A social media platform image optimization strategy learning method and system

Abstract

The application provides a social media platform image optimization strategy learning method and system, relates to the technical field of model learning and intelligent decision-making, and comprises the following steps: obtaining historical published images and corresponding information of an account, extracting visual features and constructing an account visual track; determining a stage target visual track according to a to-be-processed image, current business information, an account portrait and a current operation stage; generating a candidate action sequence and predicting a corresponding predicted visual track; selecting a target action sequence according to a visual offset cost, a track recovery cost, a business selling point salience, an image quality and a platform adaptability; executing the target action sequence to obtain a target optimized image, and updating related models according to publishing feedback and an actual visual track. The application can improve the adaptability between a social media platform image optimization strategy and account operation requirements.

Description

Technical Field

[0001] This invention relates to the field of model learning and intelligent decision-making technology, and in particular to a method and system for learning image optimization strategies for social media platforms. Background Technology

[0002] With the development of AI-generated content technology and image processing technology, an increasing number of social media platform operations are using intelligent image processing tools to optimize product images, promotional images, account cover images, and content illustrations. This aims to improve image quality, enhance content expression, and adapt to the display requirements of social media platforms. In practical applications, operators typically need to combine account positioning, content themes, marketing campaigns, and publishing platforms to perform image processing such as cropping, color adjustment, subject enhancement, text layout, background processing, or image compositing.

[0003] Existing image optimization methods typically generate processing results based on single task input, preset style parameters, or text prompts, which can improve image clarity, composition, and content presentation to a certain extent. However, in actual operation, image processing results are also affected by factors such as account positioning, content type, posting purpose, and platform display format. When the same account posts promotional content, daily content, or brand showcase content, there are often specific requirements for the way the subject is highlighted, color representation, text information, and visual style in the image. Existing technologies do not adequately utilize these factors comprehensively, which can easily lead to discrepancies between image processing results and the account's content presentation requirements or current business needs, affecting the applicability of the image processing results. Therefore, how to improve the matching degree between image optimization results on social media platforms and account operation needs has become a technical problem that needs to be solved. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, the technical problem to be solved by this invention is to propose a method and system for learning image optimization strategies for social media platforms, employing the following technical solution: This invention provides a method for learning image optimization strategies on social media platforms, comprising the following steps: S1: Obtain historical images posted by the account, along with their posting time, content type, operational stage, and business objectives. Extract the visual features of each historical image, generate a sequence of the account's visual state in chronological order, and construct the account's visual trajectory. S2: Obtain the image to be processed, current business information, account profile and current operation stage, extract the visual semantic features of the image to be processed, and determine the visual trajectory of the stage target; S3: Input the visual semantic features of the image to be processed, the current account visual state, and the visual trajectory of the stage target into the image optimization strategy model to generate multiple candidate action sequences. The visual state transition prediction model then predicts the next account visual state and the predicted visual trajectory of the subsequent release cycle corresponding to each candidate action sequence. S4: Determine the visual offset cost based on the difference between the predicted visual trajectory and the target visual trajectory of the stage, and determine the trajectory recovery cost based on the compensatory image optimization action required to revert the predicted visual trajectory to the target visual trajectory of the stage; combine the visual offset cost, trajectory recovery cost, business selling point salience, image quality and platform adaptability to calculate the strategy score of each candidate action sequence and select the target action sequence. S5: Execute the target action sequence to obtain the target optimized image, and update the account visual state sequence and account visual trajectory based on the visual features of the target optimized image; S6: Obtain the release feedback of the target optimized image and the actual visual trajectory formed by the subsequent released images. Based on the difference between the predicted visual trajectory and the actual visual trajectory corresponding to the target action sequence and the release feedback, update the visual state transition prediction model and the image optimization strategy model.

[0005] As a further improvement, in step S4, a compensatory image optimization action sequence is generated based on the state difference between the predicted visual trajectory corresponding to each of the above candidate action sequences and the above stage target visual trajectory. The compensatory image optimization action sequence includes compensatory actions for adjusting color features, subject layout features, text layout features or style embedding features. The trajectory recovery cost is determined based on the parameter changes, the area of ​​effect, and the number of actions for each of the aforementioned compensation actions.

[0006] As a further improvement, in step S3, the above-mentioned visual state transition prediction model takes the current account visual state in the above-mentioned account visual state sequence, the visual semantic features of the above-mentioned image to be processed, the above-mentioned current business information, and each of the above-mentioned candidate action sequences as input, and predicts the account visual state corresponding to the subsequent release cycle in turn, so as to form the predicted visual trajectory corresponding to each of the above-mentioned candidate action sequences.

