtaining visual content uploaded by a user, recognizing the visual content through an intelligent discrimination algorithm, obtaining a processing flow corresponding to the visual content, and obtaining a static data set and a dynamic data set; constructing a data analysis processing model to perform product feature recognition and advanced semantic analysis on the static data set to obtain a first multi-dimensional feature hierarchy; performing dynamic data segmentation, product feature recognition, time sequence feature extraction and advanced semantic analysis on the dynamic data set to obtain a second multi-dimensional feature hierarchy; analyzing the first multi-dimensional feature hierarchy and the second multi-dimensional feature hierarchy to generate a tweet generation parameter data set; and generating a personalized tweet by combining the tweet style adaptation model with a tweet generation parameter data set and cue word guidance, and performing quality adjustment on the personalized tweet through a tweet quality evaluation mechanism.

Description

Technical Field

[0001] The present invention relates to the technical field of personalized recommendation, and specifically to a dynamic tweet intelligent generation method and system based on visual feature recognition. Background Art

[0002] In recent years, with the popularization of social media, tweets have become an important means of product promotion. Through concise text and vivid visual elements, tweets can quickly attract users' attention, enhance brand awareness and product sales. However, existing tweet generation methods have significant bottlenecks in dynamic scenario response, complex feature parsing, and human-machine collaborative control, making it difficult to meet the requirements of real-time, personalization, and content consistency.

[0003] In terms of visual feature recognition technology, existing solutions mainly rely on static image analysis and are difficult to adapt to real-time changes in dynamic scenarios. For example, in product promotion, users' visual attention points change over time and scenarios, and existing technologies cannot capture and respond to these changes in a timely manner, resulting in tweet content that does not match users' interests. In addition, the complexity of visual features also poses challenges to recognition and parsing. Especially when fusing multi-modal data, existing methods often struggle to accurately extract and fuse fine-grained features, affecting the quality and attractiveness of tweet content.

[0004] The application of dynamic tweet generation technology in product promotion also has deficiencies. Existing solutions mostly use preset templates or simple rules to generate tweets and lack the ability to dynamically adjust based on real-time data. This makes it difficult to achieve personalized customization of tweet content for different users and scenarios, affecting the promotion effect. At the same time, when dealing with multi-modal data, existing technologies often only stay at shallow semantic matching, such as object name recognition, and lack the ability to perform fine-grained parsing of complex visual elements, such as program code indentation and special symbols, restricting the richness and attractiveness of tweet content.

[0005] In the product promotion scenario, tweets need to closely combine the visual features of products with text descriptions to form coherent and attractive content. However, the style and character image of tweets generated by existing technologies are easily affected by training data biases, especially in long sequence generation, where logical jumps or aesthetic mismatches are likely to occur, such as facial deformation of characters and sudden changes in scene styles. It is difficult for users to dynamically intervene in the generation process through visual markers, such as background colors and graphic identifiers, resulting in a large deviation between the generated content and users' intentions. Existing systems mainly focus on one-way generation and lack a closed-loop optimization mechanism based on visual feedback.

[0006] The application of human-machine collaborative control in tweet generation also faces challenges. Existing solutions mostly rely on manual intervention, which is inefficient and costly. In dynamic scenarios, it is difficult for humans to respond to changes in real time, resulting in lagging tweet generation and missed promotion opportunities. At the same time, existing technologies lack intelligent support in human-machine collaboration and cannot achieve efficient human-machine interaction and collaborative control, restricting the flexibility and real-time nature of tweet generation.

[0007] Therefore, a dynamic tweet intelligent generation method and system based on visual feature recognition are proposed. Summary of the Invention

[0008] The purpose of the present invention is to provide a dynamic tweet intelligent generation method and system based on visual feature recognition. By obtaining the visual content uploaded by users, identifying the visual content through an intelligent discrimination algorithm, obtaining the processing flow corresponding to the visual content, and obtaining a static data set and a dynamic data set; constructing a data analysis and processing model to perform product feature recognition and high-level semantic analysis on the static data set to obtain a first multi-dimensional feature level; performing dynamic data segmentation, product feature recognition, time series feature extraction and high-level semantic analysis on the dynamic data set to obtain a second multi-dimensional feature level; analyzing the first multi-dimensional feature level and the second multi-dimensional feature level to generate a tweet generation parameter data set; generating personalized tweets through a tweet style adaptation model in combination with the tweet generation parameter data set and prompt word guidance, and adjusting the quality of the personalized tweets through a tweet quality evaluation mechanism.

