Operation picture generation method and system and medium

Abstract

The invention discloses an operation picture generation method and system and a medium. The method comprises the following steps: constructing user image feature preferences; operation pictures are generated in batches; image content correlation quality inspection is executed; carrying out personalized operation picture matching based on an effect feedback mechanism; and displaying the matched operation diagram to the user. According to the invention, automatic generation of the personalized operation picture is realized, and the operation effect is improved.

Description

Technical Field

[0001] The present invention relates to the field of operation technology, and in particular to a method, system and medium for generating an operation picture. Background Art

[0002] In the internet marketing field, marketing graphics are a tool for attracting user attention and increasing engagement. Currently, most marketing graphics are pre-designed by designers, who combine marketing content with graphic elements and base their design on the target user group or popular design elements and styles. This approach ensures the image's aesthetic and professionalism, allowing marketing content to be presented to users in a visually appealing way.

[0003] However, existing methods for generating operational images present several challenges. Using fixed images often lacks specificity and personalization, and may not always match the needs and preferences of individual users. A mismatch between image content and user interests can lead to low click-through rates and engagement, directly impacting the conversion rate and effectiveness of operational campaigns. Furthermore, designing specialized images for each operational content increases design costs and time investment. Summary of the Invention

[0004] In response to the above technical problems, the present invention provides a method, system and medium for generating operation pictures, which realize the automatic generation of personalized operation pictures and improve the operation effect.

[0005] A first aspect of the present invention provides a method for generating an operation picture, comprising: Construct user image feature preferences; Batch generate operation pictures; Perform quality control related to image content; Personalized operational image matching based on the effect feedback mechanism; and The matched operation diagram is displayed to the user.

[0006] In a possible implementation, constructing the user image feature preference includes: Extract operational image features; Obtain user behavior data on the operation diagram; Calculate the user's behavioral preference score for the operation map; and Construct user image feature preference vector.

[0007] In a possible implementation, extracting the operation picture features includes extracting the color, color style, artistic style, and entity object categories and entity object labels in the picture content of the operation picture.

[0008] In a possible implementation, the batch generating of operation pictures includes: Analyze operational content and generate illustration requirements; Select the raw image feature range and generate feature combinations; Constructing large model prompts for raw images; and Call pre-trained multimodal large models to generate operation images in batches.

[0009] In a possible implementation, performing image content-related property inspection includes: Construct quality inspection requirement prompt words; Performing quality control related to image content; and Process unqualified pictures.

[0010] In one possible implementation, the personalized operation image matching based on the effect feedback mechanism includes: Perform content preference feature matching; Perform collaborative filtering; Remove poorly performing images; and Apply the rule strategy to determine the final matching image.

[0011] In a possible implementation, performing content preference feature matching includes calculating the similarity between the user preference feature vector and the operation map feature vector, and selecting N pictures with the highest similarity.

[0012] In a possible implementation, applying the rule strategy to determine the final matching image includes defining a strategy rule to limit the maximum number of occurrences of the same image feature value in a certain dimension in consecutive matches of the same user.

[0013] A second aspect of the present invention provides an operation picture generation system, the system comprising: A user image feature preference building module, used to build user image feature preferences; Operation picture batch generation module, used to generate operation pictures in batches; Image content related quality inspection module, used to perform image content related quality inspection; A personalized operation image matching module, used to perform personalized operation image matching based on the effect feedback mechanism; and The operation diagram display module is used to display the matched operation diagram to the user.

[0014] According to a third aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a computer, the method according to the first aspect of the embodiment of the present invention is executed.

[0015] In summary, compared with the prior art, the present invention has at least one of the following beneficial technical effects: by establishing user image feature preferences and performing personalized operational image matching based on these preferences, the attractiveness and relevance of operational images to users are increased; by batch-generating operational images and performing content-related quality checks, the efficiency and quality of operational image generation are improved; through personalized matching based on an effect feedback mechanism, dynamic optimization of operational images is achieved, improving operational effectiveness; and by applying rule-based strategies, the diversity of operational images displayed to users is ensured, avoiding visual fatigue. This present invention significantly improves the personalization and operational effectiveness of operational images, while reducing operational design costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flowchart of an embodiment of the method for generating an operation picture of the present invention.

