Thangka Style Transfer Using Adaptive Focus Convolution with Channel and Positional Attention

Thangka, as an important symbol of Tibetan culture, has traditionally been created through meticulous hand-painting, showcasing its rich cultural and historical significance. With the advancement of digital technology, the preservation and dissemination of Thangka art face new opportunities, particularly in promoting its role in global cultural exchange. Digital preservation not only protects this traditional art form from the ravages of time and environmental factors but also enables its dissemination across geographical and temporal boundaries. This study explores the use of style transfer techniques for the digital recreation of Thangka murals, aiming to achieve a fusion of traditional art forms and modern artistic styles. To address challenges in style transfer, such as maintaining semantic details and ensuring consistency in stylized colors, we propose an Adaptive Style Fusion Network (ASFNet). ASFNet employs Adaptive Focus Convolution (AFConv) to dynamically adjust convolution kernel weights, optimizing the processing of complex textures. Additionally, it integrates a Coordinate Encoding Feature Channel and Positional Attention (CPA) to enhance the capture of style image features and ensure color consistency. Experimental results on the Thangka image dataset demonstrate that our method effectively resolves issues such as structural distortion and color inconsistency found in traditional methods. ASFNet achieves high-quality style transfer while preserving the cultural and religious characteristics of Thangka art. This study not only provides a new technological approach for the digital preservation and dissemination of Thangka art but also offers valuable insights for computer-assisted artistic creation. Furthermore, both qualitative and quantitative experiments confirm that our method delivers aesthetically pleasing stylization results on Thangka images.