[0007] To further improve this, the image optimization actions in each of the above candidate action sequences include action type, area of ​​action, action parameters, and visual influence vector; The aforementioned visual impact vector is used to characterize the predicted change of each visual feature in the aforementioned account visual state by the corresponding image optimization action.

[0008] As a further improvement, in step S2, the corresponding set of stage visual states is determined from the account visual trajectory based on the current operation stage, and the set of stage visual states is adjusted based on the business objectives and selling point tags in the current business information to obtain the stage target visual trajectory.

[0009] With further improvements, the aforementioned current operational phases include the new product pre-launch phase, the key promotion phase, the daily operation phase, and the brand content phase; The aforementioned new product pre-launch phase, key promotion phase, daily operation phase, and brand content phase each correspond to a preset set of visual states and a range of visual feature adjustments.

[0010] As a further improvement, in step S1, the aforementioned visual features include comprehensive tone features, subject category features, subject layout features, white space features, text area features, text density features, and style embedding features; The visual state of the account is formed by combining the aforementioned visual features according to preset dimensions.

[0011] As a further improvement, in step S6, the trajectory prediction error is determined based on the state difference between the predicted visual trajectory corresponding to the above target action sequence and the above actual visual trajectory. The visual state transition prediction model is updated based on the trajectory prediction error, and the strategy score parameters of the image optimization strategy model are updated based on the trajectory prediction error and the release feedback.

[0012] Another aspect of the present invention provides a learning system for image optimization strategies on social media platforms, comprising: The visual trajectory construction module is used to obtain historical images posted by the account and their corresponding posting time, content type, operational stage and business goals, extract the visual features of historical images, and generate the account's visual state sequence and account visual trajectory. The phase target determination module, connected to the visual trajectory construction module, is used to acquire the image to be processed, current business information, account profile and current operation phase, extract the visual semantic features of the image to be processed, and determine the visual trajectory of the phase target. The strategy generation and prediction module, connected to the stage target determination module, is used to generate multiple candidate action sequences based on the visual semantic features of the image to be processed, the current account visual state, and the visual trajectory of the stage target, and to predict the predicted visual trajectory corresponding to each candidate action sequence. The strategy evaluation module, connected to the strategy generation and prediction module, is used to determine the visual offset cost and trajectory recovery cost based on the predicted visual trajectory, and to determine the target action sequence by combining the salience of business selling points, image quality and platform adaptability. The image processing module, connected to the strategy evaluation module, is used to execute the target action sequence to obtain the target optimized image, and update the account visual state sequence and account visual trajectory based on the visual features of the target optimized image. The model update module is connected to the visual trajectory construction module, the strategy generation and prediction module, and the image processing module, respectively. It is used to update the visual state transition prediction model and the image optimization strategy model based on the feedback of the release of the target optimized image, the actual visual trajectory, and the predicted visual trajectory.

[0013] Compared with the prior art, the beneficial effects of the present invention are: Firstly, this invention extracts visual features from historically posted images of an account, along with their corresponding posting times, content types, operational stages, and business objectives, and constructs the account's visual trajectory. This allows the image optimization process to learn strategies by combining the visual presentation of the account's existing content. Compared to methods that process only a single image or a single task input, this improves the matching degree between the image optimization results and the account's content presentation requirements.

[0014] Secondly, this invention determines the visual trajectory of the stage target based on current business information, account profile, current operation stage, and visual semantic features of the image to be processed, and generates multiple candidate action sequences, which can form multiple optional optimization paths for the current image processing task; then, it predicts the subsequent visual changes corresponding to each candidate action sequence through a visual state transition prediction model, so that the selection of image optimization strategy takes into account both the current image processing effect and the subsequent presentation of account content, thereby improving the adaptability of image processing strategy.

[0015] Third, this invention determines the visual offset cost based on the difference between the predicted visual trajectory and the stage target visual trajectory, and determines the trajectory recovery cost based on the compensatory image optimization actions required to bring the predicted visual trajectory back to the stage target visual trajectory. It also determines the target action sequence by combining the salience of business selling points, image quality, and platform adaptability. This can reduce the probability that the image optimization results deviate from the account content presentation requirements. At the same time, it updates the visual state transition prediction model and the image optimization strategy model based on the release feedback of the target optimized image and the actual visual trajectory, thereby improving the learning accuracy and stability of subsequent image optimization strategies. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 This is a flowchart of the steps of the method of the present invention; Figure 2 This is a schematic diagram of the structural framework of the system of the present invention. Detailed Implementation

[0018] To facilitate understanding by those skilled in the art, the technical solution of the present invention will now be described in further detail.