[0017] Figure 2 This is a flowchart of another embodiment of the method for generating an operation picture of the present invention.

[0018] Figure 3 This is a flowchart of another embodiment of the method for generating an operation picture of the present invention.

[0019] Figure 4 This is a flowchart of another embodiment of the method for generating an operation picture of the present invention.

[0020] Figure 5 This is a flowchart of another embodiment of the method for generating an operation picture of the present invention.

[0021] Figure 6 This is a structural diagram of an embodiment of the operation picture generation system of the present invention. DETAILED DESCRIPTION

[0022] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts shall fall within the scope of protection of the present invention.

[0023] It should be understood that the terms "first," "second," and "third," etc. in the claims, specifications, and drawings of the present disclosure are used to distinguish different objects rather than to describe a specific order. The terms "include" and "comprising" used in the specifications and claims of the present disclosure indicate the presence of the described features, wholes, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components, and / or their collections. It should also be understood that the terms used in this disclosure specification are for the purpose of describing specific embodiments only and are not intended to limit the present disclosure.

[0024] Reference Figure 1 , an embodiment of the present invention discloses a method for generating an operation picture, comprising the following steps.

[0025] S100: Constructing user image feature preferences. Specifically, feature extraction is performed on historical operational images to construct a multi-dimensional image feature vector containing color, color style (warm and cool tones, saturation, brightness), artistic style, entity object category, and label. User behavior data on these images on platforms such as apps or web pages (such as clicks, browsing, favorites, and likes) is obtained, and preference scores are assigned based on different behaviors. Based on the user behavior score and corresponding image features, combined with mean statistics and time decay coefficients, the user's preference score for each feature dimension is calculated. Ultimately, a personalized image preference vector is formed for each user, providing a basis for subsequent personalized image matching.

[0026] S200: Batch generate operational images. Specifically, a pre-trained large language model is used to analyze the main text of the operational article and automatically generate textual requirements for image matching related to the content. Operations personnel can also manually adjust this text based on this. Next, image feature dimensions and feature value ranges (such as color, style, and physical objects) are selected or included by default. All feature values are permuted and combined to generate multiple sets of image feature combinations. Based on the image matching requirements and each feature combination, corresponding text-based image prompts are constructed. Finally, the prompts are input into the pre-trained multimodal large model to batch generate operational images and record the feature vector of each image for subsequent matching.

[0027] S300: Perform quality inspection of image content. Specifically, the large model performs semantic summarization of operational content and generates corresponding quality inspection prompts based on a preset template. Each generated operational image and its corresponding quality inspection prompt are input into the multimodal large model. The model determines whether the image content meets the operational theme requirements and returns a "compliant" or "non-compliant" result. For images judged as "non-compliant", the model can choose to directly remove them or regenerate replacement images based on the original prompts and repeat the quality inspection process until all images pass quality inspection.

[0028] S400: Personalized operational image matching based on the performance feedback mechanism. Specifically, the similarity between the user's image feature preference vector and the feature vector of each operational image is calculated to select the most matching image. Next, collaborative filtering is used to reference the behavioral data of other users with similar preferences to the target user to further supplement the recommended images. Finally, based on historical behavioral feedback, images with low user preference scores are eliminated to avoid repeated display of low-conversion content. Furthermore, a rule-based policy intervention mechanism can be introduced to limit the frequency of repeated display of similar images within a certain period of time, ensuring that image recommendations are both accurate matches and diverse and fresh, thereby achieving better personalized recommendation results.

[0029] S500: Display the matched operation diagram to the user. Specifically, based on the final personalized matching result determined in step S400, the operation diagram that meets the user's preferences and has passed quality inspection is dynamically displayed to the user through various platforms such as the app, website, and vehicle-mounted display. This serves as an accompanying diagram for the operation content, encouraging users to click, browse, and participate, thereby improving the conversion rate of the operation activity and user experience.