[0009] To achieve the above object, the present invention provides the following technical solutions: A dynamic tweet intelligent generation method based on visual feature recognition, comprising: Obtaining the visual content uploaded by users, identifying the visual content through an intelligent discrimination algorithm, obtaining the processing flow corresponding to the visual content, and obtaining a static data set and a dynamic data set; the visual content includes static content and dynamic content; the processing flow includes static data processing and dynamic data processing; Constructing a data analysis and processing model to perform product feature recognition and high-level semantic analysis on the static data set to obtain a first multi-dimensional feature level; performing dynamic data segmentation, product feature recognition, time series feature extraction and high-level semantic analysis on the dynamic data set to obtain a second multi-dimensional feature level; analyzing the first multi-dimensional feature level and the second multi-dimensional feature level to generate a tweet generation parameter data set, including product type, feature keywords, audience characteristics, shooting method; Generating personalized tweets through a tweet style adaptation model in combination with the tweet generation parameter data set and prompt word guidance, and adjusting the quality of the personalized tweets through a tweet quality evaluation mechanism.

[0010] Preferably, the static data processing obtains a static data set by performing feature extraction, size standardization, and color space conversion analysis on static content; the dynamic data processing obtains a dynamic data set by performing key frame extraction, scene segmentation, motion trajectory analysis, and temporal correlation calculation analysis on dynamic content.

[0011] Preferably, the data analysis and processing model includes a multi-dimensional product feature level acquisition layer and a tweet generation parameter generation layer; The multi-dimensional product feature level acquisition layer analyzes the static data set and the dynamic data set to obtain a first multi-dimensional feature level and a second multi-dimensional feature level; the tweet generation parameter generation layer analyzes the first multi-dimensional feature level and the second multi-dimensional feature level to generate a tweet generation parameter data set.

[0012] Preferably, the specific steps of the first multi-dimensional feature level include: constructing a configurable AI large model call interface, and the interface supports dynamically switching to different product recognition large models without modifying the architecture; Analyze the static data set according to the configurable AI large model, extract low-level visual feature data, including color distribution features, material features, shape features, and spatial layout features; construct a physical attribute set based on the low-level visual feature data, including product size, color, material, style, and structural features; Perform high-level semantic feature extraction on the static data set through high-level semantic analysis, identify product categories, brand information, functional attributes, and design styles; establish a keyword system for the product, including a physical attribute keyword layer, a functional characteristic keyword layer, and an emotional connection keyword layer; through a combination of fuzzy matching and exact matching, intelligently match the extracted product features with a preset product information library to confirm the product identity and supplement the product background information; integrate the physical attribute set, high-level semantic features, and keyword system to form the first multi-dimensional feature level.

[0013] Preferably, the specific steps of the second multi-dimensional feature level include: performing frame-level feature extraction on the key frame sequence in the dynamic data set to identify the product physical features and scene elements in each frame; through temporal feature extraction, perform time dimension analysis on the frame sequence, calculate the inter-frame change rate and change pattern; retrieve the shooting technique features matching the current video feature pattern from the shooting technique feature library to identify the shooting angle, light processing, camera movement, and transition techniques; extract the dynamic interaction features in the video content, including product usage methods, operation processes, and function displays; integrate the frame-level features, temporal features, shooting technique features, and interaction features to form the second multi-dimensional feature level.

[0014] Preferably, the specific method for obtaining the tweet generation parameter dataset is as follows: Through a feature fusion algorithm, the first multi-dimensional feature level and the second multi-dimensional feature level are weighted and fused; Based on the fused feature level, the product type and sub-category are determined, and feature keywords are extracted; According to the product type and feature keywords, the target audience characteristics are inferred, including demographic characteristics, interest preferences, and consumption psychology; The shooting technique is obtained from the shooting technique characteristics of the second multi-dimensional feature level; Through a priority ranking algorithm, the tweet generation parameters are scored and ranked according to marketing data, and an influence weight value is set for each parameter; A structured tweet generation parameter dataset including product type, feature keywords, audience characteristics, and shooting techniques is generated.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By introducing a dual processing mechanism for static and dynamic visual content, the present invention effectively improves the intelligent level of tweet generation. Using visual recognition algorithms and multi-dimensional feature extraction models, in-depth analysis is carried out on the image and video content uploaded by users. It not only identifies the basic attributes of products, but also can extract semantic level information from images, and combines the temporal characteristics and shooting technique analysis in videos to construct a comprehensive product understanding system. On this basis, through a tweet style adaptation model, by integrating prompt word guidance and market positioning information, personalized tweet texts that are suitable for product characteristics and attractive are automatically generated, realizing a closed-loop process from content recognition to language output, significantly improving the relevance, expression quality, and commercial value of tweet content, and meeting the needs of personalized marketing communication in different scenarios.

[0016] 2. In the process of tweet generation, the present invention introduces a fusion mechanism of "the first multi-dimensional feature level" and "the second multi-dimensional feature level", conducts multi-dimensional feature mining on static images and dynamic video content respectively, and outputs a structured tweet generation parameter set through a feature fusion algorithm, thus significantly enriching the dimension of tweet content generation. Using dynamic visual information such as frame-level changes, camera language, and interaction behaviors in videos, the interaction methods, performance effects, and situational characteristics of products in actual use can be identified, and then the behavior preferences and consumption psychology of the product target audience can be inferred. This multi-source feature fusion strategy enables the system to flexibly adapt to different types of products and diverse communication scenarios. At the same time, combined with the priority ranking mechanism of tweet generation parameters, it improves the pertinence and effectiveness of content dissemination, thereby effectively expanding the user reach and conversion rate and enhancing the marketing effectiveness of tweets.

