CA‐CST: A Contrast Learning‐Based Artistic Style Transfer Network With Regional Self‐attention

ABSTRACT This paper proposes a culture‐aware contrastive style transfer (CA‐CST) framework to address the limitations of traditional neural style transfer in preserving culturally significant regions. Unlike conventional methods that treat all regions uniformly, CA‐CST explicitly models culturally salient areas to ensure faithful and controllable stylisation. CA‐CST integrates explicit cultural‐region masks, adaptive attention‐based modulation and contrastive learning objectives to achieve controllable and high‐fidelity stylisation. By coupling a multi‐scale cultural mask pyramid with multi‐resolution optimisation, the framework enhances cultural elements while maintaining global aesthetics. Specifically, the multi‐scale masking strategy progressively guides feature refinement, enabling precise enhancement of culturally important structures without compromising overall visual coherence. Experimental results show that CA‐CST achieves superior visual consistency and cultural fidelity, with competitive performance on PSNR (25.730), SSIM (0.658), LPIPS (0.451) and Gram Loss (0.720). Comparisons with existing approaches and ablation studies confirm its effectiveness in preserving structural integrity and highlighting cultural regions without disrupting style harmony. Overall, CA‐CST provides a robust solution for cultural heritage image stylisation, offering improved accuracy, expressiveness and controllability. The proposed framework not only advances culture‐aware style transfer research but also holds strong potential for applications in digital heritage preservation and cross‐cultural communication.