利用改进的Real-ESRGAN模型进行历史图像超分辨率重建研究  

Research on Historical Image Super-Resolution Reconstruction Using Improved Real-ESRGAN Model

作　　者：倪劼[1] 柳青远 周莉[1] Ni Jie Liu;Qingyuan;Zhou Li(Nanjing Library)

机构地区：[1]南京图书馆

出　　处：《信息与管理研究》2025年第1期65-77,共13页Journal of Information and Management

摘　　要：图书馆馆藏近代图像由于保存时间较长、拍摄条件限制等因素,往往存在分辨率低、细节模糊、对比度不均和泛黄等问题。通过分析当前图像超分辨率技术现状,提出一种改进的Real-ESRGAN模型。对馆藏近代图像进行分析,采用数据增强技术模拟图像特征,在模型的RRDB模块中引入多尺度特征融合,并结合通道注意力机制,以提高原始模型的性能。实验结果表明,改进后的模型在图像重建的整体视觉和细节部分,均比原始模型有所改善,在PSNR和SSIM指标上,相较于原始模型分别提升3db和0.0672,非常适合图书馆开展馆藏近代图像超分辨率重建任务。Images of Modern History from the library's collection often suffer from issues such as low resolution,blurred details,uneven contrast,and yellowing due to long-term preservation and limitations of early photographic techniques.By analyzing the current state of image super-resolution technology,this study proposes an improved Real-ESRGAN model.Through detailed analysis of images of Modern History from the library's collection,data augmentation techniques are employed to simulate image characteristics.Additionally,multi-scale feature fusion is integrated into the RRDB module of the model,combined with a channel attention mechanism to enhance the performance of the original model.Experimental results demonstrate that the improved model significantly outperforms the original model in both overall visual quality and detail reconstruction.Specifically,the PSNR and SSIM metrics of the improved model increase by 3 dB and 0.0672,respectively,compared to the original model,making it highly suitable for super-resolution reconstruction tasks of historical library images.

关 键 词：Real-ESRGAN 馆藏数字化 图像提升 图像超分辨率 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]