基于增强型超分辨率生成对抗网络的文物图像色彩重建

Color Reconstruction of Cultural Relics Image Based on Enhanced Super-Resolution Generative Adversarial Network

针对书画文物的褪色和画面暗旧等问题,提出了一种基于增强型超分辨率生成对抗网络的文物图像色彩重建(Color Reconstruction of Cultural Relic Images Based on Enhanced Super-Resolution Generative Adversarial Network, CR-ESRGAN)模型。该模型针对缺少成对图像的数据集问题,在双3次下采样的基础上提出了利用颜色迁移算法来生成逼真的暗旧、褪色的文物图像。同时改进了ESRGAN网络,在其生成网络中引入自注意力机制,以增强重建图像的纹理细节。在常用图像质量评价指标峰值信噪比(Peak Signal to Noise Ratio, PSNR)/结构相似性(Structural Sililarity Index, SSIM)的基础上引入颜色评价指标CIEDE2000,以更加全面、客观地评价重建图像的质量。与现有几种超分辨率算法以及其文物图像色彩修复方法相比,视觉效果和图像质量有较高的提升。

For the problems of fading of painting and calligraphy cultural relics and dark old pictures, a Color Reconstruction of Cultural Relics Images Based on Enhanced Super-Resolution Generative Adversarial Network(CR-ESRGAN) model fis proposed. To solve the problem of datasets lacking paired images, a color transfer algorithm is used by the model to generate realistic dark, faded and faded cultural relic images on the basis of bicubic downsampling. The ESRGAN network is improved, and self-attention mechanism is introduced into its generative network to enhance the texture details of the reconstructed images.Based on the commonly used image quality evaluation index Peak Signal to Noise Ratio(PSNR)/Structural Similarity Index(SSIM), the color evaluation index CIEDE2000 is introduced to evaluate the quality of the reconstructed image more comprehensively and objectively. Compared with several existing super-resolution algorithms and its cultural relic image color restoration method, the visual effect and image quality have been improved by the proposed method.