基于RDE-GAN算法的多阶段壁画图像修复

Multi-stage Mural Image Restoration Based on RDE-GAN Algorithm

针对现存的壁画图像修复方法仍存在纹理细节缺失及图像输入像素尺寸不合适的问题,提出了一种多阶段密集残差和高效注意力机制的壁画图像修复(multi stage mural image restoration based on residual dense efficient-generative adversarial networks,RDE-GAN)算法。整个网络采用编码器-解码器架构,使网络具有足够大的感受野,便于充分利用图像的特征信息。首先,利用全局感知网络得到粗略的初始结果;其次,引入较小感受野的密集残差局部过渡网络;最后,利用高效细化网络增强图像的结构信息及图像语义的连贯性。将该算法分别与其他相关算法在定性定量分析上进行比较,结果表明,在[50%,60%)掩码比例中,RDE-GAN算法的峰值信噪比(peak signal-to-noise ratio,PSNR)为32.5655 dB、结构相似性指数(structural similarity index measure,SSIM)为0.9690、学习感知图像块相似度(learned perceptual image patch similarity,LPIPS)为0.0141、生成图像与真实图像越相似指标(Fréchet inception distance,FID)为11.3027,且在其他5种掩码比例中RDE-GAN算法均优于对比算法。该研究成果能用于壁画等文化遗产的保护。

Multi-stage mural image restoration based on RDE-GAN algorithm(RDE-GAN)was proposed to address the limitations of existing methods for repairing mural images,such as missing texture details and unsuitable input pixel sizes.The entire network adopted an encoder-decoder architecture to ensure a sufficiently large receptive field for the utilization of image feature information.Firstly,a global perception network was utilized to obtain rough initial results.Secondly,a dense residual local transition network with a small receptive field was introduced.Finally,an efficient refinement network was employed to enhance the structural information of the image and the coherence of image semantics.The proposed algorithm was compared qualitatively and quantitatively with other relevant algorithms.The experimental results showed that,at[50%,60%)mask ratio,the peak signal-to-noise ratio(PSNR)of the RDE-GAN algorithm was 32.5655 dB,the structural similarity index measure(SSIM)was 0.9690,the learned perceptual image patch similarity(LPIPS)was 0.0141,and the Fréchet inception distance(FID)between generated and real images was 11.3027.Additionally,the RDE-GAN algorithm outperformed the compared algorithms in other five mask ratios.This research results can be used for the protection of cultural heritage such as murals.