Application of GANs in Ancient Architectural Heritage Image Restoration
摘要
Han-Tang Western Region beacon towers hold significant historical value but suffer severe damage, with traditional restoration lacking visual references. This paper proposes a GAN-based image restoration method combining Pix2Pix and CycleGAN models. Experiments show Pix2Pix achieves higher accuracy but requires paired samples. We design a hybrid workflow: CycleGAN first generates restoration targets, then concatenates with original images for Pix2Pix input. AI-generated virtual beacon images enhance database diversity, significantly improving restoration stability and accuracy. This research innovatively applies deep learning to architectural heritage preservation, providing a feasible method for building image restoration with limited data. It efficiently supplies visual references for traditional restoration and offers a scalable technical pathway for beacon tower virtual reconstruction, providing valuable guidance for preservation work.