AI text-to-image real-time printing device

Abstract

The application discloses an AI text-to-image real-time printing device, and relates to the technical field of image processing, which comprises a device layer and a background layer, wherein the device layer comprises a MIC radio, a screen and a printer, the background layer comprises a voice-to-text function and an AI text-to-image model, the AI text-to-image model is provided with an image processing system, the image processing system comprises an information acquisition module and a data analysis module, the information acquisition module is used to acquire multi-source data required by the printing device, including candidate image related data, user feature related data and matching degree related data, the data analysis module is used to determine a qualified material pool based on the candidate image related data, determine a user style preference distribution based on the user feature related data, and determine a final display picture based on the matching degree related data, the application guarantees the quality and theme consistency of the printing content, realizes accurate matching of the content and the preference of the current watching user group, and provides a technical framework for personalized content printing in an offline scene.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to an AI-powered real-time image printing device. Background Technology

[0002] With the rapid growth in demand for personalized content output in offline retail, public services, and cultural tourism exhibitions, AI-powered real-time text-to-image printing devices, as core terminals capable of quickly generating and outputting customized visual content, have become an important technological direction for the intelligent upgrading of offline scenarios. Currently, most mainstream AI text-to-image printing devices are equipped with general-purpose text-to-image models. By inputting a fixed theme, they generate images in corresponding styles and then directly output and print them. Some intelligent devices also connect to hardware such as cameras and sensors to collect scene information and adjust the generated content. This provides basic technical support for the printing of promotional materials and personalized commemorative content in offline scenarios, and has been widely used in scenarios such as printing promotional posters in shopping malls, printing cultural and creative commemorative items in scenic spots, and printing promotional materials for government services.

[0003] Current AI-powered image printing devices still suffer from core technological limitations in practical applications. The content generation logic of current devices lacks a systematic coupling between material quality screening and user preference matching. On one hand, the image generation stage often only uses basic theme matching for screening, failing to consider factors such as original content repetition and stylistic diversity. This easily leads to problems like high homogeneity of generated printed materials and misjudgment of original content due to interference from common elements, making it difficult to guarantee the stability and uniqueness of the output content. On the other hand, the user adaptation stage lacks the ability to accurately identify and quantify the preferences of the current viewing user group. It cannot adjust output strategies according to the content needs of different user groups, resulting in insufficient adaptation between printed content and the actual preferences of users in the current scenario, making it difficult to meet the dynamic and personalized content output needs of offline scenarios. Summary of the Invention

[0004] The purpose of this invention is to provide an AI-generated image real-time printing device, which solves the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an AI text-to-text real-time printing device, comprising a device layer and a back-end layer, wherein the device layer includes a microphone, a screen, and a printer, and the back-end layer includes a speech-to-text converter and an AI image generation model, wherein the AI ​​image generation model is equipped with an image processing system;

[0006] After the user issues a voice command, the microphone picks up the audio, which is then uploaded to the backend layer. After being processed by speech-to-text, it is imported into the AI ​​image model. The AI ​​image model processes and generates an image, which is then analyzed by the image processing system to output the optimal image content and send it to the information management platform. The information management platform then maps the image onto the screen and simultaneously starts the printer to print it.

[0007] The image processing system includes an information acquisition module, a data analysis module, and a database. The information acquisition module acquires multi-source data required for the operation of the printing equipment and sends it to the database for archiving. The database divides the data into candidate image-related data, user feature-related data, and matching degree-related data. A data preprocessing unit in the database cleans the archived data. After cleaning, the data is input into the data analysis module. The data analysis module determines a qualified material pool based on the cleaned candidate image-related data, determines the distribution of user style preferences based on the user feature-related data, and determines the final display image based on the matching degree-related data.

[0008] Optionally, the candidate image related data includes a fixed theme input by the user, key visual features, a public element feature library, a historical display image feature library, a similarity threshold, theme semantic matching degree, image quality score, and style discrimination degree;

[0009] User characteristic-related data includes user distance, user orientation, user dwell time, effective distance threshold, effective dwell time threshold, age-dimension preference probability, clothing-dimension preference probability, original group size mapping probability, interaction dwell rate, and total number of styles;

[0010] Matching data includes pass / fail scores and preference weights.

