DCAF-GAN: Enhancing historical landscape restoration with dual-branch feature extraction and attention fusion

Abstract

Historical landscape restoration has become a crucial area of research in cultural heritage preservation, and with the advancement of digital technologies, effectively restoring damaged historical images has become a critical challenge. Traditional restoration methods face difficulties in handling large occlusions, complex structural features, and maintaining high fidelity in restored images. Existing deep learning methods often focus on restoring a single feature, making it difficult to achieve high-quality reconstruction of both texture and structure. To address these challenges, we propose DCAF-GAN, a novel deep learning model that effectively restores both fine textures and global structures in damaged historical landscapes through a dual-branch encoder and a channel attention-guided fusion module. Experimental results show that DCAF-GAN achieves a PSNR of 29.12 and SSIM of 0.867 on the StreetView dataset, and a PSNR of 28.6 and SSIM of 0.854 on the Places2 dataset, significantly outperforming other models. These results demonstrate that DCAF-GAN not only provides high-quality restorations but also maintains computational efficiency. DCAF-GAN offers a promising solution for the digital preservation and restoration of cultural heritage, with significant potential for further applications.