[0030] The present invention constructs user image feature preferences, combines them with large models to automatically generate diverse, high-quality operational pictures, and ensures the consistency of images and content through multimodal quality inspection. Finally, based on a personalized matching mechanism, it accurately recommends pictures that best suit users' interests. This not only significantly improves user click-through rate and participation, and improves the conversion effect of operational activities, but also greatly reduces manual design costs, enhances user experience and brand stickiness, and achieves a dual improvement in operational efficiency and personalized services.

[0031] Further, refer to Figure 2 As an embodiment of the present invention, step S100, constructing user image feature preferences, includes: S110, extracting operation picture features; S120, obtaining user behavior data on the operation map; S130, calculating the user's behavior preference score for the operation map; and S140, constructing a user image feature preference vector.

[0032] Specifically, step S110 includes: constructing image feature dimensions, and extracting features of each dimension of the image to ultimately form a complete feature vector of the target image.

[0033] The image feature dimensions are constructed, including but not limited to the following dimensions: 1. Color (such as red, orange, yellow, green, cyan, blue, purple, colorful, etc.); 2. Color style (e.g., cool / warm colors, high / medium / low saturation, high / medium / low brightness); 3. Artistic style (such as realism, watercolor, cyberpunk, flat painting, realism, minimalism, two-dimensional, Chinese style, illustration, etc.); 4. Entity object categories and entity object labels in the image content (for example, if the image contains a car entity object, it can be divided into car entity object labels such as sports car, SUV, sedan, and commercial vehicle; if the image contains an animal entity object, it can be divided into animal entity object labels such as cat, dog, and bird).

[0034] Corresponding to the above dimensions, the features of each dimension of the image are extracted and finally form a complete feature vector of the target image, including color feature extraction, color style feature extraction, artistic style feature extraction, and entity object category and entity object label extraction in the picture content.

[0035] Color feature extraction involves defining K basic colors (e.g., K = 7, where the seven basic colors are red, orange, yellow, green, cyan, blue, and purple) and calculating which basic color each pixel in the image is closest to. An example calculation method is as follows.

[0036] Take the RGB value of the nth pixel to construct a three-dimensional vector (R n ,G n ,B n ), calculate the vector similarity with the RGB three-dimensional vector of each basic color, if the RGB three-dimensional vector of the mth basic color is (R 基础色m ,G 基础色m ,B 基础色m ), then the similarity between the two colors is: ; Calculate the similarity between the target pixel and each basic color in turn. The one with the smallest similarity value represents the closest basic color. After calculating the similar basic colors of all pixels, the proportion of each basic color is counted. One or the top N basic colors whose proportion is greater than a certain threshold can be defined as one or N theme colors of the image. When the proportion of all basic colors is less than a specific threshold, it can be considered that there is no obvious theme color feature and it can be regarded as a "colorful" color.

[0037] Color style feature extraction includes cold / warm color extraction, saturation type (high / medium / low) extraction, and brightness type (high / medium / low) extraction.

[0038] Cool / warm color extraction involves calculating the average color RGB vector for the entire image: (R 平均 ,G 平均 ,B 平均 )=( ); Among them, m is the total number of pixels, and the GRB three-dimensional vector of the n-th pixel is (R n ,G n ,B n ); Calculates the difference between the red and blue values in the image's average color | R 平均 - B 平均 |, and compare it with a certain threshold. If the difference is greater than the threshold and R 平均 If the difference is greater than the threshold and B 平均 If the difference is smaller than or equal to the threshold, it is considered a neutral color.

[0039] Saturation type (high / medium / low) extraction involves calculating the average color RGB vector of the entire image: (R 平均 ,G 平均 ,B 平均 )=( ); Among them, m is the total number of pixels, and the GRB three-dimensional vector of the n-th pixel is (R n ,G n ,B n ); Calculate the average color (R 平均 ,G 平均 ,B 平均 ) The maximum and minimum values of the three RGB channels: max_val = max(R 平均 ,G 平均 ,B 平均 ), min_val = min(R 平均 ,G 平均 ,B 平均 ); Calculate saturation: If max_val = 0, the average color is black (0,0,0), and the saturation is calculated as 0; if max_val ≠ 0, the saturation S = S is divided into three intervals, corresponding to high, medium and low saturation respectively; for example, S∈[0,0.3) is low saturation; S∈[0.3,0.7) is medium saturation; S∈[0.7,1] is high saturation.

