Spatial-frequency feature fusion network for Thangka image super-resolution

Rich religious and cultural meanings are associated with Thangka, the Tibetan sacred painting art form’s intangible cultural heritage (ICH). Thangka’s digital preservation is essential to the historical background and legacy of Tibetan Buddhist art. However, because of age degradation and preservation constraints, high-frequency information in Thangka images—like elaborate clothing patterns, ritual details, and backdrop decorations—is in danger of disappearing. The non-local texture dependencies in Thangka images are difficult for conventional reconstruction algorithms based on spatial domain characteristics, which depend on local convolution procedures, to adequately capture. As a result, this paper suggests a spatial-frequency feature fusion super-resolution (SFSR) algorithm for image super-resolution reconstruction that is based on the fusing of spatial-frequency features. This approach builds a cooperative network architecture in both the global frequency domain and the local spatial domain. A fast Fourier convolution module is used to efficiently model long-distance dependencies in images; a second-order channel attention module is introduced, which uses the second-order covariance statistics between channels to more precisely construct the nonlinear correlations between model feature channels; and a triangular window self-attention module is adopted, whose various shift mechanisms effectively model the image’s complex geometric structures and edges in various directions while minimizing artifacts. Experimental results on the Tangka dataset show that the proposed algorithm can restore texture details with high fidelity, with a PSNR value improved by 0.07 ∼ 0.37dB compared to the classic SwinIR algorithm. This achieves high-quality reconstruction of Tangka images and provides effective technical support for Tangka digital preservation.