[0011] Optionally, the data analysis module includes a candidate graph generation module, a user feature recognition module, and an optimal graph matching module.

[0012] Optionally, the candidate image generation module includes a memory repetition correction unit and a pass / fail calculation unit;

[0013] The memory repetition correction unit outputs a corrected repetition score based on key visual features, a public element feature library, a historical display image feature library, and a similarity threshold. The higher the score, the higher the image repetition.

[0014] Optionally, the qualification calculation unit outputs the qualification score of the candidate image based on the topic semantic matching degree, image quality score, style differentiation degree and repetition score, and uses the qualification score as the core indicator to determine whether the image can enter the qualified material pool.

[0015] Optionally, the user feature recognition module includes a group size correction unit and a preference probability calculation unit;

[0016] The group size correction unit outputs the corrected number of valid users based on user distance, user orientation, user dwell time, effective distance threshold, and effective dwell time threshold. The minimum number of valid users is 1, meaning that if there are no valid users, it is treated as a single-person scenario by default.

[0017] Optionally, the preference probability calculation unit outputs the user's preference weight for style s based on the preference probability of age dimension, preference probability of clothing dimension, original group size mapping probability, number of effective users, interaction dwell rate and total number of styles. The sum of the weights of all styles is 1, and the higher the weight, the higher the user's preference for that style.

[0018] Optionally, the processing logic of the optimal graph matching module is as follows:

[0019] First, the pass / fail score and preference weight are processed by boundary conditions to obtain the pass / fail score and preference weight after boundary processing. Then, a logarithmically weighted sum is performed, and finally, a non-linear mapping is performed through an activation function to obtain the final matching score. The image with the highest final matching score among all candidate images is selected as the final display image.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0021] First, this invention breaks through the limitation of traditional duplication calculation that counts all visual features equally. It innovatively separates common element features from original content features, and common elements no longer participate in duplication statistics. This fundamentally avoids the problem of original content being misjudged as duplicate when common elements such as fixed UI and brand logos account for a high proportion. It realizes duplication assessment only for original design content, ensuring that the selected candidate images retain both the necessary common identification elements of the scene and ensure that the original content has a certain degree of uniqueness and freshness. This solves the technical problem of logical distortion in the duplication assessment of existing text-based image systems.

[0022] Second, this invention employs a two-layer evaluation logic that combines hard threshold filtering with multi-index nonlinear weighted fusion. On one hand, it directly eliminates low-quality, irrelevant, and highly repetitive images that do not meet the basic requirements through triple hard thresholds of topic matching degree, image quality, and repetition, ensuring the basic quality of candidate materials. On the other hand, it uses a power product weighting method to fuse four dimensions of indicators: topic matching degree, image quality, style differentiation, and originality. This overcomes the limitation of traditional linear weighting, which can easily lead to high scores in one dimension masking defects in other dimensions. It achieves a comprehensive evaluation based on the "weakest link" effect, ensuring that images entering the candidate pool meet commercial display requirements in all dimensions.

[0023] Meanwhile, by introducing the style differentiation index, we ensure that the images in the candidate pool have sufficient style diversity, providing ample selection space for subsequent personalized matching, and solving the technical problems of unstable quality and monotonous style of Wensheng print materials from the source.

[0024] Third, this invention improves upon existing recognition processes, which are limited by relying solely on distance or dwell time for single-dimensional judgment. It innovatively introduces the concept of user orientation, dynamically adjusting the effective distance threshold. The further the user's orientation deviates from the front of the device, the closer the effective judgment distance. Combined with a soft weighting mechanism based on dwell time, it achieves accurate differentiation between interested users and passersby, fundamentally reducing the probability of miscounting passersby from the side or back. This ensures that the statistically significant number of effective users is completely consistent with the actual user group interested in the device, solving the technical problems of low user recognition accuracy and susceptibility to interference from irrelevant users in existing intelligent display devices.

[0025] Fourth, this invention enables precise quantification of the preferences of user groups of different sizes, ensuring that the generated content is highly adapted to the needs of the current viewing user group, and solving the technical problem that existing generated content is prone to drift and become disconnected from user needs.

