Artificial intelligence-based personalized image synthesis method and system

Abstract

This invention provides a personalized image synthesis method and system based on artificial intelligence. First, a basic image library containing diverse artistic styles and themes is acquired, where each image unit carries style attribute tags and theme content tags. Next, a deep learning model performs feature parsing on the basic image library to generate a style feature library and a theme feature library. Then, user-specific data is collected through an interactive input interface. The style feature library, theme feature library, and user-specific data are then input into a feature fusion network to generate a personalized synthesis control factor. Finally, based on this control factor, matching image units in the basic image library are recombined and style-adapted to generate a synthesized image that meets the user's personalized needs. This method efficiently and accurately achieves personalized image synthesis, satisfying diverse creative needs of users.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a personalized image synthesis method and system based on artificial intelligence. Background Technology

[0002] In the field of image compositing, traditional methods primarily rely on manual design and fixed template rules. For example, in artistic image creation, artists need to manually select image elements of different styles and then splice and adjust them based on their own artistic understanding and experience to achieve specific artistic effects. This method is not only inefficient but also demands extremely high levels of professional skill and creative experience from artists, making it difficult to meet the needs of large-scale, diverse image compositing.

[0003] With the development of computer technology, some computer-aided image compositing methods have gradually emerged. While these methods can utilize the computing power of computers to process and combine images to a certain extent, they still have many limitations. On the one hand, they can usually only handle a limited range of artistic styles and themes, lacking comprehensive coverage of the rich diversity of art movements and complex subjects. On the other hand, these methods often fail to fully consider the personalized needs of users, making it difficult to generate personalized images that meet their specific requirements. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a personalized image synthesis method based on artificial intelligence, the method comprising:

[0005] Obtain a basic image library containing diverse art styles and themes. The basic image library contains multiple sets of image units with style attribute tags and theme content tags. The style attribute tags are used to mark the art style characteristics to which the image unit belongs, and the theme content tags are used to mark the core content theme presented by the image unit.

[0006] The basic image material library is subjected to feature parsing operation by a deep learning model to extract style representation elements, theme composition elements and element association features in the image unit, and integrate them to generate style feature library and theme feature library.

[0007] User personalized needs data are collected through an interactive input interface. The user personalized needs data includes style preference descriptions, thematic element specifications, and emotional atmosphere preferences.

[0008] The style feature library, the theme feature library, and the user personalized demand data are input into the feature fusion network. A style feature subset and a theme feature subset that match the user demand are generated through a feature mapping mechanism. A feature association map is constructed based on the style feature subset and the theme feature subset. Personalized synthetic control factors are generated through map fusion.

[0009] Based on the personalized synthesis control factor, the matching image units in the basic image material library are recombined and style adapted. The style transfer coefficient is applied through the style transfer network, the theme element arrangement rules are applied through the element arrangement network, and the emotion rendering criteria are applied through the emotion rendering network to generate a synthetic image that meets the user's personalized needs.

[0010] In another aspect, embodiments of the present invention also provide a personalized image synthesis system based on artificial intelligence, including a processor and a machine-readable storage medium connected to the processor. The machine-readable storage medium is used to store programs, instructions, or code, and the processor is used to run the programs, instructions, or code in the machine-readable storage medium to implement the above-described method.

[0011] Based on the above, this invention constructs a basic image resource library containing diverse artistic styles and themes, and utilizes a deep learning model for feature analysis. This comprehensively extracts style representation elements, theme composition elements, and element association features from image units. By collecting personalized user data through an interactive input interface, the invention accurately grasps the user's creative intent. The style feature library, theme feature library, and personalized user data are input into a feature fusion network to generate style feature subsets and theme feature subsets matching the user's needs. A feature association graph is then constructed, generating personalized synthesis control factors, achieving a deep fusion of user needs and image features. Based on the personalized synthesis control factors, matching image units in the basic image resource library are recombined and style-adapted. Through the synergistic effect of a style transfer network, element arrangement network, and emotion rendering network, synthesized images that meet the user's personalized needs can be generated efficiently and accurately, greatly improving the efficiency and quality of image synthesis and satisfying the diverse creative needs of users. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the execution flow of the personalized image synthesis method based on artificial intelligence provided in an embodiment of the present invention.

[0013] Figure 2 This is a schematic diagram of the hardware architecture of the personalized image synthesis system based on artificial intelligence provided in an embodiment of the present invention. Detailed Implementation

[0014] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating an artificial intelligence-based personalized image synthesis method according to an embodiment of the present invention. The following is a detailed description of the artificial intelligence-based personalized image synthesis method.

[0015] Step S110: Obtain a basic image material library containing diverse art styles and themes. The basic image material library contains multiple sets of image units with style attribute tags and theme content tags. The style attribute tags are used to mark the art style characteristics to which the image unit belongs, and the theme content tags are used to mark the core content theme presented by the image unit.

[0016] This embodiment focuses on the personalized generation of landscape painting images. When acquiring the basic image resource library, it is necessary to collect various image resources that meet the requirements. First, the legality and compliance of the collection channels must be clearly defined, obtaining images from authorized art databases, publicly available digital resources from museums, professional art creation platforms, and other legitimate channels. Images involving privacy or copyright protection must be screened and removed using technical means, retaining only image units that can be legally used in this method.

[0017] During the collection process, strict control must be exercised over the diversity of images to ensure that they cover landscape paintings of various artistic styles. For example, they include classical landscape paintings that emphasize the contrast of light and shadow and rich layers of the picture, as well as Impressionist landscape paintings with bright colors and free brushstrokes, and modernist landscape paintings with simple lines and abstract style.

[0018] For each collected landscape painting image, a precise style attribute tag and theme content tag must be added. The style attribute tag should be determined based on the typical characteristics of the art movement, such as "Classicism," "Impressionism," "Modernism," and "Post-Impressionism." Each tag should accurately reflect the art movement to which the image unit belongs. A landscape painting that depicts mountains and rivers with delicate brushstrokes and natural transitions of light and shadow can be tagged with "Classicism."

[0019] Thematic content tags are determined based on the core scenic elements presented in the image, such as "high mountains and sea of ​​clouds," "lake reflections," "autumn forest," and "desert oasis." An image showing a vibrant oasis in a vast desert would be tagged with "desert oasis."

[0020] After collection, all image units are standardized. Images are uniformly converted to common image formats. Simultaneously, image resolution is adjusted to a reasonable range, ensuring sufficient detail is preserved in subsequent processing without compromising processing efficiency due to excessive data volume.

[0021] Step S120: Perform feature parsing operation on the basic image material library through a deep learning model to extract style representation elements, theme composition elements and element association features in the image unit, and integrate them to generate a style feature library and a theme feature library. The style representation elements include color combination patterns, brush stroke expression forms and layout structure features, and the theme composition elements include core object forms, scene environment composition and element interaction relationships.

[0022] This step aims to perform deep analysis of image units in a basic image library using a deep learning model, extracting key features and generating a corresponding feature library. The deep learning model employs a convolutional neural network architecture, which has been pre-trained and possesses efficient image feature extraction capabilities.

[0023] Step S121: Input each image unit in the basic image material library into the feature extraction layer of the convolutional neural network, and analyze the pixel distribution pattern of the image unit through multi-layer convolution operations to generate a feature map. The feature map includes a color distribution feature map, a texture representation feature map, and a spatial layout feature map.

[0024] Each landscape painting image unit from the basic image library is sequentially input into the feature extraction layer of a convolutional neural network. The feature extraction layer of the convolutional neural network consists of multiple convolutional layers connected in sequence, and each convolutional layer contains several convolutional kernels.

[0025] When an image unit is input into the first convolutional layer, the convolutional kernel performs a sliding convolution operation on the image to extract low-level features such as edges and lines. These low-level features are then processed by an activation function to generate a preliminary feature map.

[0026] The generated preliminary feature maps are fed into the next convolutional layer, where the kernel size and number are set according to the feature extraction requirements. This process continues to extract and combine features, resulting in more complex features. As the number of convolutional layers increases, the extracted features gradually shift from low-level to high-level, capturing more representative information from the image.

[0027] After multiple convolutional operations, the final output includes a color distribution feature map, a texture representation feature map, and a spatial layout feature map. The color distribution feature map reflects the spatial distribution of different colors in the image, with the value at each location representing the relevant characteristics of the color at that location. The texture representation feature map reflects the distribution and shape of brushstrokes and textures in the image. The spatial layout feature map shows the position and spatial relationship of each element in the image.

[0028] Step S122: Extract color combination patterns from the color distribution feature map, identify the matching relationship between the main color tone and the auxiliary color tone in the image unit, the color transition mode and the color tone ratio characteristics, and generate style representation elements of the color dimension.

[0029] To analyze the color distribution feature map, the dominant color tone of each image unit is first determined. By statistically analyzing the values ​​of each color channel in the color distribution feature map, the dominant color in the image, i.e., the dominant color tone, is identified.

[0030] Next, secondary hues are identified. Secondary hues are colors that complement the main hue in an image. By analyzing the distribution range and frequency of colors other than the main hue in the color distribution feature map, the types of secondary hues are determined.

[0031] Analyze the relationship between the primary and secondary colors to determine whether they are contrasting, adjacent, or analogous colors. For example, if the primary color is blue and the secondary color is yellow, they are contrasting colors.

[0032] Study color transition patterns and observe whether the transition from the primary color to the secondary color is gradual or abrupt, as well as the width of the transition area and the smoothness of the color change.

[0033] Calculate the hue proportion feature and count the proportion of the main hue and each auxiliary hue in the image. This proportion is determined by the ratio of the area of ​​the corresponding color region in the color distribution feature map to the total area of ​​the image.

[0034] By integrating the pairing relationship between the primary and secondary colors, the color transition methods, and the color proportion characteristics, style representation elements of the color dimension are generated.

[0035] Step S123: Extract the brushstroke representation from the texture representation feature map, identify the direction pattern, thickness variation and texture density features of the brushstrokes in the image unit, and generate style representation elements of the texture dimension.

[0036] Step S1231: Input the texture representation feature map into the multi-scale feature analysis module, extract features through convolution kernels with different receptive fields, and generate a multi-scale texture feature map. The multi-scale texture feature map includes a fine texture layer, a medium texture layer, and a coarse texture layer.