[0017] 3. The present invention introduces a tweet quality evaluation mechanism, which supports the quality evaluation and dynamic adjustment of generated personalized tweets, and realizes the automatic iterative optimization of high-quality content. This mechanism is based on an evaluation index system with multiple dimensions, which can comprehensively score the initially generated tweets, and automatically trigger parameter adjustment and content reconstruction for tweets that do not reach the quality threshold. By introducing quality threshold control and multi-round generation mechanisms, the present invention can continuously optimize tweet content until it meets the preset dissemination quality standards, ensuring that the final output content has commercial dissemination value. In addition, this mechanism also facilitates users to fine-tune the tweet style according to the marketing scenario, further improving the usability and professionalism of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a schematic flowchart of a dynamic tweet intelligent generation method based on visual feature recognition provided by the present invention; Figure 2 is a schematic structural diagram of a dynamic tweet intelligent generation system based on visual feature recognition provided by the present invention; Figure 3 is a schematic structural diagram of a data analysis and processing model provided by the present invention; Figure 4 is a schematic flowchart of tweet intelligent generation provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0020] The present invention provides a dynamic tweet intelligent generation method based on visual feature recognition. This method is applied to a dynamic tweet intelligent generation system based on visual feature recognition. The specific method and system flowcharts are referred to Figure 1 and Figure 2 .

[0021] Embodiment 1 Please refer to Figure 1 , the present invention provides a dynamic tweet intelligent generation method based on visual feature recognition, and the technical solution is as follows: Obtain the visual content uploaded by the user, identify the visual content through an intelligent discrimination algorithm, obtain the processing flow corresponding to the visual content, and obtain a static data set and a dynamic data set; the visual content includes static content and dynamic content; the processing flow includes static data processing and dynamic data processing; Construct a data analysis and processing model to perform product feature recognition and advanced semantic analysis on the static data set, and obtain the first multi-dimensional feature level; perform dynamic data segmentation, product feature recognition, time-series feature extraction and advanced semantic analysis on the dynamic data set, and obtain the second multi-dimensional feature level; analyze the first multi-dimensional feature level and the second multi-dimensional feature level to generate a tweet generation parameter data set, including product type, feature keywords, audience characteristics, and shooting methods. Generate personalized tweets through a tweet style adaptation model in combination with the tweet generation parameter data set and prompt word guidance, and perform quality adjustment on the personalized tweets through a tweet quality evaluation mechanism.

[0022] Specifically, the static data processing obtains a static data set by performing feature extraction, size normalization, and color space conversion analysis on static content; the dynamic data processing obtains a dynamic data set by performing key frame extraction, scene segmentation, motion trajectory analysis, and time-series correlation calculation analysis on dynamic content.

[0023] In this embodiment, by performing feature extraction, size normalization, and color space conversion analysis on static content, the consistency and comparability of images in subsequent feature recognition and semantic analysis processes can be effectively improved, and the analysis errors caused by image size, composition differences, or color deviations can be reduced, thereby enhancing the robustness and accuracy of static data processing; at the same time, by performing key frame extraction, scene segmentation, motion trajectory analysis, and time-series correlation calculation on dynamic content, representative and time-sensitive feature information can be extracted from dynamic visual materials such as videos, improving the expression ability and adaptability of dynamic data in scenarios such as time-series product display and interactive behavior analysis. The collaborative integration of this static and dynamic data processing process not only realizes the multi-dimensional and multi-granularity extraction of visual content information, but also provides a more comprehensive, rich, and semantically clear product feature basis for subsequent tweet generation, which is beneficial to the intelligent generation of personalized content and the optimization of dissemination effects.

[0024] Further, the data analysis and processing model includes a multi-dimensional product feature level acquisition layer and a tweet generation parameter generation layer. Refer to Figure 3 ; The multi-dimensional product feature level acquisition layer obtains the first multi-dimensional feature level and the second multi-dimensional feature level by analyzing the static data set and the dynamic data set; the tweet generation parameter generation layer generates a tweet generation parameter data set by analyzing the first multi-dimensional feature level and the second multi-dimensional feature level.

[0025] In this embodiment, by setting up a multi-dimensional product feature hierarchy acquisition layer and a tweet generation parameter generation layer, an efficient conversion from raw visual data to structured tweet generation elements is achieved. Among them, the multi-dimensional product feature hierarchy acquisition layer can extract product features of different dimensions and semantic levels from static data and dynamic data respectively, ensuring the comprehensiveness and hierarchy of product information and enhancing the understanding ability of aspects such as product attributes, scenarios, and user behaviors; while the tweet generation parameter generation layer extracts the key parameter information required for constructing high-quality tweets, such as product type, feature keywords, and audience characteristics, based on these multi-dimensional feature hierarchies, thus effectively connecting the semantic bridge between visual content and language expression. This model structure helps to enhance the system's semantic analysis ability for complex product information and achieve higher content fitness and communication accuracy when generating tweets, significantly improving the intelligent level and user acceptance of personalized content generation.