[0040] Luminosity type (high / medium / low) extraction involves calculating the average color RGB vector of the entire image: (R 平均 ,G 平均 ,B 平均 )=( ); Where m is the total number of pixels, and the GRB three-dimensional vector of the n-th pixel is (Rn ,G n ,B n ); Calculate the average color (R 平均 ,G 平均 ,B 平均 ) The maximum and minimum values of the three RGB channels: max_val = max(R 平均 ,G 平均 ,B 平均 ), min_val = min(R 平均 ,G 平均 ,B 平均 ); Calculate lightness L= , L is divided into 3 intervals, namely high, medium and low brightness; for example: L∈[0,0.3), it is low brightness; L∈[0.3,0.7), it is medium brightness, and S∈[0.7,1] is high brightness.

[0041] Artistic style feature extraction can be processed using deep learning models, especially convolutional neural networks (CNN) or its variants such as VGG, ResNet, etc. These models perform well in image feature extraction.

[0042] First, we manually annotated and collected a sample library containing a variety of artistic styles (such as realism, watercolor, cyberpunk, flat painting, realism, minimalism, anime, and Chinese style). Each style contained a large number of representative images for training deep learning models.

[0043] During training, the neural network model automatically extracts features from the image. These features may include line thickness, color usage, light and shadow treatment, and other features related to artistic style. It also learns to distinguish image features from different artistic styles.

[0044] After training, the deep learning model is able to classify the style of the input image and output its most likely artistic style label.

[0045] In some other implementations, the artistic style of an image can be directly determined by using a multimodal pre-trained large model.

[0046] Extracting entity object categories and entity object labels from the image content includes the following steps: 1. Data preparation: Collect a large amount of image data containing various physical objects, including cars (such as sports cars, SUVs, sedans, and commercial vehicles), animals (such as cats, dogs, and birds), and other physical object categories that may appear in operational images. Label these images to clearly identify the location, outline range, and major and specific sub-category labels of the physical objects in each image;

[0047] 2. Image preprocessing: Preprocess image data, including image scaling, normalization, and denoising, to improve image quality and reduce the computational burden during model training; 3. Target detection model training: Choose an appropriate object detection algorithm, such as YOLO (You Only Look Once) or Faster R-CNN (Regions with Convolutional Neural Networks), which are highly efficient and accurate in object detection. Use the prepared annotated image data to train the object detection model, enabling it to accurately identify physical objects in the image and output their categories and location information (such as coordinates and edge contours).

[0048] 4. Use the trained target detection model for target detection: Input the operational images to be feature extracted into the trained object detection model to identify the coordinate location, edge contour range, and category of the entity objects in the image; 5. Subclass label recognition model training: For each identified entity object category, further train a classification model for the subcategory label. For example, for the car category, train the model to distinguish subcategory labels such as sports cars, SUVs, and sedans. You can choose a suitable deep learning architecture, such as convolutional neural network (CNN), for image feature extraction and classification; Design the network structure, including the input layer, multiple convolutional layers (with pooling layers), fully connected layers, and output layers; Use the softmax function in the output layer to convert the network output into a probability distribution, indicating the possibility of identifying each subclass; 6. Use the trained subclass label recognition model to perform subclass label recognition: Based on the range of the entity object in the image obtained by target detection in step 3, the entity object image is captured, and the subclass label recognition model that has been trained is used to perform subclass label recognition.

[0049] After extracting the features of each dimension mentioned above, the complete feature vector of the target image can be constructed.

[0050] Furthermore, step S120, obtaining user behavior data on the operational map, includes: with user authorization, through data tracking, reporting user behavior events when users interact with target operational content on platforms such as apps and web pages. These behaviors include clicks, browsing, favorites, and likes, and recording data such as the behavior type, behavior event, user, and operational content in which the behavior occurred.