[0037] After the texture representation feature map is input into the multi-scale feature analysis module, different convolutional kernels in the multi-scale feature analysis module have different receptive field sizes. Convolutional kernels with small receptive fields are used to extract fine texture information in the image and generate fine texture layers; convolutional kernels with medium receptive fields extract medium-scale texture features and generate medium texture layers; convolutional kernels with large receptive fields capture the overall texture distribution and generate coarse texture layers.

[0038] Fine texture layers can reflect subtle changes and details in brushstrokes, such as the edge contours and local transitions of brushstrokes; medium texture layers show the transitions and connections of brushstrokes within a certain area; and coarse texture layers reflect the overall distribution pattern and general shape of brushstrokes.

[0039] Step S1232: Identify the detailed representation of brushstrokes from the fine texture layer, extract the edge contour features, distribution of turning nodes and local direction changes of the brushstrokes, and generate basic data on the direction rules of brushstrokes.

[0040] The fine texture layer is analyzed, and an edge detection algorithm is used to identify the edge contour features of the brushstrokes. By detecting changes in pixel values ​​in the texture layer, the position and shape of the brushstroke edges are determined, and the edge contour features are stored in the form of multi-dimensional feature vectors.

[0041] Identify turning points in brushstrokes—the locations where the brushstroke direction changes. Determine the distribution of these turning points by analyzing abrupt changes in pixel values ​​within the texture layer, and record the relevant feature parameters for each node.

[0042] Observe the changes in the direction of the brushstrokes in a local area, record the direction and trend of the changes, and generate basic data on the pattern of brushstroke direction, which includes detailed information on the edge contour of the brushstrokes, the distribution of turning points, and local changes in direction.

[0043] Step S1233: Identify the transition features of the brushstrokes from the medium texture layer, extract the width variation, density transition and connection mode of the brushstrokes in different areas, and generate core data of brushstroke thickness variation.

[0044] The medium texture layer effectively reflects the transition of brushstrokes across different areas. Analysis of this texture layer extracts the width variation characteristics of the brushstrokes. By measuring the width values ​​of brushstrokes at different locations, the maximum and minimum width values, as well as the range of variation, are recorded.

[0045] Identify the characteristics of the transition between light and dark strokes, analyze the color density changes of the strokes in different areas, determine whether the transition between light and dark strokes is gradual or rapid, and the size of the transition area.

[0046] This study examines the connection methods between brushstrokes, determining whether they are connected through overlapping, crossing, or continuous extension, and records relevant feature parameters. Information such as brushstroke width variations, density transitions, and connection methods are integrated to generate core data on brushstroke thickness variations.

[0047] Step S1234: Identify the overall distribution of brushstrokes from the coarse texture layer, extract the coverage, distribution density and regional correlation of the texture area, and generate key data of texture density features.

[0048] The coarse texture layer shows the overall distribution of brushstrokes. Analyzing it determines the coverage area of ​​the texture region, i.e., the size and positional distribution of the brushstrokes within the image. The coverage area is quantified by calculating the ratio of the texture region's area to the total image area.

[0049] Calculate the texture distribution density and count the number and distribution of brushstrokes per unit area. Areas with high density indicate dense brushstroke distribution, while areas with low density indicate sparse brushstroke distribution.

[0050] Analyze the correlation between texture regions to determine whether different texture regions are independent of each other or have overlapping, connected or other relationships, and record these correlation features.

[0051] By integrating information such as the coverage, distribution density, and regional correlation of texture regions, key data for texture density features are generated.

[0052] Step S1235: Perform directional clustering processing on the basic data of the stroke direction pattern, identify the main direction, secondary direction and direction conversion nodes, and generate a direction pattern descriptor.

[0053] The basic data on brushstroke direction patterns are subjected to directional clustering. A clustering algorithm is used to group brushstroke features with similar directions into one class. Through multiple iterative calculations, the central direction of the cluster is determined, which represents the primary and secondary directions of the brushstroke.

[0054] During clustering, direction transition nodes are simultaneously identified; these nodes represent the locations where brushstrokes transition between different directions. The angular ranges of the primary and secondary directions, as well as the distribution characteristics of the direction transition nodes, are recorded.

[0055] This information is then integrated to generate a direction pattern descriptor, which contains various key information about the direction of the brushstrokes.

[0056] Step S1236: Perform amplitude analysis on the core data of the stroke thickness variation, identify the range, frequency and associated region of the thickness variation, and generate a thickness variation descriptor.

[0057] Amplitude analysis is performed on the core data of stroke thickness variation to determine the range of thickness variation, i.e., the width range of the thinnest and thickest strokes. Through statistical analysis, the frequency of thickness variation is obtained, i.e., the number of times the stroke thickness changes per unit length.

[0058] Identify the associated regions of thickness variation, i.e., which areas have more frequent variations in stroke thickness and which areas have relatively gentle variations. Record the location and extent characteristics of these areas. Integrate the information such as the range of thickness variation intervals, frequency of variation, and associated regions to generate a thickness variation descriptor.

[0059] Step S1237: Perform density distribution analysis on the key data of the texture density features to identify the distribution range of high-density areas, medium-density areas and low-density areas, and generate density feature descriptors.

[0060] Density distribution analysis was performed on key data of texture density features. Density thresholds were set to divide the texture region into high-density, medium-density, and low-density regions. The distribution range of each region was determined by calculating its area and location coordinates.

[0061] Record the specific location and morphological features of high-density, medium-density, and low-density regions in the image, and integrate this information to generate a density feature descriptor.

[0062] Step S1238: Combine the direction pattern descriptor, the thickness variation descriptor, and the density feature descriptor according to the hierarchical relationship of texture features to generate style representation elements of the texture dimension. The style representation elements of the texture dimension include a brush stroke direction map, a thickness variation curve, and a density distribution matrix.

[0063] Based on the hierarchical relationship of texture features, the direction pattern descriptor, thickness variation descriptor, and density feature descriptor are combined. The stroke direction map is constructed from the direction pattern descriptor, showing the overall direction and directional changes of the stroke in the image; the thickness variation curve is generated based on the thickness variation descriptor, reflecting the change of stroke width with position; and the density distribution matrix is ​​transformed from the density feature descriptor, reflecting the distribution of texture density in the image.

[0064] Step S124: Extract layout architecture features from the spatial layout feature map, identify the position distribution, area proportion and spatial relationship between core visual elements in the image unit, and generate style representation elements of the layout dimension.

[0065] Analyzing the spatial layout feature map first involves identifying the core visual elements within the image units. Core visual elements are those that occupy an important position in the image and play a key role in expressing the theme, such as "mountains," "lakes," and "forests."

[0066] By analyzing the pixel values ​​of the spatial layout feature map, the positional distribution of core visual elements is determined, and the coordinate range of each core visual element is recorded. The ratio of the area occupied by each core visual element in the image to the total image area is calculated to obtain the region proportion feature.

[0067] Analyze the spatial relationships between core visual elements, such as their relative positions (up / down, left / right, front / back, etc.), distances, and whether there is occlusion. For example, a mountain is above a lake, the two are relatively close, and the mountain partially obscures the edge of the lake.

[0068] By integrating information such as the location distribution, area proportion, and spatial relationship between core visual elements, style representation elements of the layout dimension are generated.

[0069] Step S125: Combine the style representation elements of color dimension, texture dimension and layout dimension according to feature correlation to generate style feature vector of each image unit, and summarize all style feature vectors to generate style feature library.

[0070] We conduct feature correlation analysis on style representation elements in the color, texture, and layout dimensions to identify their intrinsic relationships. For example, the distribution of color may be related to the position of core elements in the layout, and the density of texture may change with color.

[0071] Based on these relationships, the style representation elements of the three dimensions are combined. The combination process uses feature concatenation, which concatenates the multi-dimensional feature vectors corresponding to each style representation element in a predetermined order to form the style feature vector of each image unit.

[0072] Style feature vectors contain all style information about an image unit in terms of color, texture, and layout. The style feature vectors of all image units are aggregated, stored, and managed according to predefined rules to generate a style feature library. Each entry in the style feature library is associated with a corresponding image unit in the base image resource library, facilitating subsequent retrieval and matching.

[0073] Step S126: The object detection network analyzes the core object shape in the image unit, identifies the contour features, morphological structure and detail of the object, and generates the theme composition elements of the core object dimension.

[0074] The object detection network employs a deep learning-based object detection architecture, consisting of a feature extraction network and a detection head. Image units are input into the feature extraction network, which extracts high-level features of the image through multiple convolutional and pooling operations.

[0075] The detection head processes the extracted high-level features to generate candidate regions, which may contain the core object. A classifier then classifies these candidate regions to determine whether they contain the core object and, if so, the category of that object.

[0076] For candidate regions containing the core object, their contour features are further analyzed. Through edge detection and contour extraction algorithms, the contour shape and edge details of the core object are obtained, and the contour features are presented in the form of a multi-dimensional set of coordinate points.

[0077] Analyze the morphological structure of the core object to determine its overall shape, the composition of its parts, and the connections between them. For example, the morphological structure of a large tree includes the trunk, branches, and leaves; the trunk supports the branches, and leaves grow on the branches.

[0078] Identify the detailed features of the core object, such as surface texture and color variations. Integrate information on contour features, morphological structure, and detailed representation to generate thematic elements for the core object dimension.

[0079] Step S127: The scene environment composition in the image unit is analyzed by the scene parsing network, the environmental elements, spatial structure and atmosphere features in the scene are identified, and the theme composition elements of the scene environment dimension are generated.

[0080] The scene parsing network employs a fully convolutional network architecture, enabling pixel-level image classification. Image units are input into the scene parsing network, which then performs feature extraction and semantic segmentation through multiple convolutional operations.

[0081] Semantic segmentation divides image units into different regions, each corresponding to an environmental element such as "sky," "ground," "water," and "vegetation." The types and quantities of these environmental elements are then counted to determine the composition of the scene's environment.

[0082] Analyzing the spatial structure of the scene determines the distribution and relationships of various environmental elements. For example, the sky is at the top of the image, the ground is at the bottom, water bodies are distributed in low-lying areas, and vegetation grows on the ground. Furthermore, the scene's atmospheric characteristics are identified, determined by the color, brightness, and texture of the environmental elements. For instance, an image with a pale blue sky, bright sunshine, and lush green vegetation has an overall atmosphere of "fresh and bright." Therefore, by integrating information on environmental element composition, spatial structure, and atmospheric characteristics, the thematic elements constituting the scene's environmental dimension are generated.