[0026] Further, the specific steps of the first multi-dimensional feature hierarchy include: constructing a configurable AI large model call interface, which supports dynamically switching to different product recognition large models without modifying the architecture; Analyze the static data set according to the configurable AI large model, and extract low-level visual feature data, including color distribution features, material features, shape features, and spatial layout features; construct a physical attribute set based on the low-level visual feature data, including product size, color, material, style, and structural features; Extract high-level semantic features from the static data set through high-level semantic analysis, identify product categories, brand information, functional attributes, and design styles; establish a product keyword system, including a physical attribute keyword layer, a functional characteristic keyword layer, and an emotional connection keyword layer; through a combination of fuzzy matching and exact matching, intelligently match the extracted product features with the preset product information library to confirm the product identity and supplement the product background information; integrate the physical attribute set, high-level semantic features, and keyword system to form the first multi-dimensional feature hierarchy.

[0027] In this embodiment, by introducing a configurable AI large model call interface, the flexible switching of different product recognition models is achieved without modifying the overall system architecture, thereby enhancing the adaptability to multi-category and multi-style products and improving the flexibility and scalability of static visual content processing. Combining low-level visual feature analysis and high-level semantic extraction, the system can not only accurately identify the physical attributes of products but also deeply mine semantic information such as their brands, functions, and design styles, significantly improving the comprehensiveness and depth of product feature recognition. Further, by constructing a multi-level product keyword system, the organizational structure and semantic expression ability of product feature information are effectively enhanced, facilitating the subsequent generation of more targeted and attractive tweet content. At the same time, through an intelligent matching mechanism that combines fuzzy matching and exact matching, the product identity can be accurately confirmed and missing information can be supplemented, improving the data integrity and recognition accuracy. The finally formed first multi-dimensional feature level provides a solid data foundation for personalized content generation, helps to achieve high-quality conversion from visual content to language expression, and improves the intelligence level and user experience of the system.

[0028] Further, the specific steps of the second multi-dimensional feature level include: performing frame-level feature extraction on the key frame sequence in the dynamic dataset to identify the product physical features and scene elements in each frame; through temporal feature extraction, analyzing the time dimension of the frame sequence and calculating the inter-frame change rate and change pattern; retrieving the shooting technique features matching the current video feature pattern from the shooting technique feature library to identify the shooting angle, lighting processing, camera movement, and transition techniques; extracting the dynamic interaction features in the video content, including product usage methods, operation processes, and function demonstrations; and integrating the frame-level features, temporal features, shooting technique features, and interaction features to form the second multi-dimensional feature level.

[0029] In this embodiment, by extracting frame-level features from key frame sequences in dynamic data sets, the system can fully capture the physical properties of the product and the scene background contained in each frame of the video, providing a high-precision visual basis for dynamic information analysis. Combined with the temporal feature extraction method, the inter-frame change trend and law are analyzed from the time dimension, which can accurately grasp the performance form and display rhythm of the product at different time nodes, and provide important support for understanding the dynamic evolution of video content. At the same time, the system integrates the shooting technique feature library to match and identify shooting techniques, and can effectively extract the shooting angle, light changes, lens movement and transition style in the video, thereby enhancing the perception of the content expression method. By identifying the interactive features of the product and the characters or environment in the video, the system further enhances the understanding of the product usage scenario and function demonstration, which helps to tap potential user concerns. Finally, the integration of multi-source features to construct the second multi-dimensional feature level not only enriches the dimension of product semantic expression, but also lays the data and context foundation for the subsequent generation of personalized tweets that conform to the dynamic communication logic, and improves the system's intelligent processing and tweet generation capabilities for dynamic visual content.

[0030] Furthermore, the specific method for obtaining the tweet generation parameter data set is as follows: through a feature fusion algorithm, the first multidimensional feature level and the second multidimensional feature level are weightedly fused; based on the fused feature level, product types and subcategories are determined, and feature keywords are extracted; according to the product type and feature keywords, target audience characteristics are inferred, including demographic characteristics, interest preferences and consumer psychology; shooting techniques are obtained from the shooting technique characteristics of the second multidimensional feature level; through a priority sorting algorithm, the tweet generation parameters are scored and sorted in importance according to the marketing data, and an influence weight value is set for each parameter; a structured tweet generation parameter data set containing product types, feature keywords, audience characteristics and shooting techniques is generated.

[0031] In this embodiment, by performing weighted fusion on the first multi-dimensional feature level and the second multi-dimensional feature level, the system can achieve complementary integration of static visual features and dynamic display information, thereby comprehensively and accurately depicting the multi-dimensional attribute image of the product. The fused feature level not only improves the accuracy of product type and sub-category determination, but also can extract feature keywords that are more in line with user perception and expression needs, providing high-quality semantic support for the personalized construction of tweet content. Through further analysis of the product type and keyword set, the system can intelligently infer the demographic characteristics, interest orientations, and consumption psychology of the target user group, effectively improving the matching degree and communication efficiency between the tweet content and the audience. At the same time, the system combines the shooting technique features in the dynamic content to make the generated tweet more in line with the video expression method and visual style, enhancing the communication attraction. By introducing the priority sorting algorithm to introduce marketing data to sort and assign weights to each parameter, it is ensured that the finally generated tweet generation parameter dataset not only has structure and interpretability, but also reflects the strategy optimization effect for the actual marketing scenario, thereby greatly improving the commercial value and communication efficiency of tweet generation.