[0051] Furthermore, step S130, calculating the user's behavior preference score for the operation diagram includes: Convert the user's behavior on the target operation image into a preference score. For example: click behavior + 1 point, browsing time on the details page > 10 seconds + 2 points, adding to favorites + 3 points, liking + 3 points, etc. Finally, calculate the user's cumulative behavior score for the target operation image. This can be used to construct the user's behavior preference score for all historical operation images that have generated operational behavior, and construct the user's behavior preference matrix S for all operation images, where S ij represents the behavioral preference score of user i for image j.

[0052] Furthermore, step 140, constructing a user image feature preference vector includes: The user's preference scores for all historical operation graphs that generated operational behaviors and the features of these historical operation graphs are used to calculate the user's preference vector for each image feature dimension using methods such as mean statistics. To account for the impact of time on user interests, a time decay coefficient can be multiplied when calculating preferences. For example, the specific steps for constructing a user's image feature preference vector are as follows.

[0053] 1. For user i and feature dimension k, calculate the feature value v of user i k The preference score p i (v k ) as follows: ; in, Indicates that user i has characteristic value v k The set of all pictures j that have produced behavioral operations, It represents the number of pictures in the collection.

[0054] 2. Increase the impact of time on interest preferences; In order to reflect the changes in user interests over time, a time decay factor α(t) is introduced when calculating the preference score: α(t)=e −βt ; Where t is the time interval from the last time the user took action on the image to the present, and β is a parameter that adjusts the decay rate. Therefore, the preference score p after the reference time affects the interest preferencei (v k ) is updated to:

[0055] ; Where tij is the time when user i takes action on image j.

[0056] 3. Obtain user image feature preference vector Finally, the preference vector V of user i is constructed i , which contains the preference scores of user i on all feature dimensions: V i = [p′ i (v1) , p' i (v2) , ... , p′ i (v n )]; Where n is the number of feature dimensions, and v1,v2,...,v n Represent the eigenvalues of each feature dimension respectively.

[0057] Further, refer to Figure 3 As an embodiment of the present invention, step S200, batch generation of operation pictures includes: S210, analyze the operation content and generate illustration requirements; S220, selecting a raw image feature range and generating a feature combination; S230, constructing prompt words for the large model of the raw image; and S240: Call the pre-trained multimodal large model to generate operation images in batches.

[0058] Specifically, step S210, analyzing the operation content and generating illustration requirements includes: using a large model to pre-generate illustration requirements, that is, inputting the main text content into the pre-trained large model, and requiring the large model to automatically summarize and analyze the main text of the operation article through prompt words and generate descriptive text for the illustration requirements.

[0059] In some implementations, it is also possible to support operators to manually modify / define basic requirements for image generation, such as elements that must be included, based on the automatically generated image requirements and the actual operation theme content.

[0060] Step S220, selecting a raw image feature range and generating a feature combination includes: 1. Raw image feature range selection: Based on the operator's selection output, obtain the characteristic dimension range (e.g., which colors and cold / warm color schemes need to be generated) and characteristic value range (e.g., the color range can be specifically specified as red, yellow, and green; the color style can be specifically specified as warm and neutral colors) of the operation map generated by each person. You can also select all characteristic ranges by default. 2. Based on the feature range of the generated operation map selected by the operator, exhaustively enumerate all permutations and combinations of feature values to form n possible feature combinations (for example, one possible combination is: [Red] + [Warm Colors] + [High Saturation] + [Medium Brightness] + [Illustration Style] + [SUV]).

[0061] Step S230, constructing the raw image model prompt words includes: Based on the basic requirements for image generation defined by the operator and the permutations and combinations of feature values, a prompt word generation template is created to generate n groups of prompt words for tattoo images (for example, a possible group of prompt words for a tattoo image is: Please draw a picture of the whole family driving together in autumn. Requirements: illustration style, red, warm colors, high saturation, medium brightness. The picture must include: SUV).

[0062] Step S240, calling a pre-trained multimodal large model to batch generate operation images includes: 1. Use the prompt words constructed in S230 to call the pre-trained multimodal large model to batch generate operation pictures. Input the prompt words into the cultural image large model. Generate m pictures for each group of prompt words, and generate a total of n×m pictures. 2. Record the feature combination used to generate each image, and extract features for the specified feature dimension according to the method in S110, and finally form its complete feature vector F j , so that users can subsequently match the images they are most likely to be interested in from all generated images based on image features.

