Thangka school image retrieval based on multi-attribute features
Yukai Xian, Yunjie Xiang, Xin Yang, Qijun Zhao, Xianmu Cairang
Abstract
Thangka is a traditional Tibetan art form rich in religious and cultural symbolism, with distinct styles across different schools. Accurately retrieving Thangka images by school is challenging due to subtle stylistic differences, inter-school similarity, and limited labeled data. This study proposes a deep learning-based retrieval framework combining global context, color style, and multi-granularity features. A novel multi-attribute fusion network is developed, incorporating attention mechanisms, Vision Transformers, and style representation layers. Features from different branches are integrated using graph convolutional networks to enhance representation and retrieval performance. To address dataset imbalance, targeted data augmentation and style transfer techniques are applied. Experimental results on a curated seven-school Thangka dataset demonstrate the effectiveness of our approach, achieving a mean Average Precision (mAP) of 72.51% and rank-1 accuracy of 79.26%, surpassing existing methods. This work aids digital Thangka preservation and offers insights into heritage-focused image retrieval techniques.