[0032] Furthermore, the quality adjustment of personalized tweets through the tweet quality evaluation mechanism includes: the evaluation indicators include relevance, attractiveness, readability, brand consistency, and marketing data; automatically adjust the parameters of tweets that do not reach the quality threshold and regenerate the optimized version; implement multiple rounds of tweet iteration optimization until the preset quality requirements are met; output the finally generated high-quality personalized tweets to the user interface.

[0033] In this embodiment, by introducing the tweet quality evaluation mechanism, the personalized tweets are comprehensively evaluated from multiple dimensions, including the relevance to product features and target audience, the attractiveness of content expressiveness and emotional expression, the readability of language structure and information clarity, brand consistency, and marketing data based on historical data and prediction models. This mechanism can effectively identify content that does not meet the expected quality and automatically adjust and optimize the generation parameters, thereby realizing the intelligent reconstruction of personalized tweets. The system supports multiple rounds of automatic iteration optimization processes, continuously improving the content quality, and ensuring that the tweets meet the user's preset standards in terms of style, content, and strategy. The finally output high-quality tweet content can be directly applied to various social media platforms or e-commerce channels, not only improving the user's publishing efficiency, but also significantly enhancing the brand communication effect and user conversion rate.

[0034] Through the multi-dimensional collaborative processing of static and dynamic visual content, the present invention realizes the extraction of rich and accurate product features from images and videos, laying a solid data foundation for subsequent tweet generation. The size standardization, color space conversion, and advanced semantic analysis of static content effectively improve the consistency and depth of feature recognition; the key frame extraction, temporal features, and shooting technique analysis of dynamic content comprehensively capture the diversity and interactivity of products in the time dimension and visual performance. Relying on a configurable large model call interface and a multi-level product keyword system, the system can flexibly adapt to different scenarios and product categories, efficiently generate structured tweet parameters; through feature fusion and priority ranking algorithms, it accurately matches the target audience and optimizes marketing strategies, further improving the content fit and dissemination effect. The introduction of a multi-round automated quality assessment and iterative optimization mechanism realizes the overall control of tweets in dimensions such as relevance, attractiveness, readability, and brand consistency, significantly improving the intelligence level and user conversion rate of personalized tweets. The complete schematic diagram is referred to Figure 4 .

[0035] Example Two The present invention provides a dynamic tweet intelligent generation method based on visual feature recognition, which is applied to a dynamic tweet intelligent generation system based on visual feature recognition. For details, refer to Figure 2 ; The present invention provides a dynamic tweet intelligent generation system based on visual feature recognition. This system is applied to product tweet generation scenarios such as e-commerce platforms, social media marketing, and brand promotion. The system includes a data acquisition module, a tweet data analysis and generation module, and a tweet style evaluation and adjustment module. The specific functions and working processes of each module are as follows: The data acquisition module is used to obtain the visual content uploaded by the user, identify the visual content through an intelligent discrimination algorithm, obtain the corresponding processing flow of the visual content, and obtain a static data set and a dynamic data set. In practical applications, the visual content includes product static pictures and dynamic video content; the processing flow includes static data processing and dynamic data processing. For example, users can upload product photos and usage videos of a certain smart watch, and the system will automatically identify the content type and select the corresponding processing flow.

[0036] In this embodiment, the data acquisition module first classifies and identifies the visual content uploaded by the user through the intelligent discrimination algorithm constructed by the deep learning model to determine whether it belongs to static content, such as product main pictures, detail pictures, scene pictures, etc., or dynamic content, such as product demonstration videos, usage tutorials, scene application videos, etc. For static content, the system performs preprocessing operations such as feature extraction, size standardization, and color space conversion to form a structured static data set. The feature extraction includes quantifying the color distribution, texture features, edge contours, and object shapes of the static content; the size standardization includes uniformly converting static content with different resolutions into a preset resolution format; the color space conversion includes converting the RGB color space to the HSV color space to enhance the color semantic analysis ability; For dynamic content, the system performs operations such as key frame extraction, scene segmentation, motion trajectory analysis, and temporal correlation calculation. For example, for a product display video, the system extracts 2-3 key frames per second and makes adaptive adjustments based on the degree of scene change; analyzes the object motion trajectory through the optical flow algorithm, records the display angle changes of the product in the video and the interaction methods during use; combines temporal correlation calculation to identify key segments such as representative function demonstrations and usage effects in the video, and finally forms a dynamic data set containing spatio-temporal information. The key frame extraction obtains video key frames by using the FFmpeg tool to extract the first frame per second; the scene segmentation divides the video into multiple coherent scene units by detecting significant shot transition points in the video content; the motion trajectory analysis includes calculating the position changes and moving speeds of key objects between consecutive frames; the temporal correlation calculation includes analyzing the change rules and correlation intensities of objects, colors, and compositions between frames.

