Style transfer of Dunhuang murals fusing multi-scale features

Abstract:

To address color distortion, detail blurring, and structural incoherence in existing style transfer techniques when processing Dunhuang murals——caused by highly saturated mineral pigments, intricate textures, and complex layered structures——this paper proposes a Multi-scale Dunhuang Style Transfer Network based on an improved Cycle-Consistent Generative Adversarial Network for high-quality artistic style transfer. We introduce an adaptive local dilated convolutional-net that dynamically captures detailed texture edges using deformable convolution and enhances long-range texture dependencies through dilated convolution, thereby restoring deep features to preserve brushstroke details. A dual scope net is designed to mitigate information loss and color-layer weakening during style transfer, employing a global attention branch to model overall tonal harmony and a local grouped convolution branch to reinforce stroke details. Additionally, a pathwise fusion net optimizes logical relationships and proportional coordination between elements using multi-dilation-rate depthwise separable convolutions for parallel processing and a dynamic gated fusion mechanism. Experimental results show that the proposed method achieves reductions of 5.81%, 4.36%, and 5.73% in FID, LPIPS, and L2 loss, respectively, and an improvement of 8.12% in SSIM. User studies confirm its superiority in content fidelity, style consistency, and visual appeal. This approach effectively resolves challenges in preserving color layers, texture details, and spatial layouts in Dunhuang murals transfer, offering a novel approach for Dunhuang murals digitization and innovative dissemination.