[0026] Fifth, by adopting a nonlinear mapping logic that combines logarithmic weighting with activation functions, a dual match can be achieved between image quality standards and user preferences. This ensures that the final printed output is the optimal result from the candidate pool that simultaneously meets both high quality and high adaptability, providing a precise decision-making basis for the personalized content output of AI-powered real-time image printing devices. Attached Figure Description

[0027] Figure 1 This is a flowchart of the present invention;

[0028] Figure 2 This is a flowchart of the image processing system of the present invention. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] Example 1:

[0031] Please see Figure 1This invention provides an AI-generated text and image real-time printing device, comprising a device layer and a backend layer. The device layer includes a microphone, a screen, and a printer. The backend layer includes a speech-to-text converter and an AI-generated image model. The AI-generated image model is equipped with an image processing system. After the user issues a voice command, the microphone collects the audio, which is then uploaded to the backend layer. After speech-to-text processing, the audio is imported into the AI-generated image model. The AI-generated image model processes and generates an image, which is then analyzed by the image processing system to output the optimal image content. This image content is sent to an information management platform, which is then mapped onto the screen and the printer is simultaneously started for printing.

[0032] Example 2, based on the above examples:

[0033] Please see Figure 1 and Figure 2 The image processing system includes an information acquisition module, a data analysis module, and a database. The information acquisition module acquires multi-source data required for the operation of the printing equipment and sends it to the database for archiving. The database divides the data into candidate image-related data, user feature-related data, and matching degree-related data. The data preprocessing unit in the database cleans the archived data. After cleaning, the data is input into the data analysis module. The data analysis module outputs a qualified material pool based on the cleaned candidate image-related data, outputs the user style preference distribution based on the user feature-related data, and outputs the final display image based on the matching degree-related data.

[0034] Specifically, the candidate image data includes a fixed topic t input by the user, candidate images output by the text-based image model, and the i-th key visual feature f of the current style s obtained by extracting features from the candidate images using an image feature extraction algorithm. s,i It contains K such features. A common element feature library F is constructed by pre-collecting non-original public visual elements that regularly appear in the scene, extracting their features using the same image feature extraction algorithm. p The system automatically retains the images displayed each time, extracts features using the same image feature extraction algorithm, and dynamically updates the historical image feature library F with a 7-day validity period. h The preset similarity threshold τ. The topic semantic matching degree C is obtained by calculating the cosine similarity between the input topic text features and the candidate image visual features using the CLIP cross-modal model. s,t The image quality score Q is obtained by directly evaluating candidate images using the BRISQUE no-reference image quality assessment algorithm. s The style discrimination D is obtained by calculating the average cosine distance between the feature vector of the current candidate image and the feature vectors of all other candidate images in the same batch. s .

[0035] The similarity threshold τ is set to 0.85 by default. This value is based on the general judgment standard of cosine similarity: when the cosine similarity of two visual feature vectors is ≥0.85, they can be considered to belong to the same type of visual elements, which can accurately distinguish between common elements and original features. If the value is too high, it will lead to the omission of common elements, and if it is too low, it will misclassify original elements as common elements.

[0036] Furthermore, user characteristic-related data includes user distance d. i The straight-line distance between the i-th user and the display device is directly acquired by an external millimeter-wave radar. The user's orientation is θ. i The body orientation angle of the i-th user is calculated using the front-facing camera and a human pose estimation algorithm, with 90° defined as the direction facing the display device. The user's position is continuously detected by an external millimeter-wave radar, and the time difference between the user entering and leaving the effective range is calculated to obtain the user's dwell time t within the device's effective range. i The system presets an effective distance threshold D based on the device's optimal viewing distance. th The system presets an effective dwell time threshold T based on the minimum duration of content displayed to ordinary users. th The default value is 2 seconds. The age-dimension preference probability P is obtained based on publicly available user profile datasets and after boundary truncation. a (s), obtained from the clothing-style association dataset, is the probability P of clothing dimension preference after boundary truncation. w (s). The original group size mapping probability P obtained based on offline user behavior statistics. org (s), user interaction data is collected through the touchscreen, and the interaction dwell time percentage T is calculated as the proportion of the user's dwell time on the style s preview image to the total browsing time. u,s The total number of styles S is preset based on the style types supported by the text-based image model accessed by the system.