[0037] The tweet data analysis and generation module is used to construct a data analysis and processing model to perform product feature recognition and advanced semantic analysis on the static data set to obtain a first multi-dimensional feature level; perform dynamic data segmentation, product feature recognition, temporal feature extraction, and advanced semantic analysis on the dynamic data set to obtain a second multi-dimensional feature level; analyze the first multi-dimensional feature level and the second multi-dimensional feature level to generate a tweet generation parameter data set, including product type, feature keywords, audience characteristics, and shooting methods.

[0038] In this embodiment, the tweet data analysis and generation module includes two key sub-layers: the multi-dimensional product feature level acquisition layer and the tweet generation parameter generation layer. The multi-dimensional product feature level acquisition layer first constructs a configurable AI large model call interface, which supports flexible calling of product recognition models in different fields. For example, for beauty products, a beauty-specific recognition model can be called, and for electronic devices, an electronic product recognition model can be called, so as to improve the accuracy of product recognition in specific fields.

[0039] Taking a smartwatch as an example, the system extracts low-level visual feature data from the static dataset by invoking the electronic product recognition model, including color distribution features (such as the color of the watch face and the color of the strap material), material features (such as metal, ceramic, plastic, leather, etc.), shape features (such as round or square watch faces), and spatial layout features (such as the position of buttons and the distribution of sensors). Based on these low-level visual features, the system constructs a set of physical attributes, including the watch size, color, material, style (such as sports or business), and structural features (such as the waterproof structure and the position of the heart rate sensor).

[0040] Through advanced semantic analysis, the system further extracts the high-level semantic features of the product, identifies product categories including smart sports watches and brand information; functional attributes including heart rate monitoring, blood oxygen monitoring, and sleep tracking; and design styles, such as simple and modern. The system also establishes a three-layer product keyword system: the physical attribute keyword layer includes, for example, "lightweight design" and "high-definition display screen"; the functional characteristic keyword layer includes, for example, "all-weather heart rate monitoring" and "intelligent health management"; and the emotional connection keyword layer includes, for example, "sense of technology" and "professional sports companion". Through a combination of fuzzy matching and exact matching, the system intelligently matches the extracted features with the preset product information database to confirm the product model and supplement more background information, such as the product release time and market positioning. Finally, the system integrates all the features to form the first multi-dimensional feature level.

[0041] For the dynamic dataset, the system first extracts frame-level features from the key frame sequence to identify changes in the product's physical features and scene elements in each frame. For example, it identifies the function displays of the smartwatch in different usage scenarios, such as the display of sports data during running, the waterproof performance during swimming, and the monitoring status during sleep. Through temporal feature extraction, the system analyzes the inter-frame change rate and change pattern, such as the interface switching speed, the animation transition effect, and the function response time.

[0042] The system also retrieves the shooting technique features that match the current video feature pattern from the pre-established shooting technique feature library, identifies shooting angles such as close-up shots, surround displays, and real-scene applications, the lighting processing of natural / artificial light, the camera movement methods such as steady / tracking / surround, and the transition techniques such as fade-in / fade-out / switch / overlay. In addition, the system extracts the dynamic interaction features in the video, including the interaction methods between the user and the watch, such as swiping operations and button controls; the complete operation process, such as the steps to set an alarm; and the function displays, such as receiving notifications and replying to messages. Finally, the system integrates the frame-level features, temporal features, shooting technique features, and interaction features to form the second multi-dimensional feature level.

[0043] The tweet generation parameter generation layer performs weighted fusion of the first multi-dimensional feature layer and the second multi-dimensional feature layer through a feature fusion algorithm. The algorithm automatically adjusts the weight ratio of static features and dynamic features according to different product categories. For example, fashion products may place more emphasis on static visual effects, while functional products value dynamic usage experiences more. Based on the fused feature layer, the system accurately determines the product type, including "high-end sports smartwatch" and sub-categories such as "professional outdoor sports", and extracts feature keywords, including appearance description words such as "thin and light", "scratch-resistant and wear-resistant"; function description words such as "all-day heart rate monitoring", "GPS positioning", and value description words including "professional sports analysis", "health life butler".

[0044] The system also infers the target audience characteristics based on the product type and feature keywords, including demographic characteristics such as 25 - 45 years old, medium to high income, urban residents; interest preferences such as fitness enthusiasts, outdoor sports participants, health life pursuers; and consumption psychology such as paying attention to quality, pursuing professional performance, and being concerned about health management. From the shooting technique features of the second multi-dimensional feature layer, the system identifies the main shooting methods, such as close-up display, actual usage scenario demonstration, function operation guide, etc.