[0019] This invention provides a method for learning image optimization strategies on social media platforms, applied to social media platform account content operation platforms. This platform can provide services such as product image optimization, promotional image generation, account cover creation, content image generation, and image version management to brand merchants, content operation teams, or chain stores. Social media platform accounts can be short video platform accounts, text and image content platform accounts, brand community accounts, or other accounts used for continuously publishing image content.

[0020] The platform can maintain historical data for various social media platform accounts, including posted images, posting time, content type, operational stage, business objectives, account profile, image processing records, and posting feedback. Account profiles can include information such as account positioning, content direction, target audience, brand tone, commonly used expression methods, content posting preferences, and applicable platforms. Business objectives can include goals such as product promotion, event pre-launch promotion, special sales events, brand display, user interaction, or knowledge dissemination. Posting feedback can include data such as impressions, click-through rate, interaction rate, save rate, conversion rate, dwell time, results of manual image selection, and results of manual rollback.

[0021] The platform can also call image understanding models to extract the subject category, subject position, color distribution, text area, text density, white space, human posture, background elements, composition relationship and style embedding features from the image; and call image generation models or image editing models to perform image optimization actions such as color adjustment, subject enhancement, cropping, background replacement, image compositing, digital model fusion, text layout and cover generation.

[0022] In this embodiment, the account visual state is used to characterize the visual performance of an account at a specific point in time. The account visual state can be formed by combining comprehensive color tone features, subject category features, subject layout features, white space features, text area features, text density features, and style embedding features according to preset dimensions.

[0023] For example, the first The visual status of the account at each posting point is recorded as follows: , represented as:

[0024] in, Indicates the overall color tone characteristics. Indicates the characteristics of the subject category, Indicates the main layout characteristics, This indicates the characteristic of white space in the image. Indicates the characteristics of text regions. Indicates the density characteristics of characters. This indicates style embedding features.

[0025] like Figure 1 As shown, the method includes the following steps.

[0026] Step S1: Obtain historical images posted by the account, along with their posting time, content type, operational stage, and business objectives. Extract the visual features of each historical image, generate a sequence of the account's visual state in chronological order, and construct the account's visual trajectory.

[0027] Historical images posted by an account can be retrieved from the platform's content library, account asset library, content management system, or image version library. Each account's historical images can be linked to posting time, content type, operational stage, and business objectives.

[0028] Content types can include product display images, product usage scenario images, event promotional images, brand story images, cover images, user interaction images, or illustrations for knowledge explanations. Operational phases can include new product pre-launch phases, key promotional phases, daily operations phases, and brand content phases. Business objectives can include increasing product exposure, reinforcing product selling points, conveying event information, improving content interaction, or maintaining brand image.

[0029] The platform sorts the images posted by an account in history according to the posting time and extracts visual features from each image. For content containing multiple images within the same posting period, the visual features of each image can be obtained separately first. Then, based on the image's position in the content, display duration, or user interaction data, the visual features of each image are fused to obtain the account's visual status for that posting period.

[0030] For example, for restaurant brand accounts, the platform can extract features such as the proportion of warm tones, the main area of ​​the dishes, the position of the dishes in the picture, the proportion of the price text area, the position of the brand logo, and the degree of background blur from historically posted images; for beauty brand accounts, the platform can extract features such as skin tone, the proportion of product close-ups, the posture of the people, the brightness and darkness of the picture, the density of text, and the position of product exposure.

[0031] The platform organizes account visual states into a sequence based on posting time, and constructs an account visual trajectory based on the changing relationships between adjacent account visual states. The account visual trajectory can record the direction and magnitude of change of each visual feature during continuous posting.

[0032] In step S1, the visual features include comprehensive color tone features, subject category features, subject layout features, white space features, text area features, text density features, and style embedding features; the account visual state is formed by combining these visual features according to preset dimensions.

[0033] Comprehensive color tone features can include dominant color category, color saturation, color temperature, brightness distribution, and overall color ratio. Subject category features can include the category and quantity of goods, people, food, scenes, text labels, and graphic elements. Subject layout features can include the position of the subject in the image, the proportion of the subject's area, the relative distance between subjects, and the position of the visual center of gravity. Image white space features can include the proportion of white space, the position of white space, and the color uniformity of white space. Text area features can include the number of text areas, the position of text areas, and the area of ​​text areas. Text density features can include the number of text characters per unit image area or the proportion of text areas. Style embedding features can be obtained by encoding the image using a visual coding model and are used to characterize the overall visual style of the image.

