CN-121980047-A - Art design material recommendation method based on artificial intelligence

CN121980047ACN 121980047 ACN121980047 ACN 121980047ACN-121980047-A

Abstract

The invention discloses an artistic design material recommendation method based on artificial intelligence, which belongs to the technical field of material recommendation and specifically comprises the steps of firstly, collecting a design task text, a reference image and a history log of a user, extracting features by a multi-mode encoder and fusing to generate a user intention vector, secondly, combining an image visual primitive extracted by a deep neural network and a pre-defined artistic style ontology library to construct a global design knowledge map containing physical features and abstract semantic nodes, then, projecting the intention vector into a map space, positioning related nodes to generate a user intention sub-graph, then, calculating feature propagation weights of the materials and the intention sub-graph based on a graph attention network, aggregating to generate a semantic association score, on the basis, constructing a filter mask by using a hard constraint index, removing non-compliance materials, calculating final confidence, and finally, sequencing according to the confidence level to output a recommendation list containing semantic tags, thereby realizing accurate and interpretable material recommendation.

Inventors

FENG BEIBEI
TONG XIAO
SHI YUNYUN

Assignees

河南建筑职业技术学院

Dates

Publication Date: 20260505
Application Date: 20260123

Claims (8)

1. An artistic design material recommendation method based on artificial intelligence is characterized by comprising the following steps: s1, acquiring a design task description text, a reference style image and a history interaction log of a user, extracting features by using a multi-modal encoder and splicing to generate a multi-modal intention vector of the user; S2, extracting visual primitives of the material library image by using a deep neural network, and constructing a global design knowledge graph comprising physical feature nodes and abstract semantic nodes by combining a predefined artistic style ontology library; S3, projecting the user multi-mode intention vector to a feature space where the global design knowledge graph is located, positioning semantic nodes related to the intention vector in a high-dimensional manner through a node activation algorithm, and generating a user intention subgraph; S4, calculating feature propagation weights between candidate material nodes in a material library and semantic nodes in a user intention subgraph based on a graph attention network, and aggregating to generate semantic association scores of each candidate material; s5, constructing a filter mask according to the hard constraint index in the design task description text, multiplying the semantic association scores by elements to remove the non-compliant materials, and calculating the final recommendation confidence; And S6, performing descending order arrangement on the residual materials according to the final recommendation confidence, and selecting a material set with the highest confidence to generate an artistic design material recommendation list containing the corresponding semantic tags.
2. The artistic design material recommendation method based on artificial intelligence according to claim 1, wherein in the step S1, the process of collecting the design task description text, the reference style image and the history interaction log of the user, extracting features by using a multi-modal encoder and generating the multi-modal intention vector of the user by splicing is as follows: Monitoring a design task description text input by a user and an uploaded reference style image by a design software front-end interaction interface in real time, calling a historical interaction log of the user from a background database according to a unique user identifier, and respectively performing denoising, size normalization and serialization preprocessing on the text, the image and log data; Inputting the preprocessed design task description text into a pre-training text encoder to generate text semantic vectors, inputting a reference style image into a deep convolutional neural network to extract visual feature vectors, mapping a historical interaction log into an embedded representation, and inputting the embedded representation into the convolutional neural network to generate time sequence interaction behavior vectors; And performing tensor splicing operation on the generated text semantic vector, visual feature vector and time sequence interaction behavior vector along the feature dimension, inputting the spliced joint feature vector into a multi-layer perceptron to perform nonlinear feature fusion and dimension mapping, and outputting a user multi-mode intention vector representing the current requirement of a user.
3. The method for recommending artistic design materials based on artificial intelligence according to claim 1, wherein in the step S2, the process of using the deep neural network to extract visual primitives of the material library image and combining the predefined artistic style ontology library to construct the global design knowledge graph comprising physical feature nodes and abstract semantic nodes is as follows: Creating a material entity node with a unique identifier for each material image in a material library; inputting the images of the material library into a deep neural network to extract color and texture features, mapping the features into visual primitives in a vector form and instantiating the visual primitives into physical feature nodes, and establishing attribute association edges between the material entity nodes and the corresponding physical feature nodes; analyzing the concept of style and emotion tendency in a predefined artistic style ontology library, and generating abstract semantic nodes with unique identifiers according to concept entities; Calculating the semantic category to which the physical feature node belongs by utilizing a multi-label classification algorithm, and establishing a semantic association edge between the physical feature node and the corresponding abstract semantic node; And establishing topological connection edges between abstract semantic nodes according to the ontology library level logic, and integrating all nodes and associated edges to form a global design knowledge graph.