[0045] Through a priority ranking algorithm, the system scores the importance of each tweet generation parameter based on historical marketing data and sets the influence weight value. For example, for a sports smartwatch, the system may set "professional sports function" as a high-weight keyword and "fashionable appearance" as a secondary-weight keyword. Finally, the system generates a structured tweet generation parameter dataset containing product type, feature keywords, audience characteristics, and shooting techniques, providing data support for personalized tweet generation.

[0046] The tweet style evaluation and adjustment module is used to generate personalized tweets through a tweet style adaptation model in combination with the tweet generation parameter dataset and prompt word guidance, and perform quality adjustment on the personalized tweets through a tweet quality evaluation mechanism.

[0047] The implementation process and quality adjustment process of the personalized tweet include: Build a tweet style adaptation model based on the generated parameter dataset of tweets to intelligently match the product visual features with the text expression style; design a diverse tweet template library, including tweet skeletons with different lengths, structures, and expressions; select the most suitable tweet structure from the template library according to the product type and audience characteristics; design an accurate prompt template to embed the key information in the tweet generation parameter dataset into the prompt structure; call the large AI model to generate the initial draft of the tweet to ensure that the tweet content highly matches the product visual features; perform automated quality detection on the generated tweet through the tweet quality evaluation mechanism, and the evaluation indicators include relevance, attractiveness, readability, brand consistency, and marketing effect; automatically adjust the parameters of the tweet that does not meet the quality threshold and regenerate the optimized version; achieve multi-round iteration and optimization of the tweet until the preset quality requirements are met; output the finally generated high-quality personalized tweet to the user interface and synchronously save the tweet generation record for the continuous learning of the system.

[0048] The tweet style adaptation model is based on deep learning technology and is trained by analyzing historical successful tweet samples. It can adaptively adjust the tone, rhythm, and rhetorical devices of the tweet to match the product visual presentation style.

[0049] The prompt guidance adopts a multi-level structured prompt technology to transform the product type, key features, target audience, and shooting method into an instruction format that the large AI model can efficiently process.

[0050] The tweet quality evaluation mechanism includes four dimensions: semantic relevance evaluation, emotional tone matching evaluation, audience attractiveness evaluation, and brand consistency evaluation. Each dimension is quantitatively scored through a pre-trained quality scoring model.

[0051] In this embodiment, the tweet style adaptation model first selects a basic style template that matches the product type and target audience from the preset tweet style library. For example, for a high-end smartwatch, the professional and concise style may be selected. Combining the tweet generation parameter dataset, the system constructs a customized prompt to guide the generation model to focus on the core selling points of the product and the pain points of the target audience. For example, for a sports smartwatch, the prompt may include guiding information such as "highlight the professional sports monitoring function", "emphasize the long battery life feature", and "use professional terms to establish authority".

[0052] After generating the initial version of the tweet, the system conducts a comprehensive quality evaluation through the tweet quality evaluation mechanism. The evaluation indicators include the relevance to product features (calculating the cosine similarity between the tweet and the product document through the BERT model), content attractiveness (the positive emotion value output by the sentiment analysis API), text readability (automated readability detection tool), brand consistency (judging the similarity with the brand corpus through the LSTM classification model), and expected marketing data (using the XGBoost regression model).

[0053] If certain metrics do not meet the preset quality thresholds, the system will automatically adjust the parameters and regenerate. For example, if the brand consistency score of the initial version of the tweet is low, the system will adjust the brand expression; if the readability score is not ideal, the system will optimize the language structure and expression. The system supports multiple rounds of iterative optimization until all quality metrics meet the preset standards, and finally outputs high-quality personalized tweets to the user interface for the user to use directly or further adjust.

[0054] To verify the effectiveness of the present invention, a comparative experiment was conducted, and the results are shown in Table 1; Table 1 Comparison of the product promotion effects of different tweet generation methods

[0055] As can be seen from Table 1, compared with the traditional template method and the method based on static image generation, the method of the present invention has achieved significant improvements in all key metrics. Especially in terms of click-through rate and conversion rate, it shows that the tweets generated by the present invention have stronger attraction and marketing data, greatly reducing the content creation cost and improving the marketing response speed.

[0056] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A dynamic tweet intelligent generation method based on visual feature recognition, characterized in that, Including: Obtain the visual content uploaded by the user, identify the visual content through an intelligent discrimination algorithm, obtain the processing flow corresponding to the visual content, and obtain a static data set and a dynamic data set; the visual content includes static content and dynamic content; the processing flow includes static data processing and dynamic data processing; Construct a data analysis and processing model to perform product feature recognition and advanced semantic analysis on the static data set to obtain a first multi-dimensional feature level; Perform dynamic data segmentation, product feature recognition, time series feature extraction and advanced semantic analysis on the dynamic data set to obtain a second multi-dimensional feature level; Analyze the first multi-dimensional feature level and the second multi-dimensional feature level to generate a tweet generation parameter data set, including product type, feature keywords, audience characteristics and shooting methods; Generate personalized tweets through a tweet style adaptation model in combination with the tweet generation parameter data set and prompt word guidance, and perform quality adjustment on the personalized tweets through a tweet quality evaluation mechanism.