[0037] Where the effective distance threshold D th The default value is 1.5 meters, which is the optimal viewing distance for typical offline display devices: users cannot clearly see the screen content beyond 1.5 meters, therefore it serves as the baseline value for the effective distance. Effective dwell time threshold T th The default setting is 2 seconds. This value is the shortest dwell time for a regular user to become interested in the displayed content. Dwell times of less than 2 seconds are usually unintentional passes by and do not constitute effective attention.

[0038] Furthermore, matching-related data includes the candidate graph qualification score G, calculated using the candidate graph qualification formula. s,t The user's preference weight W for style s is calculated using the user preference probability formula. u,sThe preset minimum value for boundary processing is 10. -6 .

[0039] The minimum value for boundary processing is 10. -6 A minimum value is chosen so that it does not affect the calculation results of normal scores, and can ensure the stability of logarithmic operations, thus avoiding calculation errors in the system.

[0040] Furthermore, the data analysis module includes a candidate image generation module. The core function of the candidate image generation module is to filter out low-quality, homogeneous, and mismatched images, producing a qualified material pool containing 3-5 images. This includes a memory duplication correction unit and a qualification calculation unit. The specific processing logic is as follows:

[0041] Memory repetition correction unit based on key visual features f s,i Public element feature library F p Historical image feature library F h And a similarity threshold τ, outputting the corrected repetition score R. s Repeatability score R s The value range is [0,1], and the higher the score, the higher the image repetition rate.

[0042] Specifically, the calculation logic of the memory repetition correction unit is as follows: First, among all K key features of the candidate image, the features that are common to the feature library F are counted. p The number of features with a similarity greater than the similarity threshold τ. If this number equals K, it means that all features in the image are common elements and there is no original content; therefore, it is directly assigned the value R. s =1 indicates that the image is a highly repetitive and low-quality image with no original content, and will be directly eliminated.

[0043] If non-public features exist, then only the non-public features, i.e., those related to F, will be considered. p Similarity ≤ similarity threshold τ and historical display feature library F h The sum of similarities is used as the numerator. The total number of features K minus the number of common features, which is the sum of similarities with F... p The number of features with similarity greater than the similarity threshold τ is used to calculate the total number of non-common features involved in the calculation, which serves as the denominator. The average similarity is then calculated to obtain R. s Common elements are fixed content within the scene and not part of the original design of the image. Therefore, they are deducted from the denominator to prevent a high proportion of common elements from lowering the average repetition rate and masking the repetition of original content. The specific calculation method is as follows:

[0044]

[0045] Corrected repeatability score R sUsed to quantify the degree of repetition of candidate images, it serves as one of the core indicators for subsequent qualification judgment and directly determines the screening results of original content with low repetition.

[0046] Through the above processing, the limitation of traditional duplication calculation that counts all visual features equally is overcome. It innovatively separates the features of common elements from the features of original content. Common elements no longer participate in duplication statistics, which fundamentally avoids the problem of original content being misjudged as duplicate when the proportion of common elements such as fixed UI and brand logo is high. It realizes duplication assessment only for original design content, ensuring that the selected candidate images retain the necessary common identification elements of the scene, while also ensuring that the original content has a certain degree of uniqueness and freshness. It solves the technical problem of the distortion of the duplication assessment logic in the existing text-based image system.

[0047] Furthermore, the qualification calculation unit is based on the topic semantic matching degree C. s,t Image quality score Q s Style Distinction D s and repeatability score R s Output the qualification score G of the candidate image. s,t Pass rate score G s,t The value range is [0,1], which is the core indicator for determining whether an image can enter the qualified material pool.

[0048] Specifically, the pass / fail calculation unit first performs a hard rejection judgment. Images that meet any of the following conditions are directly judged as unqualified and assigned a value of G. s,t =0:

[0049] S1: Image quality score Q s <0.6: The output range of the BRISQUE no-reference quality assessment algorithm is [0,1]. 0.6 is the critical value for distinguishing commercially viable high-quality images from low-quality images with obvious defects such as blurring and distortion. Images with visual effects below this value cannot meet the display requirements and are directly eliminated.

[0050] S2: Topic semantic matching degree C s,t <0.5: In the text-image cosine similarity calculated by the CLIP model, 0.5 is the minimum threshold for judging whether the image content is related to the topic. If it is lower than this value, it means that the image content is basically unrelated to the topic and is directly eliminated.