[0034] To reduce the impact of occasional content on an account's visual trajectory, the platform allows users to assign weights based on posting time, content interaction, and manual selection for each historically posted image. Images posted more recently, those with higher interaction rates, and those actively retained by administrators will have a higher weighting when constructing the account's visual trajectory.

[0035] Step S2: Obtain the image to be processed, current business information, account profile and current operation stage, extract the visual semantic features of the image to be processed, and determine the visual trajectory of the stage target.

[0036] The image to be processed can be an original image uploaded by the user, a product image, a photograph, a historical image, or an initial image generated by an image generation model. After receiving the image to be processed, the platform can perform subject detection, scene recognition, text region detection, image quality assessment, visual style encoding, and key selling point location to obtain the visual semantic features of the image to be processed.

[0037] Current business information may include business objectives, core selling points, product categories, product attributes, event information, target user groups, advertising platforms, and content themes. Account profiles may include account positioning, brand tone, content preferences, target audience, historical content style, and material usage preferences.

[0038] For example, if the current business information is "highlight the taste and limited-time offers of new beverages during the promotional period", and the account profile is "youthful beverage brand, commonly uses bright and warm colors, focuses on the main subject of the images, and uses concise text information", then the platform can determine based on the above information that the current task needs to strengthen the beverage subject, highlight the event information, and maintain a visual presentation that is compatible with the existing content of the account.

[0039] A phase-specific visual trajectory describes the expected range of changes in the account's visual state within the current operational phase. A phase-specific visual trajectory may include one or more phase-specific visual states, as well as the permissible range of change between adjacent phase-specific visual states.

[0040] In step S2, the corresponding set of visual states for the current operational stage is determined from the account's visual trajectory, and the set of visual states for the current stage is adjusted according to the business objectives and selling point tags in the current business information to obtain the visual trajectory of the stage objectives.

[0041] Specifically, the platform can filter historical account visual states from the account's visual trajectory that are the same as or have a correlation with the current operational stage, forming a set of stage visual states. Then, based on the current business goals and selling point tags, the visual features in the set of stage visual states are adjusted.

[0042] For example, during the promotional phase, the platform can increase the target range corresponding to the main product area, the proportion of promotional text area, and color saturation; during the brand content phase, the platform can increase the target range corresponding to the proportion of white space, the consistency of background style, and the stability of brand logo position; during the new product pre-launch phase, the platform can increase the target range corresponding to the exposure of partial products, atmosphere expression, and suspenseful text layout.

[0043] The current operational phases include the new product pre-launch phase, the key promotion phase, the daily operation phase, and the brand content phase; the new product pre-launch phase, key promotion phase, daily operation phase, and brand content phase each correspond to a preset set of visual states and a range of visual feature adjustments.

[0044] The preset visual feature adjustment range can be set by operators in the account configuration interface, or it can be automatically obtained based on historical images published during the corresponding operational phase. The platform can dynamically update the visual feature adjustment range based on the account's historical content volume, phase duration, and publishing frequency.

[0045] Step S3: Input the visual semantic features of the image to be processed, the current account visual state, and the stage target visual trajectory into the image optimization strategy model to generate multiple candidate action sequences. The visual state transition prediction model then predicts the next account visual state and the predicted visual trajectory for subsequent release cycles corresponding to each candidate action sequence.

[0046] The image optimization strategy model is used to determine the image optimization actions and the order of action execution applicable to the current task based on the visual semantic features of the image to be processed, current business information, account visual status, and stage target visual trajectory.

[0047] Candidate action sequences may include one or more image optimization actions. Image optimization actions may include color mapping, brightness adjustment, contrast adjustment, subject enhancement, local sharpening, cropping window adjustment, background replacement, background blurring, image compositing, digital model fusion, product position adjustment, text area layout, selling point text generation, cover template matching, and style transfer, among others.

[0048] For example, for a beverage product image, the first candidate action sequence could include "increasing the brightness of the image, enhancing the saturation of the beverage subject, cropping it to the platform's recommended ratio, and placing limited-time promotion text in the lower right area"; the second candidate action sequence could include "maintaining the original color tone, increasing the area of ​​the beverage subject, increasing the background blur, and placing the brand logo in the top area"; and the third candidate action sequence could include "replacing it with a promotional scene background, adding a digital model holding a beverage element, and placing promotional information in the bottom area".

[0049] In step S3, the visual state transition prediction model takes the current account visual state in the account visual state sequence, the visual semantic features of the image to be processed, the current business information, and each candidate action sequence as input, and predicts the account visual state for the corresponding subsequent release cycle in turn, so as to form the predicted visual trajectory corresponding to each candidate action sequence.