2. The intelligent dynamic tweet generation method based on visual feature recognition according to claim 1, characterized in that: The static data processing obtains a static data set by performing feature extraction, size standardization and color space conversion analysis on the static content; the dynamic data processing obtains a dynamic data set by performing key frame extraction, scene segmentation, motion trajectory analysis and time series correlation calculation analysis on the dynamic content.

3. The intelligent dynamic tweet generation method based on visual feature recognition according to claim 1, characterized in that: The data analysis and processing model includes a multi-dimensional product feature level acquisition layer and a tweet generation parameter generation layer; The multi-dimensional product feature level acquisition layer obtains a first multi-dimensional feature level and a second multi-dimensional feature level by analyzing the static data set and the dynamic data set; The tweet generation parameter generation layer generates a tweet generation parameter data set by analyzing the first multi-dimensional feature level and the second multi-dimensional feature level.

4. The intelligent dynamic tweet generation method based on visual feature recognition according to claim 1, characterized in that: The specific steps of the first multi-dimensional feature level include: constructing a configurable AI large model call interface, and the interface supports dynamically switching to different product recognition large models without modifying the architecture; Analyze the static data set according to the configurable AI large model, extract low-level visual feature data, including color distribution features, material features, shape features and spatial layout features; construct a physical attribute set based on the low-level visual feature data, including product size, color, material, style and structural features; Advanced semantic feature extraction is performed on the static data set through advanced semantic analysis to identify product categories, brand information, functional attributes, and design styles; a keyword system for products is established, including a physical attribute keyword layer, a functional characteristic keyword layer, and an emotional connection keyword layer; through a combination of fuzzy matching and exact matching, the identified product features are intelligently matched with a preset product information library to confirm the product identity and supplement the product background information; the physical attribute set, advanced semantic features, and keyword system are integrated to form a first multi-dimensional feature level.

5. The intelligent dynamic tweet generation method based on visual feature recognition according to claim 1, wherein: The specific steps of the second multi-dimensional feature level include: performing frame-level feature extraction on the key frame sequence in the dynamic data set to identify the product physical features and scene elements in each frame; through temporal feature extraction, performing time dimension analysis on the frame sequence to calculate the inter-frame change rate and change pattern; retrieving the shooting method features matching the current video feature pattern from the shooting method feature library to identify the shooting angle, light processing, camera movement, and transition techniques; extracting the dynamic interaction features in the video content, including product usage methods, operation processes, and function displays; integrating the frame-level features, temporal features, shooting method features, and dynamic interaction features to form a second multi-dimensional feature level.

6. The intelligent dynamic tweet generation method based on visual feature recognition according to claim 1, wherein: The specific method for obtaining the tweet generation parameter data set: through a feature fusion algorithm, the first multi-dimensional feature level and the second multi-dimensional feature level are weighted and fused; based on the fused feature level, the product type and sub-category are determined, and feature keywords are extracted; according to the product type and feature keywords, the target audience characteristics are inferred, including demographic characteristics, interest preferences, and consumption psychology; the shooting method is obtained from the shooting method features of the second multi-dimensional feature level; through a priority ranking algorithm, the tweet generation parameters are scored and ranked according to the marketing data, and an influence weight value is set for each parameter; a structured tweet generation parameter data set including the product type, feature keywords, audience characteristics, and shooting method is generated.

7. The intelligent dynamic tweet generation method based on visual feature recognition according to claim 1, wherein: Quality adjustment of personalized tweets through a tweet quality evaluation mechanism includes: the evaluation indicators include relevance, attractiveness, readability, brand consistency, and marketing data; tweets that do not reach the quality threshold are automatically adjusted for parameters and regenerated into an optimized version; multiple rounds of tweet iteration optimization are achieved until the preset quality requirements are met; the finally generated high-quality personalized tweets are output to the user interface.

8. A dynamic tweet intelligent generation system based on visual feature recognition, characterized in that, Including: A data acquisition module, configured to acquire the visual content uploaded by the user, identify the visual content through an intelligent discrimination algorithm, obtain the processing flow corresponding to the visual content, and obtain a static data set and a dynamic data set; the visual content includes static content and dynamic content; the processing flow includes static data processing and dynamic data processing; A tweet data analysis and generation module, which is used to construct a data analysis and processing model to perform product feature recognition and advanced semantic analysis on the static data set, and obtain the first multi-dimensional feature level; Perform dynamic data segmentation, product feature recognition, time-series feature extraction and advanced semantic analysis on the dynamic data set to obtain the second multi-dimensional feature level; Analyze the first multi-dimensional feature level and the second multi-dimensional feature level to generate a tweet generation parameter data set, including product type, feature keywords, audience characteristics, shooting methods; A tweet style evaluation and adjustment module, which is used to generate personalized tweets through a tweet style adaptation model in combination with the tweet generation parameter data set and prompt word guidance, and perform quality adjustment on the personalized tweets through a tweet quality evaluation mechanism.

Images