Research on the Construction of Digital Art Graphics Classification and Retrieval Framework under Machine Learning Technology

To improve the classification and retrieval accuracy of digital art graphics with complex style attributes, this study proposes a joint framework integrating style and structural embeddings. A multi-scale residual network with channel attention is employed for feature extraction, and dual-branch representations are constructed for stylistic semantics and geometric structure. A multitask classification model and a dual-tower cross-modal retrieval structure are designed. A two-level retrieval strategy—style-based indexing and structure-based matching—supports efficient similarity fusion, while contrastive learning enhances embedding alignment. Experimental results show that the framework surpasses existing methods in multi-label classification, embedding clustering, and cross-modal retrieval, particularly in distinguishing similar styles and supporting text-guided queries. These findings validate the effectiveness and scalability of the dual-branch embedding approach for complex art graphic semantics.