[0051] S3: Repeatability R s >0.3: A repetition score of over 0.3 indicates that the original content has an average similarity of more than 30% with the content displayed in the past 7 days, indicating excessive homogenization. To ensure the freshness of the displayed content, it will be directly eliminated.

[0052] For images that pass the hard elimination criterion, a multi-indicator weighted fusion calculation is used to determine their passability score. The specific calculation formula is as follows:

[0053]

[0054] The weighting coefficient is determined by the following logic: topic semantic matching degree C s,t The weight, 0.4, is the core of all evaluation dimensions. The displayed content must first meet the requirements of the input theme; therefore, it is assigned the highest weight. Image quality score Q s With a weight of 0.3, visual quality is the foundation of the display effect, second only to theme matching in importance, and therefore assigned the second highest weight. Style Distinctiveness D s A weight of 0.15 ensures diversity in the candidate pool, providing sufficient selection space for subsequent personalized matching; its importance is weaker than the first two dimensions. Originality (1 - R) s A weight of 0.15 is assigned to ensure the originality and freshness of the content, which is as important as the style differentiation.

[0055] The use of a product-based approach instead of linear weighting is due to the "barrel effect" relationship among the four dimensions: an extremely low score in any one dimension will lead to a significant drop in the overall pass rate, thus preventing images with extremely high scores in one dimension but extremely low scores in others from being mistakenly selected. The use of power-law weighting instead of direct multiplication by weights ensures that when the scores of each dimension are all within the [0,1] range, the final result also falls within the [0,1] range, while maintaining the non-linear constraint relationship between the indicators.

[0056] Passability score G s,t Images with a score of ≥0.75 enter the qualified image pool, with 3-5 qualified images retained per round. The passing score is set at 0.75, a reasonable threshold derived from weighted logic: when all four dimensions meet the passing level, the weighted score is approximately 0.75, ensuring that images entering the candidate pool meet the basic requirements in all dimensions. The candidate pool size is set at 3-5 images, which is the optimal range balancing style diversity and computational efficiency for subsequent matching. Too few images will result in insufficient style options, while too many will increase the computational cost of subsequent matching. If there are fewer than 3 qualified images meeting the criteria, the passing score is relaxed to 0.6, and the top 3 images with the highest scores enter the candidate pool; if there are still fewer than 3 images with scores above 0.6, new candidate images are generated.

[0057] After the above processing, a two-layer evaluation logic of hard threshold filtering and multi-index nonlinear weighted fusion is adopted. On the one hand, by using triple hard thresholds of topic matching degree, image quality, and repetition, low-quality, irrelevant, and highly repetitive images that do not meet the basic requirements are directly eliminated, ensuring the basic quality of candidate materials. On the other hand, a power product weighting method is used to fuse the four dimensions of topic matching degree, image quality, style differentiation, and originality. This breaks through the limitation of traditional linear weighting, which is prone to high scores in one dimension masking defects in other dimensions. It achieves a comprehensive evaluation based on the "barrel effect," where a low score in any of the four dimensions will lead to a significant drop in the overall pass rate, ensuring that the images entering the candidate pool meet the requirements for commercial display in all dimensions. At the same time, by introducing the style differentiation index, the images in the candidate pool are guaranteed to have sufficient style diversity, providing ample selection space for subsequent personalized matching. This solves the technical problems of unstable quality and monotonous style of Wensheng printable materials from the source.

[0058] Furthermore, the data analysis module also includes a user feature recognition module. The core function of the user feature recognition module is to eliminate interference from passing users and calculate the style preference distribution of the current effective user group. This includes a group size correction unit and a preference probability calculation unit. The specific processing logic is as follows:

[0059] Group size adjustment unit based on user distance d i User orientation θ i User dwell time t i Effective distance threshold D th and the effective stay duration threshold T th Output the corrected number of effective users n c Number of effective users n c The minimum value is 1, meaning that if there are no valid users, it will be treated as a single-user scenario by default.