[0050] Visual state transition prediction models can be implemented using recurrent neural networks, gated recurrent unit networks, long short-term memory networks, temporal Transformer networks, graph neural networks, or other prediction models capable of processing temporal state data.

[0051] For the Candidate action sequence The visual state transition prediction model can predict the visual state of the current account. Visual semantic features of the image to be processed Current business information and candidate action sequences Predict the account's visual appearance for the next release cycle:

[0052] Where F represents the visual state transition prediction model. This represents the predicted visual state of the next account corresponding to the j-th candidate action sequence.

[0053] The platform can continue to use the predicted visual state of the next account as the input state for subsequent release cycles. By combining the target state at the corresponding time point in the stage target visual trajectory, the platform can sequentially obtain the predicted visual state results of multiple subsequent release cycles, thereby forming the predicted visual trajectory corresponding to the candidate action sequence.

[0054] It should be understood that for subsequent release cycles, the platform will use the predicted account visual state as the input state for the next release cycle, and the target state at the corresponding time point in the stage target visual trajectory as the stage constraint. The predicted visual trajectory is used to characterize the changing trend of the account visual state under the action of the candidate action sequence, and is not used to determine the specific content of the subsequently released image.

[0055] The image optimization actions in each candidate action sequence include action type, area of ​​effect, action parameters, and visual impact vector; the visual impact vector is used to characterize the predicted change of each visual feature in the account's visual state by the corresponding image optimization action.

[0056] The action type identifies the category to which the image optimization action belongs. The area of ​​effect identifies the region in the image to be processed corresponding to the image optimization action; this can be represented by pixel coordinate ranges, region masks, subject identifiers, or text region identifiers. Action parameters represent the execution intensity, direction, or range of the specific action. For example, the action parameters for a color mapping action may include color temperature adjustment values, saturation adjustment values, and brightness adjustment values; the action parameters for a cropping window adjustment action may include cropping ratio, window coordinates, and subject retention ratio.

[0057] The visual impact vector can include changes in overall tone, subject area, subject position, white space ratio, text area, text density, and style embedding. The platform can train visual impact vectors corresponding to various image optimization actions based on historical image processing records, image optimization action parameters, and visual feature changes corresponding to the optimized image.

[0058] Step S4: Determine the visual offset cost based on the difference between the predicted visual trajectory and the target visual trajectory of the stage, and determine the trajectory recovery cost based on the compensatory image optimization actions required to revert the predicted visual trajectory to the target visual trajectory of the stage; combine the visual offset cost, trajectory recovery cost, business selling point salience, image quality and platform adaptability to calculate the strategy score of each candidate action sequence and select the target action sequence.

[0059] Visual offset cost is used to characterize the degree of deviation between the predicted visual trajectory formed after the execution of the candidate action sequence and the target visual trajectory of the stage. The platform can assign corresponding weights to each visual feature according to the importance of visual features such as overall color tone, main layout, text area, text density, and style embedding.

[0060] For the j-th candidate action sequence, the platform can calculate the feature difference between the predicted account visual state and the stage target visual state in the h-th subsequent release cycle, and determine the visual offset cost D_j based on the time weight corresponding to each subsequent release cycle.

[0061] Trajectory recovery cost is used to characterize the cost of compensatory image optimization actions required to restore the subsequent account visual state to the stage target visual trajectory when the candidate action sequence causes the predicted visual trajectory to deviate from the stage target visual trajectory.

[0062] In step S4, a compensatory image optimization action sequence is generated based on the state difference between the predicted visual trajectory corresponding to each candidate action sequence and the target visual trajectory of the stage. The compensatory image optimization action sequence includes compensatory actions for adjusting color features, subject layout features, text layout features, or style embedding features. The trajectory recovery cost is determined based on the parameter change amount, effective area, and number of actions of each compensatory action.

[0063] Specifically, the platform can generate corresponding compensation actions for visual features in the predicted visual trajectory that deviate from the target visual trajectory at a given stage. For example, when a candidate action sequence causes the text density in subsequent images to consistently exceed the text density range of the target visual trajectory, compensatory image optimization actions may include reducing the text area, decreasing the amount of text information, or adjusting the text position. When a candidate action sequence causes the overall color tone of subsequent images to consistently deviate from the target range, compensatory image optimization actions may include adjusting the color temperature, reducing saturation, or restoring the preset primary color ratio.

[0064] The cost of trajectory recovery can be determined based on the parameter changes of each compensation action, the area of ​​effect, the number of actions, and the execution dependencies between actions. The greater the parameter changes, the larger the area of ​​effect, and the more compensation actions, the higher the corresponding trajectory recovery cost.

