A Vision Transformer with Channel-Spatial Attention for Thangka Image Classification

Thangka, as an important carrier of Tibetan culture, holds significant importance for the preservation and study of cultural heritage through its image classification. However, traditional image classification methods often struggle to comprehensive capture the intricate textures, distinctive color schemes and dense religious symbols inherent in Thangka artworks. This makes It difficult to achieve ideal classification results. To address these challenges, this paper proposes a Thangka image classification method based on Vision Transformer(Vit). Different from traditional Convolutional Neural Network (CNN) and standard ViT, the proposed model, MSA-ViT, introduces a multi-scale feature fusion mechanism to effectively integrate feature information at different scales in Thangka images, enabling the model to balance local details and global semantics. Furthermore, we incorporate channel-spatial attention mechanisms to allow the model to focus on more discriminative feature regions and enhance feature representation capacity. Experimental evaluations conducted on a self-constructed Thangka dataset demonstrate that the proposed method achieves classification accuracy improvements of 6.2 and 8.1 percentage points over traditional CNN and standard Vit, respectively, validating its efficacy in Thangka image classification tasks.