[0060] Specifically, the group size correction unit first performs a dynamic effective distance determination, and the user's effective determination distance is dynamically adjusted according to their orientation. The greater the user's orientation deviates from directly facing the screen (i.e., the larger the angle of 90°), the shorter the dynamic effective distance. When the user is completely facing away from the device, the effective distance is shortened to 70% of the baseline value, i.e., 1.05 meters. The logic behind this setting is that the further the user's orientation deviates from facing the screen, the lower the probability that they will pay attention to the screen. Therefore, it is necessary to narrow the effective determination range and reduce the probability of misjudging passing users. The correction margin of 0.3 is based on human behavior statistics. For every 30° increase in the deviation angle, the probability of the user paying attention to the screen decreases by about 10%, corresponding to a 10% reduction in the effective distance. Therefore, the total correction margin is set at 30%. It should be emphasized that this value is an initial value and can be modified by the user as needed in specific applications.

[0061] Secondly, the dwell time is weighted. When the dwell time is less than the threshold, the weight decreases linearly with the dwell time. When the dwell time is greater than or equal to the threshold, the weight is 1. This achieves soft filtering of users with short dwell times and avoids misjudgment of threshold values ​​caused by relying solely on hard thresholds.

[0062] The weighted sum of the valid user determination results is rounded to the nearest integer, and then the result is multiplied by 1 to obtain the maximum value, which gives the final number of valid users n. c Rounding is used to obtain an integer count of valid users, which aligns with the user counting logic in real-world scenarios. Setting the minimum value to 1 ensures that when no valid user is detected, the system automatically reverts to the default single-user scenario, preventing the system from being unable to output content when there are no valid users. The specific calculation formula is as follows:

[0063]

[0064] Corrected number of effective users n c It accurately reflects the current user base that is paying attention to the display devices, providing a basis for subsequent preference adjustments for different group sizes.

[0065] Through the above processing, the limitations of most existing recognition systems, which rely solely on distance or dwell time for judgment, are overcome. The system innovatively introduces the concept of user orientation, dynamically adjusting the effective distance threshold. The further the user's orientation deviates from the front of the device, the closer the effective judgment distance. Combined with a soft-weighting mechanism based on dwell time, it achieves accurate differentiation between interested users and passersby, fundamentally reducing the probability of miscounting people passing by from the side or back. This ensures that the statistically significant number of effective users is completely consistent with the actual user group interested in the device. It solves the technical problems of low user recognition accuracy and susceptibility to interference from irrelevant users in existing intelligent display devices, providing accurate scenario-based information for subsequent preference calculations.

[0066] Furthermore, the preference probability calculation unit is based on the age dimension preference probability P. a (s), Clothing Dimension P Probability w (s), original group size mapping probability P org (s), number of effective users n c Interaction dwell time percentage T u,s Given the total number of styles S, output the user's preference weight W for style s. u,s Preference weight W u,s The value range is [0,1], and the sum of the weights of all styles is 1. The higher the weight, the higher the user's preference for that style.

[0067] Specifically, the preference probability calculation unit first performs group size preference correction, based on the number of effective users n. c Mapping probability P to the original group size org(s) is modified to obtain the constrained conditional probability P of style s under the modified group size. g (s), the revised rule is:

[0068] K1: When n c When = 1, niche / personalized style: P g (s)=P org (s) + 0.3; Popular style: P g (s)=P org The (s) -0.1 is because in a single-person scenario, users tend to choose content that matches their unique personal aesthetic rather than content that is generally accepted by the public.

[0069] K2: When 2≤n c When ≤3, the popular style: P g (s)=P org (s) + 0.25 is because small groups tend to choose popular styles that are acceptable to most people when making decisions, thus reducing group preference conflicts.

[0070] K3: When n c ≥4, High-impact / Festive style: P g (s)=P org The reason for (s)+0.3 is that in large-group scenarios, content needs to be able to quickly attract everyone's attention, and a style with high saturation and strong visual impact is more in line with the needs of the scenario.

[0071] The logic behind setting the correction range to 0.25-0.3 is to reflect the preference differences between different group sizes without completely covering the preference features of other dimensions, thus ensuring a balance of multi-dimensional preferences. All correction results are truncated using a boundary function to ensure that the probability values ​​always fall within the reasonable range of [0,1].