[0083] Step S128: The interaction relationships of elements in the image unit are analyzed by the relationship recognition network, the positional association, interaction state and dependency pattern between the core object and the environmental elements and the objects are identified, and the theme constituent elements of the relationship dimension are generated.

[0084] Step S1281: Input the image unit into the relationship recognition network to encode the features of the core object and environmental elements, and generate object feature vectors and environmental feature vectors.

[0085] The relationship recognition network includes a feature encoding module, which consists of multiple fully connected layers. After image units are input into the relationship recognition network, the core objects and environmental elements are first located and extracted.

[0086] The feature encoding module performs feature encoding on the extracted core objects and environmental elements respectively. For core objects, features such as shape, color, and texture are extracted during the encoding process; for environmental elements, features such as type, distribution, and attributes are extracted during the encoding process.

[0087] After feature encoding, object feature vectors and environment feature vectors are generated. Both vectors are multi-dimensional feature vectors, containing key feature information of the core object and environmental elements, respectively.

[0088] Step S1282: Calculate the spatial positional relationship between the object feature vector and the environment feature vector, identify the relative position, orientation relationship and distance features of the core object in the environment, and generate a position association descriptor.

[0089] The system extracts spatially relevant feature components from the object's feature vector and the environment's feature vector. These components contain the coordinate information of the core object and environment elements in the image coordinate system. By calculating the relative difference between the center coordinates of the core object and the boundary coordinates of the environment elements, the system determines the relative position of the core object within the environment, such as whether the core object is located to the left, right, above, or below the environment elements.

[0090] Analyze the relationship between the orientation of the core object and the distribution direction of environmental elements to determine the orientation relationship. For example, whether the main orientation of the core object is consistent with, perpendicular to, or at an angle to the extension direction of the environmental elements.

[0091] The shortest distance between the core object and environmental elements is calculated, derived from the geometric distance formula between coordinate points, and quantified into multi-dimensional numerical components. Relative position, orientation relationship, and distance features are integrated into a position association descriptor, where each element corresponds to a quantified feature of a positional relationship.

[0092] Step S1283: Analyze the action interaction features between the core object and other objects, identify the type, intensity and duration of the interaction actions, and generate an interaction state descriptor.

[0093] When multiple core objects exist within an image unit, the interaction between these core objects is analyzed. By analyzing the action-related feature components in the object's feature vector, the type of interaction action is identified, such as approaching, moving away, contacting, or surrounding.

[0094] The intensity of an interactive action is determined by the magnitude of changes in the object's characteristics during the action. For example, the faster the distance between objects changes, the higher the interaction intensity; the greater the range of characteristic changes, the higher the interaction intensity.

[0095] Analyze the continuity of interactive actions in the temporal dimension (for sequential image units) or the distribution range in the spatial dimension (for single-frame image units) to determine persistence features. In a single-frame image, persistence features are determined by the completion degree and the range of influence of the action; for example, the wider the area covered by the interactive action, the more significant the persistence feature. Therefore, the type, intensity, and persistence features of interactive actions are integrated into an interaction state descriptor.

[0096] Step S1284: Analyze the functional relationships between the core object and the environmental elements, identify the support, containment and complementary relationships between the elements, and generate a dependency pattern descriptor.

[0097] Analyze the functional relationship between the core object and environmental elements to determine whether a supporting relationship exists. For example, does the core object depend on environmental elements for its existence? If a tree grows on the ground and the ground supports the tree, then a supporting relationship exists between them.

[0098] Identify containment relationships, that is, whether the surrounding elements completely or partially surround the core object, such as islands in a lake being surrounded by water, where the water and islands have a containment relationship.

[0099] Determining a complementary relationship involves determining whether the core object and environmental elements functionally complement each other to form a complete scene. For example, an oasis in the desert forms an ecological complementarity between the desert and the oasis, thus indicating a complementary relationship between the two.

[0100] The identification results of support, inclusion, and complement relationships are quantified into feature parameters and integrated to generate dependency pattern descriptors.

[0101] Step S1285: Perform association analysis on the location association descriptor, the interaction state descriptor, and the dependency pattern descriptor to identify the inherent connections between the three types of relationship descriptors and generate a relationship association matrix.

[0102] The feature parameters in the location association descriptor, interaction state descriptor, and dependency pattern descriptor are compared pairwise to calculate the correlation between the feature parameters of different descriptors. For example, the proximity feature in the location association may be highly correlated with the contact action feature in the interaction state, and the support relationship may be associated with specific relative position features.

[0103] Based on the correlation analysis results, a relationship association matrix is ​​constructed, where each element represents the association strength of feature parameters between different descriptors. The association strength is calculated using the co-occurrence frequency and consistency of variation of feature parameters; higher values ​​indicate a stronger association.

[0104] Step S1286: Based on the relational matrix, the three relation descriptors are integrated to generate topic components of the relation dimension. The topic components of the relation dimension include a positional relation graph, an interaction state sequence, and a dependency pattern model.

[0105] Based on the association strength in the relational matrix, the location association descriptor, interaction state descriptor, and dependency pattern descriptor are weighted and integrated. Feature parameters with high association strength are given higher weights during the integration process to highlight their importance in the relational dimension.

[0106] The integrated location association features are constructed into a location relationship graph. The nodes in the graph represent core objects and environmental elements, and the edges represent the location relationships and strengths between them.

[0107] The integrated interactive state features are transformed into an interactive state sequence, which displays the interaction process between core objects in chronological or logical order.

[0108] The integrated dependency pattern features are constructed into a dependency pattern model, which describes the functional dependency relationship between the core object and environmental elements in a structured way.

[0109] By integrating positional relationship maps, interaction state sequences, and dependency pattern models, the thematic components of the relationship dimension are formed, comprehensively reflecting the interactive relationships of elements in the image unit.

[0110] Step S129: Integrate the theme components of the core object dimension, scene environment dimension and relationship dimension according to feature importance to generate a theme feature vector for each image unit, and summarize all theme feature vectors to generate a theme feature library.

[0111] The importance of features in the thematic elements across the core object dimension, scene environment dimension, and relationship dimension is assessed. The assessment is based on the contribution of each dimension's features to expressing the image's thematic content. For example, features in the core object dimension directly determine the core content of the theme, and their importance is relatively high; features in the scene environment dimension provide background support for the theme, and their importance is secondary; features in the relationship dimension reflect the connections between elements, and their importance fluctuates depending on the specific thematic content.

[0112] Based on the importance assessment results, corresponding weights are assigned to the thematic components of each dimension. Dimensions with higher importance retain more feature details during the integration process and therefore have relatively larger weight values.

[0113] By employing a feature concatenation method, the three-dimensional thematic components are concatenated together according to their weights to form a thematic feature vector for each image unit. Thematic feature vectors contain all thematic information about the image unit regarding its core object, scene environment, and element relationships.

[0114] Thematic feature vectors of all image units are aggregated to establish a thematic feature library. The thematic feature library adopts the same storage and indexing mechanism as the style feature library to ensure that each thematic feature vector is accurately associated with the corresponding image unit in the basic image material library.

[0115] Step S1210: Analyze the intrinsic relationship between the feature vectors in the style feature library and the theme feature library through feature association network, identify the co-occurrence pattern of style features and theme features, and generate an element association feature table. The element association feature table is used to record the matching relationship between different style features and theme features.

[0116] The Feature Association Network adopts a graph neural network architecture, consisting of an input layer, hidden layers, and an output layer. Style feature vectors from a style feature library and topic feature vectors from a topic feature library are used as input data and simultaneously fed into the input layer of the Feature Association Network.

[0117] The input layer standardizes the input feature vectors to bring feature parameters of different dimensions to the same order of magnitude. Standardization is achieved through feature scaling, mapping the feature parameters to a predefined numerical range.

[0118] The hidden layer contains multiple graph convolutional layers, which model the relationships between feature vectors through graph convolution operations. The graph convolutional layer treats each feature vector as a node in a graph, and determines the weight of the edges by calculating the similarity between nodes. The weight values ​​reflect the strength of the association between style features and topic features.

[0119] Through multi-layer graph convolution operations, feature association networks can capture deep co-occurrence patterns between style features and thematic features. For example, certain color combination patterns (style features) often co-occur with "autumn forest" (thematic feature), forming a stable co-occurrence pattern.

[0120] The output layer parses the processing results of the hidden layer to generate an element association feature table. This table records the matching relationships between different style feature vectors and topic feature vectors, including matching feature parameters, association strength, and co-occurrence frequency. The element association feature table is stored in a tabular structure for easy retrieval and retrieval during subsequent feature fusion.

[0121] Step S130: Collect user personalized needs data through an interactive input interface. The user personalized needs data includes style preference description, theme element specification, and emotional atmosphere preference. The style preference description is used to reflect the user's specific preference for art style, and the theme element specification is used to determine the core visual elements that the user wants to include.

[0122] This step guides users to accurately express their personalized needs for landscape painting images by building an intuitive and easy-to-use interactive input interface.

[0123] Step S131: Construct an interactive input interface. The interactive input interface includes a style preference selection area, a theme element specification area, and an emotional atmosphere adjustment area. The style preference selection area provides style example images and style feature description text. The theme element specification area supports element category selection, feature description input, and reference image upload. The emotional atmosphere adjustment area provides emotional vocabulary options and atmosphere example images.

[0124] The interactive input interface adopts a graphical user interface design with a clear overall layout and simple operation process. The style preference selection area is located on the left side of the interface, displaying example images of landscape paintings in various artistic styles. For example, the "Classical" style examples showcase landscapes with delicate brushstrokes and rigorous composition, while the "Impressionist" style examples showcase landscapes with bright colors and variations in light and shadow. Each example image is accompanied by a descriptive text describing the style characteristics, detailing the typical features of that style, such as "Classicism: Emphasis on the accuracy of lines and the sense of layering in the picture, with stable colors and soft contrasts of light and shadow."