4. The artistic design material recommendation method based on artificial intelligence according to claim 1, wherein in the step S3, the process of projecting the user multi-modal intention vector to the feature space where the global design knowledge graph is located, locating semantic nodes related to the intention vector in a high-dimensional manner through a node activation algorithm, and generating the user intention subgraph is as follows: Utilizing a multi-layer perceptron to construct a feature mapping layer, and converting a user multi-mode intention vector into a projection vector consistent with the embedding dimension of the global design knowledge graph node; calculating dot products of the projection vectors and the embedded vectors of all semantic nodes in the global design knowledge graph, and taking the operation result as an associated activation value for measuring the node correlation; performing descending order arrangement on semantic nodes according to the associated activation values, selecting a head node set, and marking the set as a core activation node responding to user intention; and traversing the first-order neighbor nodes and the connecting edges of the global design knowledge graph retrieval core activation node, and combining to generate a user intention subgraph representing the current specific requirement of the user.
5. The artistic designing material recommendation method based on artificial intelligence according to claim 1, wherein in the step S4, the process of calculating the feature propagation weight between the candidate material nodes in the material library and each semantic node in the user intention subgraph based on the graph attention network and aggregating to generate the semantic association score of each candidate material is as follows: Setting material entity nodes corresponding to candidate materials to be recommended in a material library as central nodes, connecting semantic nodes in user intention subgraphs, and constructing a local topological structure for executing a message transmission mechanism; the feature vectors of the central node and the semantic nodes are spliced and input into the attention layer of the graph, and the original attention coefficients for representing the importance among the nodes are calculated by using a shared weight matrix; Activating an original attention coefficient by adopting a linear unit with leakage correction, and executing normalized exponential function operation to generate a characteristic propagation weight in a probability distribution form; and performing weighted summation on the feature association degree of the semantic nodes according to the feature propagation weight, and aggregating and outputting semantic association scores representing the matching degree of the candidate materials and the user intention.
6. The artistic design material recommendation method based on artificial intelligence according to claim 5, wherein the feature vectors of the spliced central node and the semantic nodes are input into the attention layer of the graph, and the specific way of calculating the original attention coefficient representing the importance among the nodes by using the shared weight matrix is as follows: respectively performing linear mapping on the feature vectors of the central node and the semantic node by using a shared linear transformation matrix preset in the attention layer of the graph, performing tensor stitching on the two groups of mapped feature vectors along the channel dimension, and generating node pair joint feature vectors containing bidirectional feature information; The node pair joint feature vectors are input into a single-layer feedforward neural network, dot product operation is carried out on the joint feature vectors by utilizing the shared attention weight vectors, and un-normalized scalar values are output to serve as original attention coefficients for representing importance among the nodes.
7. The artistic design material recommendation method based on artificial intelligence according to claim 1, wherein in the step S5, the process of constructing a filter mask according to the hard constraint index in the design task description text, performing the element multiplication operation on the semantic association score to remove the non-compliant material, and calculating the final recommendation confidence is as follows: Analyzing the design task description text by using the sequence annotation model, extracting resolution threshold, file format and color mode parameters, and generating a structured hard constraint index set; Searching candidate material metadata, performing logic comparison with the hard constraint index set, and respectively assigning one or zero according to whether all constraint conditions are met or not to construct a binary filter mask vector; Performing element-wise multiplication operation of the binary filter mask vector and the semantic association score vector, and forcibly zeroing the non-compliance material score to generate a corrected semantic association score vector; And executing maximum and minimum normalization processing on the corrected semantic association score vector, linearly mapping the numerical value to a unit interval, and outputting final recommendation confidence level for representing the final availability of the material.
8. The artistic design material recommendation method based on artificial intelligence according to claim 1, wherein in the step S6, the remaining materials are arranged in descending order according to the final recommendation confidence, and the process of selecting the material set with the highest confidence to generate the artistic design material recommendation list including the corresponding semantic tag is as follows: Performing descending order arrangement on the residual materials according to the final recommendation confidence by using a rapid ordering algorithm, and intercepting a preset number of materials at the top of an ordering sequence to form a preferred material set; Tracing the connection relation of each material node in the optimized material set in the user intention subgraph, extracting the abstract semantic node name with the highest feature propagation weight, and generating a corresponding semantic label; And carrying out structural association on the optimized material image data and the corresponding semantic tags, and assembling according to the confidence sequence to generate an artistic design material recommendation list.

