Image-driven batik product knowledge graph construction

A text-based knowledge graph is limited to providing a basic textual introduction and retrieval of intangible cultural heritage, which hinders the in-depth exploration of the intricate artistic connotations of batik products. This study focuses on batik products and employs an image-driven approach to construct a knowledge graph. Initially, we conduct image data collection and processing for batik products, establishing an ontology specific to the domain of batik using an image-driven approach that defines classes and relationships within the domain. Subsequently, we construct a knowledge graph for batik products by identifying entities and relationships. We employ a method to obtain representative pattern entities through image clustering, where patterns of the same type are clustered again to reveal evolutionary relationships based on clustering results. Subsequently, the DySnake convolution module is integrated into the backbone network of the object detection model ResNet. Moreover, a channel gate control mechanism based on low-rank tensor fusion is proposed to establish an automatic extraction object detection model for batik patterns and facilitate knowledge graph automatic updates. Finally, we visualize the knowledge graph using Gephi and Neo4j. Through comparative experiments and ablation experiments on our self-built batik dataset with multiple object detection models, we validate the effectiveness of our methods.