[0125] The designated theme element area is located in the middle of the interface, featuring a drop-down menu for element category selection, including several core landscape element categories such as "mountains," "lakes," "forests," and "deserts." A text input box is also provided, allowing users to enter specific descriptions of the theme element, such as "a steep mountain peak covered in snow." Additionally, a reference image upload button allows users to upload reference images containing the target theme element.

[0126] The mood / atmosphere adjustment area is located on the right side of the interface, listing various mood / atmosphere options such as "tranquil," "magnificent," "warm," and "desolate," with each mood / atmosphere corresponding to a set of example images. For example, the example image for "tranquil" shows a calm lake, soft lighting, and sparse vegetation, creating a serene atmosphere.

[0127] At the bottom of the interface are a confirmation button and a reset button. After completing the input, the user can click the confirmation button to submit the required data, and click the reset button to clear all the input and start over.

[0128] Step S132: Monitor the user's operation behavior in the style preference selection area, record the style identifier corresponding to the style example image selected by the user, collect the style feature description text input by the user, perform semantic parsing on the style feature description text, extract style keywords and feature descriptions, and integrate the style identifier with the style keywords and feature descriptions into a style preference description.

[0129] The system monitors user actions in the style preference selection area in real time through the interface interaction monitoring module. When a user clicks on a style example image, the system automatically records the style identifier corresponding to that example image, such as "F001" representing classicism and "F002" representing impressionism.

[0130] If the user enters a description of the style characteristics in the text input box next to the style example image, such as "I hope the colors of the picture are bright, the brushstrokes are free, like Impressionism but more lively", then the natural language processing module will be invoked to perform semantic parsing on the text.

[0131] The semantic parsing process first performs word segmentation, breaking the text down into multiple word units. Then, it identifies key words such as nouns and adjectives through part-of-speech tagging. Finally, it combines a knowledge base in the field of art styles to extract style keywords such as "vibrant colors," "free brushstrokes," and "bright."

[0132] Feature mapping is performed on style keywords to transform them into quantifiable feature descriptions. For example, "vibrant colors" corresponds to the feature description of high color saturation, and "free brushstrokes" corresponds to the feature description of brushstrokes without fixed direction.

[0133] By linking and integrating style identifiers, style keywords, and feature descriptions, a structured style preference description is formed, which includes all information about the user's preferences for art styles.

[0134] Step S133: Respond to the user's input operation in the designated topic element area. If the user selects an element category, record the selected element category identifier and feature parameters. If the user inputs a feature description, extract element features from the description text to generate element feature parameters. If the user uploads a reference image, perform feature parsing on the reference image to extract core element features. Summarize the information obtained above to generate the designated topic element.

[0135] When a user selects an element category, such as "mountain", from the drop-down menu in the designated area of ​​the theme element, the system records the identifier "E001" corresponding to the element category and displays the default feature parameter options under that category, such as "number of peaks", "mountain shape", "whether there is snow", etc. The user can select or modify these feature parameters according to their needs, and the system records the modified parameter values ​​synchronously.

[0136] If a user enters a description of the theme elements in the text input box, such as "a clear lake with dense forests on its shore and small boats on its surface," the element feature extraction module will be invoked to process the text. First, semantic understanding is performed to identify the core elements as "lake," "forest," and "boat." Then, features for each element are extracted; for example, the feature for "lake" is "clear," the feature for "forest" is "dense," and the feature for "boat" is "located on the lake surface." These features are then converted into element feature parameters.

[0137] When a user uploads a reference image, the system calls the image feature analysis module to process it. The reference image first undergoes preprocessing, including resizing and format conversion, and then is input into a pre-trained object detection model to identify the core elements in the reference image, such as "waterfall" and "rocks" in the reference image.

[0138] Feature extraction is performed on the identified core elements, including morphological features (such as the shape of the waterfall's water flow and the shape of the rocks), color features (such as the color of the waterfall's water and the color of the rocks), and positional features (such as the relative position of the waterfall and the rocks). These features are then integrated into core element features.

[0139] The information obtained by users through element category selection, feature description input, or reference image upload is summarized, duplicate or conflicting feature parameters are removed, and a unified theme element specification is formed, which includes feature information of all core visual elements that the user wants to include.

[0140] Step S134: Capture the user's selection behavior in the emotional atmosphere adjustment area, record the emotional words and atmosphere example images selected by the user, and perform emotional feature analysis to determine the user's preferred emotional atmosphere type and performance characteristics, and generate emotional atmosphere preferences.

[0141] The system captures user selection behavior in the emotional atmosphere adjustment area through the interface interaction monitoring module. When a user clicks on a certain emotional word (such as "magnificent") or the corresponding atmosphere example image, the system records the selected emotional word and example image identifier.

[0142] The sentiment feature analysis module is invoked to analyze the selected sentiment words and atmosphere example images. For sentiment words, the corresponding sentiment dimension is determined by combining the sentiment dictionary, such as "magnificent" corresponding to the sentiment dimensions of "grand" and "open".

[0143] Visual feature analysis was performed on the atmospheric example images to extract visual features related to emotions, such as the color tone of the image (the "magnificent" example images are mostly dark blue, ochre and other heavy colors), composition method (mostly using a distant view composition to show a vast space), and element distribution (such as large areas of sky, mountains, etc.).

[0144] By associating the emotional dimension of emotional vocabulary with the visual features of example images, we can determine the type of emotional atmosphere preferred by users (such as "grand") and specific expressive features (such as rich colors, open composition, and inclusion of large natural elements).

[0145] This information is integrated to generate emotional atmosphere preferences, including emotional types and corresponding visual performance feature parameters.

[0146] Step S135: Perform correlation analysis on the style tendency description, the theme element specification, and the emotional atmosphere preference to generate structured user personalized demand data.

[0147] Correlation analysis was performed on the characteristic parameters in style preference description, theme element specification, and emotional atmosphere preference to identify the intrinsic relationships among the three. For example, the "Impressionist" style (style preference) selected by a user may be associated with "sunny lake" (theme element) and "warm" (emotional atmosphere), because the bright colors of Impressionism are suitable for depicting a warm, sunny lake scene.

[0148] Based on the correlation analysis results, the consistency of the feature parameters of the three parts is checked, and conflicting features are removed. For example, if the style preference requires "stable colors" while the emotional atmosphere preference requires "bright colors," the system prompts the user to make changes, or reconciles the settings based on the user's historical selection weights.

[0149] The verified feature parameters are organized according to the set structure to generate structured user personalized demand data. It adopts a hierarchical feature system, with the top layer divided into three parts: style, theme, and emotion. Each part contains corresponding sub-feature parameters.

[0150] Step S140: Input the style feature library, the theme feature library and the user personalized demand data into the feature fusion network, generate a style feature subset and a theme feature subset that match the user demand through the feature mapping mechanism, construct a feature association map based on the style feature subset and the theme feature subset, and generate a personalized synthesis control factor through map fusion. The personalized synthesis control factor includes style transfer coefficient, theme element arrangement rules and emotion rendering criteria.

[0151] This step uses a feature fusion network to accurately match and deeply fuse the user's personalized needs with pre-built style feature libraries and theme feature libraries, generating personalized synthesis control factors to guide image synthesis and ensuring that the synthesized image can accurately meet the user's personalized needs.

[0152] Step S141: The style preference description in the user personalized demand data is input into the text feature encoding network and converted into a style demand feature vector.

[0153] The text feature encoding network adopts a Transformer architecture, which includes an embedding layer, a multi-head self-attention layer, and a feedforward neural network layer. Textual information from style tendency descriptions (such as style keywords and feature descriptions) is input into the embedding layer. The embedding layer converts each word in the text into a fixed-dimensional word vector. The word vectors are generated based on a pre-trained word embedding model, which can capture the semantic information of words.

[0154] The word vector sequence is input into a multi-head self-attention layer, which calculates the attention weights between words in parallel through multiple attention heads to capture the semantic relationships between different words in style description. For example, there is a high attention weight between "colorful" and "impressionism", indicating that the two are closely related in style description.

[0155] The output of the multi-head self-attention layer is processed by a feedforward neural network layer. This feedforward neural network consists of two linear transformation layers and one activation function layer, performing non-linear transformations and dimensional adjustments on the features to ultimately generate a fixed-dimensional style requirement feature vector. This style requirement feature vector contains all the semantic information about the user's style preferences and can be matched with style feature vectors in a style feature library.

[0156] Step S142: The style requirement feature vector is matched with the style feature vector in the style feature library by the feature matching module. Style feature vectors that match the style requirement feature vector with a preset condition are found and a style feature subset is formed.

[0157] Step S1421: The style requirement feature vector and the style feature vector in the style feature library are input into the vector comparison layer of the feature matching module to perform a dimension-by-dimensional comparison of the vector features and identify the feature similarity in the same dimension.

[0158] The vector comparison layer of the feature matching module receives the style requirement feature vector and the style feature vector in the style feature library, and performs dimension alignment on the two vectors to ensure that they have the same number of dimensions and the same meaning of dimensions.

[0159] The similarity between two vectors is calculated dimension by dimension, based on the degree of difference in feature parameters. For example, for the color saturation dimension, if the parameter values ​​of the style requirement feature vector are small in difference from the parameter values ​​of a certain style feature vector, the similarity in that dimension is high; otherwise, it is low.

[0160] The feature similarity of each dimension is recorded as a numerical component, forming a similarity vector.

[0161] Step S1422: Calculate the overall matching degree between the style requirement feature vector and the style feature vector based on the dimensional feature similarity, and generate a matching degree value.

[0162] After obtaining the similarity vector, the overall matching degree needs to be calculated based on the feature similarity of each dimension. First, a corresponding weight is assigned to each dimension, and the magnitude of the weight is determined according to the importance of that dimension in the style features. For example, the weight of the dimension related to color combination patterns in Impressionist style matching may be higher than that of some dimensions of brushstroke expression.

[0163] Each component in the similarity vector is multiplied by its corresponding weight to obtain the weighted similarity for each dimension. Then, all weighted similarities for each dimension are summed and divided by the sum of the weights to obtain the overall matching score, which reflects the overall similarity between the style requirement feature vector and the style feature vector. The higher the score, the higher the matching degree between the two.

[0164] Step S1423: Select style feature vectors with matching scores higher than the matching score threshold as candidate style feature vectors.