Description

Art design material recommendation method based on artificial intelligence Technical Field The invention relates to the technical field of material recommendation, in particular to an artistic design material recommendation method based on artificial intelligence. Background The rapid development of artificial intelligence technology is increasingly widely applied in the field of art design, and the art design material recommendation technology is used as a key support for improving the design efficiency, and has become a hotspot direction for industrial research and application. Currently, an art design material recommendation technology based on artificial intelligence mainly relies on core technologies such as big data processing, computer vision and natural language processing, and provides material retrieval and recommendation services for users by carrying out feature extraction and classification management on mass art design materials. In the prior art, related recommendation schemes are mostly developed around the matching of user input information and material characteristics, visual characteristics such as colors, textures, compositions and the like of materials are extracted by means of an image recognition technology, or design demand description input by a user is processed by a text analysis technology, so that basic material screening and pushing are realized, the method is widely applied to various artistic creation scenes such as plane design, digital drawing, visual transmission and the like, and a rich material resource acquisition channel is provided for designers. However, existing art design material recommendation techniques have significant drawbacks in capturing the implicit authoring intent of the user. The technology mainly relies on keywords or basic visual labels explicitly input by a user to perform shallow feature matching, and cannot construct deep semantic association between a design task and materials. Because design requirements are often implicit in unstructured multimodal descriptions, existing algorithms have difficulty crossing the cognitive gap between underlying physical features and high-level abstract semantics (e.g., style guidance, emotional tendency, cultural context), and lack efficient modeling and cross-modal alignment mechanisms for dynamic design intent. The recommendation result is often limited to be similar in appearance due to the fact that the recommendation result is lack of semantic understanding, real expectations of users cannot be responded on the creative kernel, and design efficiency and accurate acquisition of creation inspiration are severely restricted. Disclosure of Invention The invention aims to provide an art design material recommending method based on artificial intelligence, which solves the problems in the background technology: The aim of the invention can be achieved by the following technical scheme: an art design material recommendation method based on artificial intelligence comprises the following steps: s1, acquiring a design task description text, a reference style image and a history interaction log of a user, extracting features by using a multi-modal encoder and splicing to generate a multi-modal intention vector of the user; S2, extracting visual primitives of the material library image by using a deep neural network, and constructing a global design knowledge graph comprising physical feature nodes and abstract semantic nodes by combining a predefined artistic style ontology library; S3, projecting the user multi-mode intention vector to a feature space where the global design knowledge graph is located, positioning semantic nodes related to the intention vector in a high-dimensional manner through a node activation algorithm, and generating a user intention subgraph; S4, calculating feature propagation weights between candidate material nodes in a material library and semantic nodes in a user intention subgraph based on a graph attention network, and aggregating to generate semantic association scores of each candidate material; s5, constructing a filter mask according to the hard constraint index in the design task description text, multiplying the semantic association scores by elements to remove the non-compliant materials, and calculating the final recommendation confidence; And S6, performing descending order arrangement on the residual materials according to the final recommendation confidence, and selecting a material set with the highest confidence to generate an artistic design material recommendation list containing the corresponding semantic tags. In the step S1, the process of collecting the design task description text, the reference style image and the history interaction log of the user, extracting the characteristics by using the multi-modal encoder and splicing to generate the multi-modal intention vector of the user is as follows: Monitoring a design task description text input by a user and an upload