[0072] Next, multi-dimensional preference fusion is performed, using a product form to fuse the preference probabilities of the three basic dimensions. This is because the three dimensions are independent and complementary; styles that satisfy preferences across multiple dimensions are assigned a higher probability, consistent with the cumulative effect of user preferences. An interaction preference correction term (1+T) is then introduced. u,s ), T u,s This represents the percentage of time users spend on the preview image for the corresponding style, ranging from [0,1]. Therefore, the correction term ranges from [1,2]. The logic behind this design is that users' explicit interaction behavior is a direct feedback of their preferences. The longer the user stays on a style, the higher the probability of their preference can be, up to double, and this preference has a higher weight than the basic statistical dimension of preference.

[0073] Finally, normalization is performed by summing the numerator scores of all styles and using the sum as the normalization denominator, so that the final sum of the weights of all styles is 1, which facilitates the subsequent calculation and comparison of matching scores. The specific calculation formula is as follows:

[0074]

[0075] The obtained user preference weight W u,s The comprehensive quantification of the current user group's preference for different styles is the core basis for matching candidate images with user preferences.

[0076] User preference weight W u,s The processing approach integrates preference features from three basic statistical dimensions: age, clothing, and group size. It also incorporates explicit user interaction and dwell time data as a preference enhancement factor, achieving multi-dimensional preference fusion. The differentiated adjustment logic based on group size adapts to content preference patterns in different scenarios: individual, small group, and large group. Specifically, in individual scenarios, it strengthens the preference weight of niche, personalized styles to meet the unique aesthetic needs of individual users; in small group scenarios, it strengthens the preference weight of popular styles to meet the universal needs of group decision-making; and in large group scenarios, it strengthens the preference weight of high-impact styles to meet the attention-grabbing needs of large groups.

[0077] This allows for the precise quantification of preferences among user groups of different sizes, ensuring that the generated content is highly compatible with the needs of the current viewing user group, and solving the technical problem that existing generated content is prone to drift and become disconnected from user needs.

[0078] Furthermore, the data analysis module also includes an optimal image matching module. The core function of the optimal image matching module is to combine the quality of candidate images with user preferences, output the optimal matching image, and optimize the score mapping logic so that the full score corresponds to a result close to 1. The specific processing logic is as follows:

[0079] First, boundary processing is performed, and the pass / fail score G is... s,t And preference weights W u,s The minimum value is set to 10. －6 To avoid meaningless results due to zero values ​​in subsequent logarithmic calculations, the pass / fail score G after boundary processing is obtained. c And the preference weights W after boundary processing c The calculation formula is:

[0080]

[0081]

[0082] Then, a logarithmically weighted sum is performed, and the pass rate G is calculated. c And preference weights W cThe weights for both are 3, indicating that both have the same impact on the final matching score: ensuring the quality of the displayed image is acceptable and that it conforms to user preferences are equally important. The logarithmic function maps the score in the [0,1] interval to the (-∞, 0] interval, amplifying the differences in the low-score interval and reducing the differences in the high-score interval, which aligns with the user's perceptual logic: when the score is low, a small improvement in quality or preference will bring a significant perceptual difference. When the score is high, further improvements will result in a smaller perceptual difference. The logic of the offset +6 is: when G c =1, W c When the sum is 1 (full score scenario), all logarithmic terms are 0, and the summation result is 6. This ensures that the value input into the subsequent function is positive, making the output in the full score scenario close to 1, which conforms to the user's intuitive understanding of a high score.

[0083] Finally, a nonlinear mapping is performed using an activation function, the formula for which the activation function is calculated is:

[0084]

[0085] This function can map any input value to the (0,1) interval. Its S-shaped curve can achieve the mapping effect of "high sensitivity in the medium score interval and low sensitivity in the low / high score interval": when the matching degree is medium, changes in quality or preference will cause significant changes in score; when the matching degree is extremely low or extremely high, the score changes tend to be gradual, which is consistent with the matching decision logic of real-world scenarios.

[0086] Final match score M u,s,t The calculation formula is:

[0087]

[0088] G in perfect score scenario s,t =1, W u,s =1,M u,s,t The score is approximately 0.997, close to 1 but not equal to 1. This aligns with users' understanding of a perfect score while avoiding the problem of indistinguishable high-quality images due to score saturation. Ultimately, M is selected from all candidate images. u,s,t The image with the highest score is selected as the final display image, achieving an optimal balance between quality and preference.