[0165] A matching score threshold is preset, which is determined based on the style matching accuracy requirements in the actual application scenario. The calculated overall matching score is compared with this threshold. If the overall matching score of a style feature vector is higher than the threshold, it is selected and included in the candidate style feature vector set.

[0166] Step S1424: Select vectors that meet preset conditions from the candidate style feature vectors to form a style feature subset. The style feature subset includes multiple style feature vectors and their respective matching degree information.

[0167] The preset conditions include a limit on the number of candidate style feature vectors and a ranking requirement based on matching degree. First, the candidate style feature vectors are ranked from highest to lowest according to their overall matching degree. Then, based on the preset upper limit, a subset of the top-ranked candidate style feature vectors is selected to form a style feature set.

[0168] For example, if the preset maximum number is ten, then the top ten candidate style feature vectors are selected. Simultaneously, the overall matching degree value corresponding to each vector is retained in the style feature subset, so that this information can be referenced in subsequent steps for further processing and analysis.

[0169] Step S143: The topic element specified input element feature encoding network in the user personalized demand data is converted into a topic demand feature vector.

[0170] The element feature encoding network consists of an embedding layer and multiple fully connected layers. The subject element specification contains user descriptions, selections, or reference image information about the core visual elements, and this information is fed into the embedding layer.

[0171] The embedding layer converts text descriptions into text embedding vectors, element category selections into category embedding vectors, and the feature parsing results of the reference image into image embedding vectors. These embedding vectors are then input into the subsequent fully connected layer, which fuses and processes them through nonlinear transformations to ultimately generate a topic requirement feature vector.

[0172] The theme requirement feature vector is a multi-dimensional vector, with each dimension corresponding to a feature attribute of the theme element, such as element type, shape feature, color feature, etc.

[0173] Step S144: The feature matching module matches the topic requirement feature vector with the topic feature vector in the topic feature library, finds the topic feature vector that matches the topic requirement feature vector with a preset condition, and forms a topic feature subset.

[0174] Step S1441: Input the topic requirement feature vector and the topic feature vector in the topic feature library into the topic comparison layer of the feature matching module to perform a structured comparison of topic features and identify the matching status of core topic elements.

[0175] The topic comparison layer of the feature matching module first performs structured parsing on the topic requirement feature vector and the topic feature vector in the topic feature library, extracting the core topic element information contained therein. This core topic element information includes the element type, key attributes, and interrelationships.

[0176] Then, the core theme elements of both are compared one by one to identify matching and non-matching elements. For example, if the theme requirement feature vector contains two core theme elements, "lake" and "willow," and a certain theme feature vector contains "lake" and "poplar," then "lake" is a matching element, while "willow" and "poplar" are non-matching elements. At the same time, the number of matching core theme elements and the degree of matching are recorded.

[0177] Step S1442: Calculate the theme matching degree between the theme requirement feature vector and the theme feature vector based on the matching of core theme elements, and generate a theme matching degree value.

[0178] The theme matching degree is calculated based on the number and degree of matching of core theme elements. First, a weight is assigned to each core theme element, with the weight determined by the element's importance in the theme, and more important elements receiving higher weights.

[0179] For matching core topic elements, their weights are summed; for non-matching elements, a set discount weight is applied based on their similarity to the requirement elements, and this weight is also summed. Then, the sum is divided by the total weight of the core topic elements in the topic requirement feature vector to obtain the topic matching degree value.

[0180] For example, if the total weight of the core topic elements in the topic demand feature vector is one hundred, the total weight of the matching elements is sixty, and the total discount weight of the non-matching elements is fifteen, then the topic matching degree value is (sixty plus fifteen) divided by one hundred, which gives 0.75.

[0181] Step S1443: Select topic feature vectors with topic matching scores higher than the topic matching score threshold as candidate topic feature vectors.

[0182] Set a topic matching threshold, determined based on the strictness of topic matching. Compare the topic matching score of each topic feature vector with this threshold. If the score is higher than the threshold, the topic feature vector is selected as a candidate topic feature vector.

[0183] Step S1444: Select vectors from the candidate topic feature vectors that meet the preset conditions for topic matching and have unique topic performance to form a topic feature subset. The topic feature subset includes multiple topic feature vectors and their respective topic matching information.

[0184] The preset conditions include minimum requirements for topic matching and vector diversity. First, it ensures that the topic matching values ​​of the selected candidate topic feature vectors are all within a preset high range. Second, to ensure the diversity of topic feature subsets and avoid selecting vectors with overly similar topic performance, the topic similarity between candidate vectors is calculated to filter out vectors with unique topic performance.

[0185] For example, if two candidate topic feature vectors have extremely high topic similarity, it means that the topics they represent are quite similar. In this case, the one with the higher topic matching degree is selected. The final topic feature subset contains multiple topic feature vectors, each with its own topic matching degree information.

[0186] Step S145: Extract the image unit identifiers corresponding to the style feature subset and the theme feature subset, identify the common identifiers that exist simultaneously in the image unit identifiers corresponding to the two subsets, establish the association relationship between style features and theme features based on the common identifiers, and construct a feature association graph. The nodes of the feature association graph represent features, and the edges represent the association strength between features.

[0187] Each style feature vector in the style feature subset corresponds to one or more image units in the basic image material library. The identifiers of these image units are extracted by querying the index system. Similarly, the image unit identifiers corresponding to each theme feature vector in the theme feature subset are extracted.

[0188] The image unit identifiers corresponding to the style feature subset and the theme feature subset are compared to identify common identifiers. The image units corresponding to the common identifiers meet both the user's style requirements and the user's theme requirements.

[0189] Based on common identifiers, corresponding style and theme features are identified, and a relationship is established between them. The strength of the relationship is determined by a comprehensive calculation based on the number of common identifiers, style matching degree, and theme matching degree. For example, the more common identifiers and the higher the matching degree, the stronger the relationship.

[0190] A feature association graph is constructed using style features and theme features as nodes and association strength as edges. The feature association graph visually illustrates the association between style features and theme features.

[0191] Step S146: Input the emotional atmosphere preference in the user's personalized needs data into the emotional feature encoding network and convert it into an emotional feature vector. The emotional feature vector contains feature parameters of multiple emotional dimensions.

[0192] The sentiment feature encoding network consists of a text encoding layer and a feature mapping layer. Users' emotional atmosphere preferences may exist in the form of textual descriptions, such as "tranquil," "passionate," and "melancholy." These textual descriptions are input into the text encoding layer, which converts the text into initial vectors using word embedding techniques.

[0193] The initial vector is input into the feature mapping layer, which processes the initial vector through a multi-layer neural network, mapping it to multiple sentiment dimensions to generate sentiment feature vectors. Sentiment dimensions include, but are not limited to, "pleasure," "activity," and "tension," and the feature parameters of each dimension reflect the user's preference for that sentiment dimension.

[0194] Step S147: Input the feature association map and the sentiment feature vector into the map fusion module, and adjust the association strength in the feature association map through the sentiment feature weighting mechanism to generate a feature association map that integrates sentiment factors.

[0195] The graph fusion module first analyzes the feature parameters of each sentiment dimension in the sentiment feature vector to determine the influence weight of each sentiment dimension on the association between style features and theme features. For example, a preference for the sentiment atmosphere of "tranquility" may increase the weight of the association between the style feature of "soft colors" and the theme feature of "lake".

[0196] Based on the influence weights, the association strength of edges in the feature association graph is adjusted. The adjustment method involves multiplying the original association strength by the corresponding influence weight to obtain the new association strength. Through this sentiment feature weighting mechanism, the feature association graph incorporates users' emotional atmosphere preferences, generating a feature association graph that integrates emotional factors.

[0197] Step S148: Extract key feature association paths from the fused feature association map, and generate style transfer coefficients, theme element arrangement rules and sentiment rendering criteria based on the key paths.

[0198] A path analysis algorithm was used to analyze the fused feature association map, identifying key feature association paths with high association strength and reasonable path length. These key feature association paths reflect the important relationships between style features, theme features, and sentiment features.

[0199] The variation patterns of style features are extracted from the association paths of key features to generate style transfer coefficients. Style transfer coefficients indicate the degree to which each style dimension needs adjustment when converting image units from a base image library into images that meet the user's style requirements.

[0200] Based on the distribution and association of topic features in the key feature association path, rules for the arrangement of topic elements are formulated. These rules include regulations regarding the positional distribution, size ratio, and interrelationships of topic elements.

[0201] Based on the association between emotional features and style / theme features in the key feature association path, emotional rendering criteria are determined. These criteria define how to render the desired emotional atmosphere for the user through color, lighting, composition, and other methods during image compositing.

[0202] Step S149: The style transfer coefficient, the theme element arrangement rules, and the emotion rendering criteria are integrated according to the control logic to generate a personalized synthesis control factor. The personalized synthesis control factor includes the correlation and application order of each control item.

[0203] This study analyzes the intrinsic relationships between style transfer coefficients, subject element arrangement rules, and emotion rendering criteria to determine their regulatory logic in the image compositing process. For example, subject element arrangement rules are the foundation of style transfer and emotion rendering; style transfer needs to be carried out based on subject element arrangement, while emotion rendering is a further optimization of the results of the former two processes.

[0204] Following the control logic, the three control items are integrated, clarifying the relationships and application order among them. For example, the theme element arrangement rules are applied first, followed by the style transfer coefficient, and finally the emotion rendering criteria. The integrated result is the personalized synthesis control factor, which is the core parameter set guiding the image synthesis process.

[0205] Step S150: Based on the personalized synthesis control factor, the matching image units in the basic image material library are recombined and style adapted. The style transfer coefficient is applied through the style transfer network, the theme element arrangement rules are applied through the element arrangement network, and the emotion rendering criteria are applied through the emotion rendering network to generate a synthetic image that meets the user's personalized needs.

[0206] Step S151: Analyze the personalized synthesis control factors, determine the specific content and application scope of the style transfer coefficient, thematic element arrangement rules and emotional rendering criteria, and establish the correspondence between the control factors and the processing network.

[0207] The personalized compositing control factors are analyzed to extract the specific values ​​of style transfer coefficients, detailed clauses of subject element arrangement rules, and specific requirements of emotional rendering criteria. The application scope of each control item is clarified; for example, style transfer coefficients are applicable to adjusting the style features of image units such as color and texture, while subject element arrangement rules are applicable to the position and layout of core objects in image units.