[0065] Business selling point salience is used to characterize the visibility of products, event information, human actions, or core text information related to the current business objective in the image being processed. The platform can determine the business selling point salience based on the area proportion, clarity, contrast, positional priority, text readability, and relevance to the main image subject of the selling point object within the image.

[0066] Image quality can include sharpness, noise level, brightness uniformity, subject integrity, edge naturalness, text readability, and the naturalness of image composition. The platform can determine image quality using image quality assessment models, rule-based assessment models, or quality prediction models trained on manually evaluated samples.

[0067] Platform fit is used to characterize the degree of match between candidate optimized images and the display requirements of the target social media platform. Platform fit can include image size ratio, cover safety area, risk of text obscuring, integrity of the main subject display, matching degree of the platform's recommended image size, and matching degree of content review rules.

[0068] For the Given a candidate action sequence, the platform can calculate the strategy score as follows: :

[0069] in, This represents the saliency of the business selling point corresponding to the j-th candidate action sequence. Indicates image quality, Indicates platform compatibility. Indicates the cost of visual offset. Indicates the cost of trajectory recovery. , , , and This indicates the preset weighting coefficient.

[0070] The platform selects the candidate action sequence with the highest policy score that also meets the image quality threshold, platform adaptability threshold, and visual offset cost threshold as the target action sequence. If multiple candidate action sequences have the same or similar policy scores, the platform may prioritize the candidate action sequence with lower trajectory recovery cost.

[0071] Step S5: Execute the target action sequence to obtain the target optimized image, and update the account visual state sequence and account visual trajectory based on the visual features of the target optimized image.

[0072] The image processing module performs corresponding processing on the image to be processed sequentially based on the action type, area of ​​effect, and action parameters in the target action sequence. For example, the image processing module can first perform subject enhancement and color adjustment, and then perform cropping window adjustment, background replacement, or text area layout; or it can first perform image compositing or digital model fusion, and then perform color coordination and cover layout based on the fusion result.

[0073] The platform can assign version numbers to target optimized images and save the image to be processed, target action sequence, target optimized image, execution time, image processing parameters and strategy score, so as to facilitate subsequent version backtracking, manual selection recording and model training.

[0074] After the target optimized image is generated, the platform extracts the visual features of the target optimized image and writes these visual features into the account visual state sequence as the account visual state for the current publishing cycle. At the same time, the platform updates the account visual trajectory based on the relationship between the newly written account visual state and the previous account visual state.

[0075] Step S6: Obtain the release feedback of the target optimized image and the actual visual trajectory formed by the subsequent released images. Based on the difference between the predicted visual trajectory and the actual visual trajectory corresponding to the target action sequence and the release feedback, update the visual state transition prediction model and the image optimization strategy model.

[0076] After optimizing the image for publication, the platform can obtain feedback from the social media platform's data interface, content management system, business statistics system, or manual review interface. This feedback can include impressions, click-through rate, engagement rate, save rate, conversion rate, dwell time, manual selection results, manual rollback results, and content review results.

[0077] The actual visual trajectory formed by subsequent image releases can be obtained by acquiring images released by accounts within a preset number of release cycles after the target optimized image is released, and generating a sequence of subsequent account visual states according to the visual feature extraction method in step S1.

[0078] In step S6, the trajectory prediction error is determined based on the state difference between the predicted visual trajectory and the actual visual trajectory corresponding to the target action sequence; the visual state transition prediction model is updated based on the trajectory prediction error; and the strategy scoring parameters of the image optimization strategy model are updated based on the trajectory prediction error and the published feedback.

[0079] Specifically, the platform can calculate the differences between visual features such as overall color tone, main layout, text area, text density, and style embedding for the account visual state corresponding to the posting period in the predicted visual trajectory and the actual visual trajectory, and summarize them according to preset weights to obtain the trajectory prediction error.

[0080] The smaller the trajectory prediction error, the more accurate the visual state transition prediction model is in predicting subsequent visual changes caused by candidate action sequences. The platform can use the current account visual state, visual semantic features of the image to be processed, current business information, target action sequence, predicted visual trajectory, and actual visual trajectory as training samples to update the model parameters of the visual state transition prediction model.

[0081] The platform can also calculate actual strategy benefits based on published feedback. For example, it can determine business revenue based on click-through rate, interaction rate, conversion rate, manual selection results, and review results. For target action sequences with high business revenue, good image quality, high platform compatibility, and small trajectory prediction errors, the platform can increase the strategy scoring parameters corresponding to the action type, action parameters, and action combinations. For target action sequences with low business revenue, high trajectory recovery costs, or numerous manual rollback attempts, the platform can decrease the corresponding strategy scoring parameters.