[0089] The above processing achieves an optimal balance between candidate image quality and user preferences. Specifically, it employs a non-linear mapping logic combining logarithmic weighting and activation functions. On one hand, logarithmic transformation amplifies the differences in low-score intervals and reduces the differences in high-score intervals, aligning with users' perception of content quality. On the other hand, a function maps the matching score to the 0-1 interval, ensuring that the final score matches the user's intuitive perception of the matching degree. Furthermore, in the perfect score scenario, the score is close to 1 but not saturated, avoiding the problem of indistinguishable high-quality images due to score saturation.

[0090] This achieves a dual match between image quality standards and user preferences, ensuring that the final printed output is the optimal result from the candidate pool that simultaneously meets both high quality and high adaptability. This provides a precise decision-making basis for the personalized content output of AI-powered real-time image printing devices.

[0091] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-powered real-time image printing device, comprising a device layer and a backend layer, characterized in that: The device layer includes a microphone, a screen, and a printer; the backend layer includes a speech-to-text conversion and an AI image generation model, and the AI ​​image generation model is equipped with an image processing system. After the user issues a voice command, the microphone picks up the audio, which is then uploaded to the backend layer. After being processed by speech-to-text, it is imported into the AI ​​image model. The AI ​​image model processes and generates an image, which is then analyzed by the image processing system to output the optimal image content and send it to the information management platform. The information management platform then maps the image onto the screen and simultaneously starts the printer to print it. The image processing system includes an information acquisition module, a data analysis module, and a database. The information acquisition module acquires multi-source data required for the operation of the printing equipment and sends it to the database for archiving. The database divides the data into candidate image-related data, user feature-related data, and matching degree-related data. After cleaning, the data is input into the data analysis module. The data analysis module determines a qualified material pool based on the cleaned candidate image-related data, determines the distribution of user style preferences based on the user feature-related data, and determines the final display image based on the matching degree-related data.

2. The AI-powered real-time image printing device according to claim 1, characterized in that: The candidate image-related data includes user-input fixed themes, key visual features, a public element feature library, a historical image feature library, similarity thresholds, theme semantic matching degree, image quality score, and style discrimination degree. User characteristic-related data includes user distance, user orientation, user dwell time, effective distance threshold, effective dwell time threshold, age-dimension preference probability, clothing-dimension preference probability, original group size mapping probability, interaction dwell rate, and total number of styles; Matching data includes pass / fail scores and preference weights.

3. The AI-powered real-time image printing device according to claim 2, characterized in that: The data analysis module includes a candidate graph generation module, a user feature recognition module, and an optimal graph matching module.

4. The AI-powered real-time image printing device according to claim 3, characterized in that: The candidate image generation module includes a memory duplication correction unit and a pass / fail calculation unit; The memory repetition correction unit outputs a corrected repetition score based on key visual features, a public element feature library, a historical display image feature library, and a similarity threshold. The higher the score, the higher the image repetition.

5. The AI-powered real-time image printing device according to claim 4, characterized in that: The qualification calculation unit outputs the qualification score of the candidate image based on the topic semantic matching degree, image quality score, style differentiation degree and repetition score. The qualification score is used as the core indicator to determine whether an image can enter the qualified material pool.

6. The AI-powered real-time image printing device according to claim 5, characterized in that: The user feature recognition module includes a group size correction unit and a preference probability calculation unit; The group size correction unit outputs the corrected number of valid users based on user distance, user orientation, user dwell time, effective distance threshold, and effective dwell time threshold. The minimum number of valid users is 1, meaning that if there are no valid users, it is treated as a single-person scenario by default.

7. The AI-powered real-time image printing device according to claim 6, characterized in that: The preference probability calculation unit outputs the user's preference weight for style s based on the preference probability of age dimension, preference probability of clothing dimension, original group size mapping probability, number of effective users, interaction dwell rate and total number of styles. The sum of the weights of all styles is 1. The higher the weight, the higher the user's preference for that style.

8. The AI-powered real-time image printing device according to claim 7, characterized in that: The processing logic of the optimal graph matching module is as follows: First, the pass / fail score and preference weight are processed by boundary conditions to obtain the pass / fail score and preference weight after boundary processing. Then, a logarithmically weighted sum is performed, and finally, a non-linear mapping is performed through an activation function to obtain the final matching score. The image with the highest final matching score among all candidate images is selected as the final display image.

Images