[0208] Based on the function and application scope of each control item, a correspondence is established with the processing network, namely, the style transfer coefficient corresponds to the style transfer network, the theme element arrangement rule corresponds to the element arrangement network, and the emotion rendering criterion corresponds to the emotion rendering network.

[0209] Step S152: Based on the theme element arrangement rules, select image units containing corresponding theme elements from the basic image material library as matching image units, sort them according to theme relevance, and determine the recombined image unit sequence.

[0210] Based on the core thematic elements specified in the thematic element layout rules, a search is performed in the basic image material library to select image units containing these core thematic elements as matching image units. For example, if the thematic element layout rules require the inclusion of both "mountain peaks" and "streams," then image units containing both elements are selected.

[0211] The matched image units are sorted according to their thematic relevance to the arrangement rules of the subject elements. Thematic relevance is calculated based on factors such as the degree of matching, quantity, and distribution of core subject elements within the image units. After sorting, a reorganized sequence of image units is obtained, with the image units at the beginning of the sequence showing higher relevance to the subject requirements.

[0212] Step S153: Input the image unit sequence into the element arrangement network, and arrange the matching image units spatially according to the position arrangement scheme, size setting scheme and hierarchical relationship scheme in the theme element arrangement rules, determine the position, size and stacking order of each image unit in the synthesized image, and generate a preliminary image layout.

[0213] For example, in step S1531, the theme element arrangement rules are parsed, and the position arrangement scheme, size setting scheme and hierarchical relationship scheme are extracted. The position arrangement scheme includes the target position information of each theme element, the size setting scheme includes the size range information of each theme element, and the hierarchical relationship scheme includes the stacking order information between theme elements.

[0214] The rules for arranging the main elements are analyzed in detail, separating the positional arrangement scheme, the size setting scheme, and the hierarchical relationship scheme. The positional arrangement scheme clarifies the target coordinate range of each core main element in the composite image. For example, the "mountain peak" is located in the upper half of the image, and the "stream" extends from the mountain peak to the lower half of the image.

[0215] The sizing scheme specifies the size range for each core theme element, such as the ratio of the width of a "mountain peak" to the image width and the ratio of its height to the image height. The hierarchy scheme determines the stacking order between different theme elements, such as stacking "trees" on top of "ground" and "clouds" on top of "sky".

[0216] Step S1532: Mark the subject element for each matching image unit in the image unit sequence, and identify the position information and size information of the core subject element in the image unit. The position information is used to indicate the relative position of the core subject element in the image unit, and the size information is used to indicate the size of the core subject element.

[0217] Using object detection and localization techniques, each matching image unit in the image unit sequence is processed to mark the core subject elements. Coordinate localization is used to determine the relative position information of the core subject elements within the image unit. This relative position information is represented by the coordinate proportion of the element within the image unit; for example, the top-left corner of a "mountain peak" element has a coordinate proportion of (0.2, 0.1), and the bottom-right corner has a coordinate proportion of (0.8, 0.6).

[0218] At the same time, the size information of the core theme element is calculated. The size information is represented by the width ratio and height ratio of the element in the image unit. For example, the width ratio of the "mountain peak" element is 0.6 and the height ratio is 0.5.

[0219] Step S1533: Based on the target position information in the position arrangement scheme, calculate the positional deviation between the position information of the core subject element in each matching image unit and the target position information, adjust the position of the image unit through the position adjustment mechanism to make the position of the core subject element consistent with the target position information, and determine the position of each image unit.

[0220] The relative position information of the core subject elements in the matched image unit is converted into absolute position information, and then compared with the target position information in the position arrangement scheme to calculate the position deviation. The position deviation includes horizontal deviation and vertical deviation.

[0221] The position adjustment mechanism calculates the adjustment amount based on the positional deviation and performs translation processing on the image unit to gradually bring the position of the core subject element closer to the target position information. After multiple fine adjustments, until the position of the core subject element matches the target position information, the position of the image unit in the composite image is determined.

[0222] Step S1534: Based on the size range information in the size setting scheme and the size information of the core theme element in the image unit, calculate the size adjustment ratio of the image unit, adjust the size of the image unit through the size adjustment mechanism, so that the size of the core theme element conforms to the size range information, and determine the size of each image unit.

[0223] The size information of the core subject element in the image unit is compared with the size range information in the size setting scheme. If the size of the core subject element exceeds or falls below the size range, a size adjustment ratio is calculated. The size adjustment ratio is calculated based on the ratio of the target size to the current size.

[0224] The resizing mechanism scales the image units according to the calculated scaling ratio. During scaling, the aspect ratio of the image units remains constant to avoid image distortion. After adjustment, the size of the core subject element conforms to the size range information, and the size of that image unit in the composite image is then determined.

[0225] Step S1535: Based on the superposition order information in the hierarchical relationship scheme, sort all matching image units hierarchically to determine the superposition order of each image unit in the synthesized image. Image units corresponding to elements with high theme importance are in the upper layer, and those with low theme importance are in the lower layer.

[0226] The hierarchical relationship scheme clearly defines the importance ranking of different thematic elements, and the matching image units are hierarchically sorted according to this ranking. Image units corresponding to core thematic elements with high thematic importance are assigned higher levels and are located in the upper layer of the composite image; those with low thematic importance are assigned lower levels and are located in the lower layer.

[0227] For example, if in the hierarchical relationship scheme, "mountain peak" is more important than "stream", and "stream" is more important than "stone", then the image unit containing "mountain peak" is at the top layer, the one containing "stream" is at the middle layer, and the one containing "stone" is at the bottom layer.

[0228] Step S1536: According to the determined position, size and stacking order, place all matching image units into the preset canvas space to generate the initial image layout.

[0229] The size and aspect ratio of the preset canvas space are determined based on common image output requirements. Following the previously determined position, size, and stacking order of each image unit, the image units are placed one by one into their corresponding positions within the canvas space.

[0230] During placement, ensure that the positional relationship between image units conforms to the requirements of the subject element layout rules, the stacking order is correct, and there are no hierarchical errors. After placement, an initial image layout is formed.

[0231] Step S1537: Perform a spatial rationality analysis on the initial image layout to identify whether there is spatial conflict or waste between image units. If there is spatial conflict, adjust the position or size of the relevant image units. If there is waste, fill the blank area by supplementing elements to optimize the image layout.

[0232] Spatial rationality analysis includes spatial conflict detection and space waste detection. Spatial conflict detection is achieved by calculating the overlapping area between image units. If the overlapping area exceeds a preset threshold, a spatial conflict is determined to exist. For image units with spatial conflicts, the conflict is eliminated by fine-tuning their position or appropriately adjusting their size.

[0233] Space waste detection is achieved by calculating the area of ​​blank areas in the canvas space that are not covered by image units. If the area of ​​blank areas exceeds a preset ratio, space waste is determined to exist. For areas with space waste, auxiliary element image units related to the theme, such as "birds," "rocks," and "grass," are selected from the basic image material library. These auxiliary element image units are then filled into the blank areas according to the supplementary element placement principle in the theme element arrangement rules to optimize the image layout.

[0234] Step S1538: Perform visual balance analysis on the optimized image layout and adjust the positional distribution of image units.

[0235] Visual balance analysis starts with the overall layout of the image, considering factors such as the distribution density, perceived weight, and visual center of gravity of image units on the canvas. By calculating the distribution center of gravity of image units in the horizontal and vertical directions, it determines whether the visual center of gravity of the image is in a reasonable position. If the visual center of gravity deviates too much from the center of the canvas, the position of the relevant image units needs to be adjusted.

[0236] For example, if there are too many image units on the left side of an image, creating an overly heavy visual effect that shifts the visual center of gravity to the left, then some of the image units on the left can be moved to the right, or the size of the image units on the right can be adjusted to increase their visual weight, thus balancing the visual center of gravity of the image. At the same time, observe whether the spacing between the image units is uniform, avoiding areas of overcrowding or excessive sparseness, and fine-tune their positions to make the overall layout more harmonious.

[0237] Step S1539: Convert the adjusted image layout into structured layout data containing the position, size and hierarchy information of each image unit to generate a preliminary image layout.

[0238] The position information (represented by coordinates in the canvas coordinate system), size information (width and height), and layer information (overlay order number) of each image unit in the adjusted image layout are extracted and organized. This information is then organized into structured layout data according to a preset data format. This structured layout data can be stored in tabular form, where each row corresponds to the relevant information of one image unit, and each column corresponds to attributes such as position, size, and layer.

[0239] The preliminary image layout generated in the above manner records the spatial distribution and hierarchical relationship of all image units in the synthesized image.

[0240] Step S154: Extract the style features of each matching image unit in the preliminary image layout, compare them with the style feature subset, identify style difference features, and perform style transfer processing on the matching image units through a style transfer network based on the style difference features and style transfer coefficients to adjust the color combination, brush stroke expression and layout structure of the image units so that the style of all image units remains consistent.

[0241] First, for each matching image unit in the initial image layout, the same deep learning model as in step S120 is used to extract its style features, including color combination patterns, brushstroke expression forms and layout structure features, to generate the current style feature vector of each image unit.

[0242] These current style feature vectors are compared with the style feature subset. The comparison process is similar to the vector comparison method in step S1421, calculating the similarity dimension by dimension to identify differences in dimensions such as color, texture, and layout, thus forming style difference features. Style difference features are represented by multi-dimensional difference vectors, with the value of each dimension reflecting the degree of difference of the corresponding style feature.

[0243] The style transfer network comprises a color conversion module, a brushstroke adjustment module, and a layout optimization module. The color conversion module adjusts the color combinations of image units based on the color difference component of the style difference features and the style transfer coefficient. For example, if the difference features indicate that the dominant hue of an image unit is inconsistent with the dominant hue of the style feature subset, and the corresponding color adjustment has a high weight in the style transfer coefficient, then the hue, saturation, and brightness of that image unit are adjusted to bring its dominant hue closer to the dominant hue of the style feature subset, while maintaining a natural color transition.

[0244] The brushstroke adjustment module modifies the brushstroke representation of image units based on brushstroke differences and style transfer coefficients in the style difference features. For areas where the brushstroke direction does not match the target style, the direction and density distribution of the brushstroke are adjusted; for parts where the brushstroke thickness does not meet the requirements, it is adjusted according to the thickness variation curve to match the brushstroke representation with the brushstroke features of the style feature subset.

