CLIP-ViL with Contrastive Learning for Automated Semantic Labeling of Digital Art

With the rapid development of digital media technology, the number of digital artworks and intricate textile designs continues to grow, creating demand for efficient and accurate semantic annotation methods for management, retrieval, and dissemination. Existing digital art annotation methods suffer from low semantic alignment accuracy, insufficient generalization, and weak adaptability to artistic style diversity. To address these issues, this paper proposes an automated semantic labeling method for digital art based on the CLIP-ViL model combined with contrastive learning. The multimodal feature extraction framework of CLIP-ViL is first optimized to enhance the extraction of artistic style and semantic concept features. A cross-modal contrastive learning module is then designed to achieve fine-grained alignment between visual features and semantic labels. Finally, a multi-scale feature fusion strategy is introduced to improve recognition of multi-level semantic information. Experiments on ArtEmis, WikiArt, and DigitalArt-1M show that the method achieves average annotation accuracy of 92.3%, exceeding traditional methods by 18.7 percentage points and standalone CLIP by 9.5 percentage points. Recall, F1-score, and mAP reach 89.6%, 90.9%, and 91.8%, respectively.