[0063] Further, refer to Figure 4 As an embodiment of the present invention, step S300, performing image content-related property inspection, includes: S310, constructing quality inspection requirement prompt words; S320, performing image content-related quality inspection; and S330, processing unqualified pictures.

[0064] Specifically, because the operational images generated by the large model have a certain degree of randomness, some batches of generated images may deviate from the original operational content. Therefore, during the quality inspection process, the quality inspection requirements for the images and content are input into the multimodal large model, allowing the multimodal large model to determine whether they meet the quality inspection requirements.

[0065] Among them, step S310, constructing the quality inspection requirement prompt words, includes: Use the big model to summarize the main text of the operation content, and then fill it into the quality inspection prompt text template; For example: if the operational content extracted from the main text is summarized as "Participate in the autumn travel prize event and share your autumn self-driving travel stories", then fill it into the prompt word template to form a complete prompt word: "Please judge whether this picture is suitable as a picture for the operational theme 'Participate in the autumn travel prize event and share your autumn self-driving travel stories'? Please tell me directly 'yes' or 'no', and do not give any other unnecessary replies."

[0066] Step S320, performing image content-related quality inspection, including: The generated quality inspection requirement text and the image to be inspected are input into the multimodal large model, and the response result of the large model is obtained and analyzed.

[0067] Step S330, processing unqualified images, includes: For images that are judged as "non-compliant" by the final large model, they can be directly filtered out, or new images can be generated using the original generated prompt words, and the quality inspection operation of the above step S320 can be repeated until all images are replaced with images that meet the quality inspection requirements.

[0068] Further, refer to Figure 5 As an embodiment of the present invention, step S400, performing personalized operation picture matching based on the effect feedback mechanism, includes: S410, performing content preference feature matching; S420, performing collaborative filtering; S430, removes images with poor results; and S440: Apply the rule strategy to determine the final matching image.

[0069] Specifically, personalized operational images are recommended and matched through a combination of content preference feature matching, collaborative filtering, elimination of images with poor performance, and rule strategy intervention.

[0070] In step S410, performing content preference feature matching includes: Based on the above, a feature vector can be constructed for each generated operation diagram, with F j represents the eigenvector of graph j; Based on the above, construct the preference feature vector of each user, V i represents the preference score vector of user i for all feature dimensions; Use cosine similarity to calculate the user preference feature vector V iand each generated operation graph feature vector F j The similarity between them (for multiple images generated by the same set of prompt words, only one is randomly selected for similarity calculation): ; Take the N pictures with the greatest similarity and add them to the list of operation pictures to be displayed to the user.

[0071] Step S420, performing collaborative filtering, includes: 1. User collection construction: Construct a user set U, which includes all users who have generated positive behavioral feedback on the target operation content; 2. Preference similarity calculation: For each user u∈U, calculate its preference feature vector V u The similarity between the preference feature vector Vi of the target user i is: ; 3. Select the N users with the highest similarity: Select the N users with the highest similarity as the similar user set H; 4. Weighted calculation of preference score: According to the behavioral preference score p of user u for operation picture j in the similar user set H u (j), predict the preference score of target user i for each operation map j: ; where w u is the weight of user u, which can be set as similarity sim(V u ,V i ) ; 5. Select the M pictures with the highest scores and add them to the list of operation pictures to be displayed to the user.

[0072] Step S430, removing pictures with poor quality, includes: For the graphs with very low user behavior score preferences after being displayed to users in the past, they are removed to prevent these graphs with low conversion rates from being displayed to more users.

[0073] Step S440, applying the rule strategy to determine the final matching image, including: In order to match the operation diagrams that users are most interested in as much as possible while ensuring the diversity of image features, discovering more user interests and preferences, and avoiding visual fatigue caused by matching operation diagrams with similar features for users, it is necessary to make adjustments through certain rule strategies to finally determine the operation diagrams that match users.

[0074] Specifically, first define the policy rules: The limit is set to a maximum number of occurrences (b) of the same image feature value (e.g., red) in a consecutive image match for the same user. For example, a red image can appear at most once in three consecutive matches for the same user. In other implementations, multiple rules can be defined simultaneously.

