ResNet transfer learning and CycleGAN style transfer based historical scene image restoration and enhancement system
Historical photographs represent invaluable cultural heritage, yet they commonly suffer from multiple degradations including noise, scratches, color fading, and physical damage due to aging and improper storage. While recent deep learning-based restoration methods have demonstrated impressive technical performance, they tend to over-enhance historical images, transforming them into modern-looking photographs and consequently losing their characteristic historical aesthetic and authenticity. This paper presents a novel dual-pathway restoration framework that effectively removes various degradations while explicitly preserving the original historical visual style. Our approach employs a VGG-19-based style extraction branch parallel to the main restoration pathway, utilizing adaptive instance normalization (AdaIN) layers to inject historical style features into the decoder. We introduce a comprehensive loss function combining reconstruction loss, perceptual loss, and a novel style preservation loss computed through Gram matrix distance, which distinguishes desirable historical characteristics from undesirable degradations. A two-stage progressive training strategy ensures stable convergence and optimal performance. Extensive experiments on 10,000 historical photograph pairs spanning 1920s-1980s demonstrate that our method achieves superior quantitative results (PSNR: 28.73 dB, SSIM: 0.891) compared to state-of-the-art approaches including DnCNN, ESRGAN, and Bringing Old Photos Back to Life, while maintaining authentic historical visual characteristics suitable for digital heritage reconstruction applications.
Paper
Full text
ResNet transfer learning and CycleGAN style transfer based historical scene image restoration and enhancement system
Semantic Scholar · Computer Science · 2026
Abstract
Historical photographs represent invaluable cultural heritage, yet they commonly suffer from multiple degradations including noise, scratches, color fading, and physical damage due to aging and improper storage. While recent deep learning-based restoration methods have demonstrated impressive technical performance, they tend to over-enhance historical images, transforming them into modern-looking photographs and consequently losing their characteristic historical aesthetic and authenticity. This paper presents a novel dual-pathway restoration framework that effectively removes various degradations while explicitly preserving the original historical visual style. Our approach employs a VGG-19-based style extraction branch parallel to the main restoration pathway, utilizing adaptive instance normalization (AdaIN) layers to inject historical style features into the decoder. We introduce a comprehensive loss function combining reconstruction loss, perceptual loss, and a novel style preservation loss computed through Gram matrix distance, which distinguishes desirable historical characteristics from undesirable degradations. A two-stage progressive training strategy ensures stable convergence and optimal performance. Extensive experiments on 10,000 historical photograph pairs spanning 1920s-1980s demonstrate that our method achieves superior quantitative results (PSNR: 28.73 dB, SSIM: 0.891) compared to state-of-the-art approaches including DnCNN, ESRGAN, and Bringing Old Photos Back to Life, while maintaining authentic historical visual characteristics suitable for digital heritage reconstruction applications.