[0245] The layout optimization module fine-tunes the layout architecture of image units based on layout differences and style transfer coefficients in the style difference features. For example, if the position of the core visual element of an image unit deviates from the target layout, its position is adjusted appropriately without affecting the overall arrangement, so that the layout architecture better conforms to the layout requirements of the style feature subset.

[0246] Through the processing of the style transfer network, the style features of all matching image units are unified, ensuring a consistent style tone in color, brushstrokes, and layout.

[0247] Step S155: Input the style-transformed image layout into the emotion rendering network. Based on the tone adjustment scheme, lighting and shadow setting scheme and atmosphere enhancement scheme in the emotion rendering criteria, perform emotion atmosphere rendering processing on the image layout, adjust the overall tone, add lighting and shadow effects and enhance atmosphere features.

[0248] The emotion rendering network consists of a tone adjustment layer, a lighting and shadow generation layer, and an atmosphere enhancement layer. The tone adjustment layer adjusts the overall tone of the style-transformed image layout according to the tone adjustment scheme in the emotion rendering guidelines. For example, if the preferred emotional atmosphere is "tranquil and serene," the tone adjustment scheme might require an overall tone leaning towards cool tones (such as blue or cyan). The tone adjustment layer then adjusts the color components of each pixel in the image to shift the overall tone towards cooler tones while maintaining color harmony between different image units.

[0249] The lighting and shadow generation layer adds appropriate lighting and shadow effects to the image layout based on the lighting and shadow settings. First, it determines the position and type of the light source, such as natural light or artificial light. Then, based on the characteristics of the light source, it calculates the lit and shadowed surfaces of each object in the image, generating a lighting and shadow distribution map. Brightness is increased in the lit areas to enhance contrast and make details in those areas clearer; brightness is decreased in the shadowed areas to increase shadow effects and enhance the image's depth and three-dimensionality.

[0250] The atmosphere enhancement layer, based on the atmosphere enhancement scheme, strengthens specific areas in the image layout to highlight emotional atmosphere characteristics. For example, for an "autumn forest" theme with a "warm and peaceful" emotional atmosphere preference, the atmosphere enhancement layer will enhance the orange-yellow tones of the fallen leaves in the forest and increase the brightness of the light spots formed by sunlight filtering through the leaves, making the warm atmosphere more prominent. At the same time, for areas less associated with the emotional atmosphere, their visual weight is appropriately reduced to avoid distracting from the main atmosphere.

[0251] Through the processing of the emotion rendering network, the emotional atmosphere of the image layout is effectively enhanced, making it more in line with the user's emotional atmosphere preferences.

[0252] Step S156: Perform detail optimization on the image layout after style transfer and emotion rendering. Eliminate splicing marks between image units through edge blending. Adjust the color connection between adjacent areas through color transition processing. Perform overall feature check on the optimized image to identify whether there are areas with missing subject elements or inconsistent styles. If so, make local adjustments according to the personalized synthesis control factor to supplement missing subject elements or correct areas with inconsistent styles. Then, output the adjusted image as a synthesized image that meets the user's personalized needs.

[0253] The detailed optimization process begins with edge blending. For the stitching edges between image units, an edge detection algorithm is used to identify the edge's position and shape. Then, through pixel interpolation and blurring, the pixel values ​​at the edges of adjacent image units gradually transition, eliminating obvious stitching artifacts. For example, for the edges of two adjacent image units, the pixel values ​​at the edges are weighted and averaged according to their distance from the edge, making the pixel value changes on both sides of the edge smoother.

[0254] Color transition processing is used to adjust the color transition between adjacent areas. It analyzes the color differences between adjacent image units; if the differences are significant, a gradient adjustment is applied to the color in the transition area to make the color change from one image unit to another natural and smooth. For example, if one image unit has a greenish tint and an adjacent image unit has a yellowish tint, a gradient from green to yellow is set in the transition area to avoid abrupt color changes.

[0255] The overall feature check is performed using feature extraction and comparison. Thematic and style features of the optimized image are extracted and compared with the thematic element specifications and style tendency descriptions in the user's personalized requirements data to identify whether there are missing thematic elements, such as the absence of the "small bridge" element, or areas where the style features are inconsistent with the target style, such as the brushstroke style of a certain part not matching the overall impressionistic style.

[0256] If there are missing thematic elements, the corresponding thematic element image units are selected from the basic image material library according to the thematic element arrangement rules in the personalized synthesis control factor and added to the appropriate position; if there are areas with inconsistent styles, the style of the area is locally adjusted according to the style transfer coefficient to make it consistent with the overall style.

[0257] After the above processing, the adjusted image is output as the final composite image, which meets the user's personalized needs in terms of subject matter, artistic style, and emotional atmosphere.

[0258] Figure 2 The illustration shows a schematic diagram of the hardware structure of an AI-based personalized image synthesis system 100 for implementing the above-described AI-based personalized image synthesis method, as provided in an embodiment of the present invention. Figure 2 As shown, the AI-based personalized image synthesis system 100 may include a processor 110, a machine-readable storage medium 120, a bus 130, and a communication unit 140.

[0259] Machine-readable storage medium 120 may store data and / or instructions. In some embodiments, machine-readable storage medium 120 may store data acquired from an external terminal. In some embodiments, machine-readable storage medium 120 may store data and / or instructions used by the AI-based personalized image synthesis system 100 to perform or use in order to accomplish the exemplary methods described in this invention.

[0260] In a specific implementation, one or more processors 110 execute computer-executable instructions stored in the machine-readable storage medium 120, enabling the processor 110 to execute the AI-based personalized image synthesis method as described in the above method embodiment. The processor 110, the machine-readable storage medium 120, and the communication unit 140 are connected via a bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0261] The specific implementation process of processor 110 can be found in the various method embodiments executed by the artificial intelligence-based personalized image synthesis system 100 described above. The implementation principle and technical effect are similar, and will not be repeated here.

[0262] Furthermore, embodiments of the present invention also provide a readable storage medium containing computer-executable instructions. When a processor executes the computer-executable instructions, the above-described personalized image synthesis method based on artificial intelligence is implemented.

[0263] It should be noted that, in order to simplify the description of this invention and thus aid in the understanding of one or more embodiments, the foregoing description of the embodiments of this invention sometimes combines multiple features into a single embodiment, drawing, or description thereof. Similarly, it should be noted that, in order to simplify the description of this invention and thus aid in the understanding of one or more embodiments, the foregoing description of the embodiments of this invention sometimes combines multiple features into a single embodiment, drawing, or description thereof.

Claims

1. A personalized image synthesis method based on artificial intelligence, characterized in that, The method includes: Obtain a basic image library containing diverse art styles and themes. The basic image library contains multiple sets of image units with style attribute tags and theme content tags. The style attribute tags are used to mark the art style characteristics to which the image unit belongs, and the theme content tags are used to mark the core content theme presented by the image unit. The basic image material library is subjected to feature parsing operation by a deep learning model to extract style representation elements, theme composition elements and element association features in the image unit, and integrate them to generate style feature library and theme feature library. User personalized needs data are collected through an interactive input interface. The user personalized needs data includes style preference descriptions, thematic element specifications, and emotional atmosphere preferences. The style feature library, the theme feature library, and the user personalized demand data are input into the feature fusion network. A style feature subset and a theme feature subset that match the user demand are generated through a feature mapping mechanism. A feature association map is constructed based on the style feature subset and the theme feature subset. Personalized synthetic control factors are generated through map fusion. Based on the personalized synthesis control factor, the matching image units in the basic image material library are recombined and style adapted. The style transfer coefficient is applied through the style transfer network, the theme element arrangement rules are applied through the element arrangement network, and the emotion rendering criteria are applied through the emotion rendering network to generate a synthetic image that meets the user's personalized needs.

2. The personalized image synthesis method based on artificial intelligence according to claim 1, characterized in that, The step involves performing feature parsing on the basic image material library using a deep learning model to extract style representation elements, theme composition elements, and element association features from image units, and integrating them to generate a style feature library and a theme feature library, including: Each image unit in the basic image material library is input into the feature extraction layer of a convolutional neural network. The pixel distribution pattern of the image unit is analyzed through multiple convolution operations to generate a feature map. The feature map includes a color distribution feature map, a texture representation feature map, and a spatial layout feature map. Extract color combination patterns from the color distribution feature map, identify the matching relationship between the main color and the auxiliary color in the image unit, the color transition mode and the color proportion feature, and generate style representation elements of the color dimension. Extract the brushstroke representation from the texture representation feature map, identify the direction pattern, thickness variation and texture density features of the brushstrokes in the image unit, and generate style representation elements of the texture dimension. The layout architecture features are extracted from the spatial layout feature map, and the positional distribution, area proportion and spatial relationship between core visual elements in the image unit are identified to generate style representation elements of the layout dimension. The style representation elements of color, texture and layout dimensions are combined according to feature correlation to generate style feature vectors for each image unit, and all style feature vectors are aggregated to generate a style feature library. By analyzing the core object morphology in image units through an object detection network, the contour features, morphological structure and detail representation of the object are identified, and the theme composition elements of the core object dimension are generated. The scene parsing network analyzes the scene environment composition in the image unit, identifies the environmental elements, spatial structure and atmosphere features in the scene, and generates the theme composition elements of the scene environment dimension. By analyzing the interaction relationships of elements in image units through a relation recognition network, the core object and environmental elements, the positional association, interaction state and dependency pattern between objects are identified, and the theme constituent elements of the relation dimension are generated. The thematic components of the core object dimension, scene environment dimension, and relationship dimension are integrated according to feature importance to generate the theme feature vector of each image unit, and all theme feature vectors are aggregated to generate a theme feature library. The intrinsic relationship between feature vectors in the style feature library and the theme feature library is analyzed by feature association network analysis, the co-occurrence pattern of style features and theme features is identified, and an element association feature table is generated. The element association feature table is used to record the matching relationship between different style features and theme features.