[0075] Next, the final matching graph is confirmed based on the rules, including: 1. Merge candidate graph lists: Based on the above points 1 and 2, a list of operational graphs to be displayed to the user in descending order of similarity scores and a list of operational graphs to be displayed to the user in descending order of user preference scores are obtained. The two lists are interleaved and merged into one group; 2. Rule Check: Based on the rules defined above, if the first operation diagram is found to be inconsistent with the rules, the next operation diagram will be checked in turn until a consistent operation diagram is found as the final matching operation diagram for the user. 3. Backup matching: If all the images in the list do not meet the rules, one of the first three images can be randomly selected as the final operation image matched for the user.

[0076] Through the above steps, it is possible to generate and match personalized operational images for each user, ensuring that the images displayed to users not only meet their preferences, but also ensure diversity and visual freshness, thereby improving operational results.

[0077] Reference Figure 6 The embodiment of the present invention also discloses an operation picture generation system, including a user image feature preference construction module 1, an operation picture batch generation module 2, an image content relevance quality inspection module 3, a personalized operation picture matching module 4 and an operation picture display module 5.

[0078] The user image feature preference building module 1 is used to build user image feature preferences.

[0079] Specifically, the user image feature preference construction module 1 first extracts multi-dimensional features from historical operation pictures, and constructs a picture feature vector including color, color style (warm and cool colors, saturation, brightness), artistic style, and the category and label of entity objects in the picture; then, through embedding points, it collects user behavior data such as clicks, browsing, collections, and likes of pictures on platforms such as APP and web pages, and assigns corresponding preference scores to different behaviors; then, combined with the user's behavior score for pictures with specific features, the preference value of each image feature dimension is calculated according to the statistical mean, and a time decay factor is introduced to dynamically reflect changes in user interests; finally, a feature preference vector reflecting the user's image interest tendency is formed, providing an accurate basis for subsequent picture matching and recommendation.

[0080] The operation picture batch generation module 2 is used to generate operation pictures in batches.

[0081] Specifically, the operation picture batch generation module 2 uses a pre-trained large language model to analyze the content of the operation article, and automatically generates text with picture requirements that are consistent with the content. Operation personnel can also make manual adjustments based on this; then select the image feature range to be generated (such as color, color style, artistic style, entity object category, etc.), and the system arranges and combines the selected feature values to construct multiple groups of image feature combinations; then generates corresponding text image prompts based on each group of feature combinations and picture requirements; finally, inputs the prompts into the text image large model, generates operation pictures in batches, and records the feature information of each picture as the basis for subsequent screening and recommendation.

[0082] The image content-related property inspection module 3 is used to perform image content-related property inspection.

[0083] Specifically, the image content-related quality inspection module 3 generates a quality inspection prompt word template based on the content of the operation article, automatically extracts the operation theme through the pre-trained large model and fills it into the prompt word; then the generated prompt word is input into the multimodal large model together with each picture to be tested, allowing it to judge whether the picture content is consistent with the operation theme, and outputs the judgment result of "compliant" or "non-compliant"; for pictures that do not meet the requirements, the system can automatically eliminate them, or regenerate alternative pictures based on the original prompt words and repeat the quality inspection to ensure that the visual content of the pictures finally retained is highly consistent with the operation purpose.

[0084] The personalized operation picture matching module 4 is used to perform personalized operation picture matching based on the effect feedback mechanism.

[0085] Specifically, the personalized operation picture matching module 4 calculates the similarity between the user image feature preference vector and the feature vector of the generated picture, and selects the picture that best matches the user preference; secondly, through the collaborative filtering algorithm, it refers to the behavioral data of other users with similar preferences to the target user to further supplement the recommendation results; at the same time, it eliminates pictures with poor user feedback in history to avoid repeated recommendations of low-conversion content; finally, it introduces a rule strategy intervention mechanism to control the diversity of recommendation results, such as limiting the number of consecutive displays of pictures with the same features, to ensure that users obtain personalized picture display results that are both in line with their interests and visually fresh.