[0082] In one specific embodiment, a beverage brand account is in a promotional phase and needs to publish a promotional image for a new beverage. The platform obtains historical images published by the account within recent periods, extracts the overall color tone, beverage subject position, text density, background elements, and style embedding features corresponding to each image, and constructs the account's visual state sequence and visual trajectory.

[0083] The platform acquires images of beverages to be processed, current business information, and account profiles. Current business information includes highlighting new product flavors, limited-time offers, and the target platform's recommended image format; account profiles include a preference for youthful content, warm color schemes, centered subject matter, and concise text. Based on the visual state set corresponding to each promotional stage, and combined with the selling point tags corresponding to new product flavors and limited-time offers, the platform determines the target visual trajectory for each stage.

[0084] The image optimization strategy model generates multiple candidate action sequences. One candidate action sequence includes increasing the brightness of the beverage subject, improving the overall color saturation, cropping the image to the ratio recommended by the target platform, and placing promotional information at the bottom. Another candidate action sequence includes replacing the promotional background, adding a digital model holding a beverage, and placing brand information and promotional information at the top and bottom, respectively.

[0085] The visual state transition prediction model predicts the account's visual state for subsequent release cycles corresponding to each candidate action sequence. The platform found that while candidate action sequences incorporating digital models and multi-regional text layouts have high selling point salience, they cause the main body layout, text density, and style embedding features to deviate from the stage's target visual trajectory, resulting in high trajectory recovery costs. Therefore, the platform selected candidate action sequences corresponding to main body enhancement, color adjustment, cropping, and bottom promotional information layout as the target action sequences.

[0086] The platform executes the target action sequence to obtain the target optimized image, and after the image is published, it obtains the click-through rate, interaction rate, conversion rate, and manual selection results. The platform continues to obtain the actual visual trajectory formed by subsequent published images, compares the actual visual trajectory with the predicted visual trajectory corresponding to the target action sequence, and updates the visual state transition prediction model and image optimization strategy model based on the trajectory prediction error and publishing feedback.

[0087] like Figure 2 As shown, another aspect of the present invention provides a learning system for image optimization strategies on social media platforms, including a visual trajectory construction module, a stage goal determination module, a strategy generation and prediction module, a strategy evaluation module, an image processing module, and a model update module.

[0088] The visual trajectory construction module is used to obtain historical images posted by an account and their corresponding posting time, content type, operational stage, and business objectives, extract the visual features of historical images, and generate a sequence of account visual states and an account visual trajectory.

[0089] The phase target determination module is connected to the visual trajectory construction module. It is used to acquire the image to be processed, current business information, account profile and current operation phase, extract the visual semantic features of the image to be processed, and determine the visual trajectory of the phase target.

[0090] The strategy generation and prediction module is connected to the stage target determination module. It is used to generate multiple candidate action sequences based on the visual semantic features of the image to be processed, the current account visual state, and the visual trajectory of the stage target, and to predict the predicted visual trajectory corresponding to each candidate action sequence.

[0091] The strategy evaluation module is connected to the strategy generation and prediction module. It is used to determine the visual offset cost and trajectory recovery cost based on the predicted visual trajectory, and to determine the target action sequence by combining the salience of business selling points, image quality and platform adaptability.

[0092] The image processing module is connected to the strategy evaluation module. It is used to execute the target action sequence to obtain the target optimized image, and update the account visual state sequence and account visual trajectory based on the visual features of the target optimized image.

[0093] The model update module is connected to the visual trajectory construction module, the strategy generation and prediction module, and the image processing module, respectively. It is used to update the visual state transition prediction model and the image optimization strategy model based on the feedback of the release of the target optimized image, the actual visual trajectory, and the predicted visual trajectory.

[0094] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, any modifications or substitutions to the present invention without departing from the inventive concept should fall within the protection scope of the present invention.