3. The personalized image synthesis method based on artificial intelligence according to claim 2, characterized in that, The step of extracting brushstroke representation from the texture representation feature map, identifying the direction pattern, thickness variation, and texture density features of brushstrokes in image units, and generating style representation elements for the texture dimension includes: The texture representation feature map is input into the multi-scale feature analysis module, and features are extracted through convolution kernels with different receptive fields to generate a multi-scale texture feature map. The multi-scale texture feature map includes a fine texture layer, a medium texture layer and a coarse texture layer. The fine texture layer is used to identify the detailed expression of brushstrokes, extract the edge contour features, distribution of turning nodes and local directional changes of brushstrokes, and generate basic data on the brushstroke directional patterns. The transition features of the brushstrokes are identified from the medium texture layer, and the width variation, density transition and connection mode of the brushstrokes in different areas are extracted to generate core data of brushstroke thickness variation. The overall distribution of brushstrokes is identified from the coarse texture layer, and the coverage, distribution density, and regional correlation of the texture region are extracted to generate key data of texture density features. The basic data of the brushstroke direction pattern is subjected to directional clustering processing to identify the main direction, secondary direction and direction conversion nodes, and generate a direction pattern descriptor. The core data of the stroke thickness variation is analyzed to identify the range, frequency and associated region of the thickness variation, and a thickness variation descriptor is generated. The key data of the texture density features are analyzed for density distribution to identify the distribution range of high-density, medium-density and low-density regions, and a density feature descriptor is generated. The direction pattern descriptor, the thickness variation descriptor, and the density feature descriptor are combined according to the hierarchical relationship of texture features to generate style representation elements of the texture dimension. The style representation elements of the texture dimension include a brushstroke direction map, a thickness variation curve, and a density distribution matrix.

4. The personalized image synthesis method based on artificial intelligence according to claim 2, characterized in that, The process involves parsing the interaction relationships between elements in an image unit using a relationship recognition network, identifying the positional associations, interaction states, and dependency patterns between core objects and environmental elements, and generating thematic constituent elements of the relationship dimension, including: The image unit is input into the relationship recognition network to encode the features of the core object and environmental elements, generating object feature vectors and environmental feature vectors. Calculate the spatial relationship between the object feature vector and the environment feature vector, identify the relative position, orientation, and distance features of the core object in the environment, and generate a position association descriptor; Analyze the action interaction features between the core object and other objects, identify the type, intensity and duration of the interaction actions, and generate an interaction state descriptor; Analyze the functional relationships between core objects and environmental elements, identify the supporting, containing, and complementary relationships between elements, and generate dependency pattern descriptors; The location association descriptor, the interaction state descriptor and the dependency pattern descriptor are analyzed to identify the intrinsic relationship between the three types of relationship descriptors and generate a relationship association matrix. Based on the relationship association matrix, the three types of relationship descriptors are integrated to generate topic components of the relationship dimension. The topic components of the relationship dimension include a positional relationship graph, an interaction state sequence, and a dependency pattern model.

5. The personalized image synthesis method based on artificial intelligence according to claim 1, characterized in that, The process of collecting personalized user needs data through an interactive input interface includes style preference descriptions, thematic element specifications, and emotional atmosphere preferences, including: An interactive input interface is constructed, which includes a style preference selection area, a theme element specification area, and an emotional atmosphere adjustment area. The style preference selection area provides style example images and style feature description text. The theme element specification area supports element category selection, feature description input, and reference image upload. The emotional atmosphere adjustment area provides emotional vocabulary options and atmosphere example images. Monitor the user's operation behavior in the style preference selection area, record the style identifier corresponding to the style example image selected by the user, collect the style feature description text input by the user, perform semantic parsing on the style feature description text, extract style keywords and feature descriptions, and integrate the style identifier with the style keywords and feature descriptions into a style preference description. In response to user input in the designated area of ​​the topic element, if the user selects an element category, the selected element category identifier and feature parameters are recorded; if the user inputs a feature description, element features are extracted from the description text to generate element feature parameters; if the user uploads a reference image, feature parsing is performed on the reference image to extract core element features, and the above-obtained information is summarized to generate the designated topic element. Capture the user's selection behavior in the emotional atmosphere adjustment area, record the emotional words selected by the user and the example image of the atmosphere, and perform emotional feature analysis to determine the user's preferred emotional atmosphere type and expression characteristics, and generate emotional atmosphere preferences; The style preference description, the theme element specification, and the emotional atmosphere preference are correlated and analyzed to generate structured user personalized demand data.

6. The personalized image synthesis method based on artificial intelligence according to claim 1, characterized in that, The process involves inputting the style feature library, the theme feature library, and the user's personalized needs data into a feature fusion network; generating a style feature subset and a theme feature subset matching the user's needs through a feature mapping mechanism; constructing a feature association graph based on the style feature subset and the theme feature subset; and generating personalized synthesis control factors through graph fusion. This includes: The style preference description in the user's personalized needs data is input into a text feature encoding network and converted into a style needs feature vector. The feature matching module matches the style requirement feature vector with the style feature vector in the style feature library, and finds style feature vectors that match the style requirement feature vector with a preset condition, forming a style feature subset. The topic elements in the user personalized demand data are specified as input element feature encoding networks and converted into topic demand feature vectors. The feature matching module matches the topic requirement feature vector with the topic feature vector in the topic feature library, finds the topic feature vector that matches the topic requirement feature vector with a preset condition, and forms a topic feature subset. Extract the image unit identifiers corresponding to the style feature subset and the theme feature subset, identify the common identifiers that exist simultaneously in the image unit identifiers corresponding to the two subsets, establish the association relationship between style features and theme features based on the common identifiers, and construct a feature association graph, where the nodes of the feature association graph represent features and the edges represent the association strength between features; The emotional atmosphere preference in the user's personalized needs data is input into the emotional feature encoding network and converted into an emotional feature vector, which contains feature parameters of multiple emotional dimensions. The feature association map and the sentiment feature vector are input into the map fusion module. The association strength in the feature association map is adjusted through the sentiment feature weighting mechanism to generate a feature association map that integrates sentiment factors. Key feature association paths are extracted from the fused feature association map, and style transfer coefficients, theme element arrangement rules and emotion rendering criteria are generated based on the key paths. The style transfer coefficient, the theme element arrangement rules, and the emotion rendering criteria are integrated according to the control logic to generate a personalized synthesis control factor. The personalized synthesis control factor includes the correlation and application order of each control item.

7. The personalized image synthesis method based on artificial intelligence according to claim 6, characterized in that, The step of matching the style requirement feature vector with the style feature vectors in the style feature library through the feature matching module, and finding style feature vectors that meet the preset conditions for matching degree with the style requirement feature vector, forming a style feature subset, includes: The style requirement feature vector is compared with the style feature vector in the style feature library by inputting it into the vector comparison layer of the feature matching module to perform a dimension-by-dimensional comparison of the vector features and identify the feature similarity in the same dimension. Calculate the overall matching degree between the style requirement feature vector and the style feature vector based on the dimensional feature similarity, and generate a matching degree value; Style feature vectors with matching scores higher than the matching score threshold are selected as candidate style feature vectors. Vectors that meet preset conditions are selected from the candidate style feature vectors to form a style feature subset, which contains multiple style feature vectors and their respective matching degree information.

8. The personalized image synthesis method based on artificial intelligence according to claim 6, characterized in that, The step of matching the topic requirement feature vector with the topic feature vectors in the topic feature library through the feature matching module, and finding topic feature vectors that meet the preset conditions for matching degree with the topic requirement feature vector, forming a topic feature subset, includes: The topic requirement feature vector and the topic feature vector in the topic feature library are input into the topic comparison layer of the feature matching module for structured comparison of topic features to identify the matching status of core topic elements; Calculate the theme matching degree between the theme requirement feature vector and the theme feature vector based on the matching of core theme elements, and generate a theme matching degree value; Feature vectors of topics with a topic matching degree higher than the topic matching degree threshold are selected as candidate topic feature vectors; Vectors that meet the preset conditions for topic matching and have unique topic performance are selected from the candidate topic feature vectors to form a topic feature subset. The topic feature subset contains multiple topic feature vectors and their respective topic matching information.

9. The personalized image synthesis method based on artificial intelligence according to claim 1, characterized in that, The process of recombining and style-adapting matching image units in the basic image material library based on the personalized synthesis control factor, applying style transfer coefficients through a style transfer network, applying subject element arrangement rules through an element arrangement network, and applying emotion rendering criteria through an emotion rendering network, generates a synthesized image that meets the user's personalized needs, including: The personalized synthesis control factors are analyzed to determine the specific content and application scope of style transfer coefficient, theme element arrangement rules and emotional rendering criteria, and to establish the correspondence between control factors and processing networks. Based on the theme element arrangement rules, image units containing corresponding theme elements are selected from the basic image material library as matching image units, sorted according to theme relevance, and the recombined image unit sequence is determined. The image unit sequence is input into the element arrangement network. Based on the position arrangement scheme, size setting scheme and hierarchical relationship scheme in the theme element arrangement rules, the matching image units are spatially arranged to determine the position, size and stacking order of each image unit in the synthesized image, and a preliminary image layout is generated. The style features of each matching image unit in the preliminary image layout are extracted and compared with the style feature subset to identify style difference features. Based on the style difference features and style transfer coefficient, the matching image units are style-transformed through a style transfer network to adjust the color combination, brush stroke expression and layout structure of the image units so that the style of all image units is consistent. The style-transformed image layout is input into the emotional rendering network. Based on the tone adjustment scheme, lighting and shadow setting scheme and atmosphere enhancement scheme in the emotional rendering criteria, the image layout is processed for emotional atmosphere rendering, adjusting the overall tone, adding lighting and shadow effects and enhancing atmosphere features. The layout of the image after style transfer and emotional rendering is optimized in detail. Edge blending is used to eliminate splicing marks between image units, color transition is used to adjust the color connection between adjacent areas, and the optimized image is checked for overall features to identify areas with missing subject elements or inconsistent styles. If such areas exist, local adjustments are made according to the personalized synthesis control factors to supplement missing subject elements or correct areas with inconsistent styles. The adjusted image is then output as a synthesized image that meets the user's personalized needs.

10. A personalized image synthesis system based on artificial intelligence, characterized in that, The device includes a processor and a memory, the memory being connected to the processor. The memory is used to store programs, instructions, or code, and the processor is used to run the programs, instructions, or code in the memory to implement the artificial intelligence-based personalized image synthesis method according to any one of claims 1-9.

Images