[0086] The operation diagram display module 5 is used to display the matched operation diagram to the user.

[0087] Specifically, based on the final recommendation results determined in the personalized operation picture matching module 4, the operation map display module 5 dynamically displays the operation pictures that are highly matched with the user's interests and preferences and have passed quality inspection to the corresponding users in the operation positions of various platforms such as APP, website, and car computer, as pictures for the operation content, to guide users to generate interactive behaviors such as clicks, browsing, and participation, thereby effectively improving the exposure rate, click-through rate and overall conversion effect of the operation activities.

[0088] The embodiment of the present invention also discloses a readable storage medium.

[0089] A readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the operation picture generation method described in any one of the above embodiments.

[0090] It is understood that computer-readable storage media may include any entity or device capable of carrying a computer program, recording media, USB flash drives, removable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), and software distribution media. A computer program includes computer program code. Computer program code may be in source code form, object code form, an executable file, or some intermediate form. Computer-readable storage media may include any entity or device capable of carrying a computer program code, recording media, USB flash drives, removable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), and software distribution media.

[0091] In certain embodiments of the present invention, an electronic device may include a controller or processor. The controller is a single-chip microcomputer chip that integrates a processor, memory, a communication module, and the like. The processor may refer to the processor contained in the controller. The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), off-the-shelf programmable gate arrays (FPGAs), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, and the like.

[0092] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code comprising one or more executable instructions for implementing a specific logical function or process step, and the scope of the preferred embodiments of the present invention includes alternative implementations in which functions may be performed out of the order shown or discussed, including performing functions in a substantially simultaneous manner or in the reverse order depending on the functions involved, which should be understood by those skilled in the art to which the embodiments of the present invention pertain.

[0093] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the composition and steps of each example according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0094] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for generating an operation picture, characterized in that: include: Construct user image feature preferences; Batch generate operation pictures; Perform quality control related to image content; Personalized operation picture matching based on effect feedback mechanism; as well as The matched operation diagram is displayed to the user.

2. The method for generating an operation picture according to claim 1, wherein: The constructing of user image feature preferences includes: Extract operational image features; Obtain user behavior data on the operation diagram; Calculate the user's behavioral preference score for the operation map; and Construct user image feature preference vector.

3. The method for generating an operation picture according to claim 2, wherein: The extracting of operation picture features includes extracting the color, color style, artistic style of the operation picture and the entity object category and entity object label in the picture content.

4. The method for generating an operation picture according to claim 2, wherein: The batch generation of operation pictures includes: Analyze operational content and generate illustration requirements; Select the raw image feature range and generate feature combinations; Constructing large model prompts for raw images; and Call pre-trained multimodal large models to generate operation images in batches.

5. The method for generating an operation picture according to claim 1, wherein: The performing of image content-related quality inspection includes: Construct quality inspection requirement prompt words; Performing quality control related to image content; and Process unqualified pictures.

6. The method for generating an operation picture according to claim 1, wherein: The personalized operation image matching based on the effect feedback mechanism includes: Perform content preference feature matching; Perform collaborative filtering; Remove poorly performing images; and Apply the rule strategy to determine the final matching image.

7. The method for generating an operation picture according to claim 6, wherein: The performing content preference feature matching includes calculating the similarity between the user preference feature vector and the operation map feature vector, and selecting N pictures with the highest similarity.

8. The method for generating an operation picture according to claim 6, wherein: The applying rule strategy to determine the final matching image includes defining a strategy rule to limit the maximum number of occurrences of the same image feature value in a certain dimension in continuous matching of the same user.

9. An operation picture generation system, characterized in that: The system comprises: A user image feature preference building module, used to build user image feature preferences; Operation picture batch generation module, used to generate operation pictures in batches; Image content related quality inspection module, used to perform image content related quality inspection; A personalized operation image matching module, used to perform personalized operation image matching based on the effect feedback mechanism; and The operation diagram display module is used to display the matched operation diagram to the user.

10. A computer-readable storage medium, characterized in that A computer program is stored thereon, and when the computer program is run by a computer, the operation picture generation method according to any one of claims 1 to 8 is executed.

Images