Claims

1. A method for learning image optimization strategies on social media platforms, characterized in that, Includes the following steps: S1: Obtain historical images posted by the account, along with their posting time, content type, operational stage, and business objectives. Extract the visual features of each historical image, generate a sequence of the account's visual state in chronological order, and construct the account's visual trajectory. S2: Obtain the image to be processed, current business information, account profile and current operation stage, extract the visual semantic features of the image to be processed, determine the corresponding stage visual state set from the account visual trajectory according to the current operation stage, and adjust the stage visual state set according to the business objectives and selling point tags in the current business information to obtain the stage target visual trajectory. S3: Input the visual semantic features of the image to be processed, the current account visual state, and the visual trajectory of the stage target into the image optimization strategy model to generate multiple candidate action sequences. The visual state transition prediction model then predicts the next account visual state and the predicted visual trajectory of the subsequent release cycle corresponding to each candidate action sequence. S4: Determine the visual offset cost based on the difference between the predicted visual trajectory and the stage target visual trajectory, and determine the trajectory recovery cost based on the compensatory image optimization action required to revert the predicted visual trajectory to the stage target visual trajectory. By comprehensively considering visual offset cost, trajectory recovery cost, business selling point salience, image quality, and platform compatibility, a strategy score is calculated for each candidate action sequence, and the target action sequence is selected. S5: Execute the target action sequence to obtain the target optimized image, and update the account visual state sequence and account visual trajectory based on the visual features of the target optimized image; S6: Obtain the release feedback of the target optimized image and the actual visual trajectory formed by the subsequent released images. Based on the difference between the predicted visual trajectory and the actual visual trajectory corresponding to the target action sequence and the release feedback, update the visual state transition prediction model and the image optimization strategy model.

2. The social media platform image optimization strategy learning method according to claim 1, characterized in that, In step S4, a compensatory image optimization action sequence is generated based on the state difference between the predicted visual trajectory corresponding to each candidate action sequence and the stage target visual trajectory. The compensatory image optimization action sequence includes compensatory actions for adjusting color features, subject layout features, text layout features, or style embedding features. The trajectory recovery cost is determined based on the parameter changes, the area of ​​effect, and the number of actions of each compensation action.

3. The method for learning image optimization strategies on social media platforms according to claim 1, characterized in that, In step S3, the visual state transition prediction model takes the current account visual state in the account visual state sequence, the visual semantic features of the image to be processed, the current business information, and each candidate action sequence as input, and predicts the account visual state for the corresponding subsequent release period in turn, so as to form the predicted visual trajectory corresponding to each candidate action sequence.

4. The social media platform image optimization strategy learning method according to claim 3, characterized in that, The image optimization actions in each of the candidate action sequences include action type, region of action, action parameters, and visual influence vector; The visual impact vector is used to characterize the predicted change in each visual feature of the account's visual state caused by the corresponding image optimization action.

5. The method for learning image optimization strategies on social media platforms according to claim 1, characterized in that, The current operational phase includes the new product pre-launch phase, the key promotion phase, the daily operation phase, and the brand content phase; The new product pre-launch phase, the key promotion phase, the daily operation phase, and the brand content phase each correspond to a preset set of visual states and a range of visual feature adjustments.

6. The method for learning image optimization strategies on social media platforms according to claim 1, characterized in that, In step S1, the visual features include comprehensive tone features, subject category features, subject layout features, white space features, text area features, text density features, and style embedding features; The account's visual state is formed by combining the aforementioned visual features according to preset dimensions.

7. The social media platform image optimization strategy learning method according to claim 3, characterized in that, In step S6, the trajectory prediction error is determined based on the state difference between the predicted visual trajectory and the actual visual trajectory corresponding to the target action sequence; The visual state transition prediction model is updated based on the trajectory prediction error, and the strategy score parameters of the image optimization strategy model are updated based on the trajectory prediction error and the release feedback.

8. A learning system for image optimization strategies on social media platforms, characterized in that, include: The visual trajectory construction module is used to obtain historical images posted by the account and their corresponding posting time, content type, operational stage and business goals, extract the visual features of historical images, and generate the account's visual state sequence and account visual trajectory. The stage target determination module, connected to the visual trajectory construction module, is used to acquire the image to be processed, current business information, account profile and current operation stage, extract the visual semantic features of the image to be processed, determine the corresponding stage visual state set from the account visual trajectory according to the current operation stage, and adjust the stage visual state set according to the business goals and selling point tags in the current business information to obtain the stage target visual trajectory. The strategy generation and prediction module, connected to the stage target determination module, is used to generate multiple candidate action sequences based on the visual semantic features of the image to be processed, the current account visual state, and the visual trajectory of the stage target, and to predict the predicted visual trajectory corresponding to each candidate action sequence. The strategy evaluation module, connected to the strategy generation and prediction module, is used to determine the visual offset cost and trajectory recovery cost based on the predicted visual trajectory, and to determine the target action sequence by combining the salience of business selling points, image quality and platform adaptability. The image processing module, connected to the strategy evaluation module, is used to execute the target action sequence to obtain the target optimized image, and update the account visual state sequence and account visual trajectory based on the visual features of the target optimized image. The model update module is connected to the visual trajectory construction module, the strategy generation and prediction module, and the image processing module, respectively. It is used to update the visual state transition prediction model and the image optimization strategy model based on the feedback of the release of the target optimized image, the actual visual trajectory, and the predicted visual trajectory.

Images