基于生成对抗网络的彩色图像去马赛克修复研究

Research on color image de-mosaic restoration based on generative adversarial networks

针对传统生成对抗网络在图像去马赛克时存在梯度消失和网络结构欠缺对图像高阶特征学习的问题,提出一种基于生成对抗网络的彩色图像去马赛克改进算法,并确定去马赛克修复阈值.该方法借鉴Pix2Pix算法和SRGAN算法的结构特点,在Pix2Pix算法基础上加入VGG19内容损失计算进行改进;通过在数据集COCO(2014版)上批量添加10个规格的全局马赛克,然后以清晰图像和马赛克图像成对的方式输入网络进行训练.实验结果表明,改进算法Pix2Pix-VGG19在阈值9×9内,较Pix2Pix算法的PSNR值平均提高了0.3545,较SRGAN算法的PSNR值平均提高了1.5707,提升了彩色图像去马赛克的修复效果.

This work propose an improved de-mosaic algorithm to address the issues of gradient disappearance and network structure lack of learning high-order features of images in Generative Adversarial Networks.The idea is to borrow Pix2Pix and SRGAN algorithms and introduce the VGG19 content loss function for improvement.10 global mosaics are added into COCO’14 data set.Training is implemented by feeding the network by clear image mosaic image pair.Experimental results show that PSNR is increased by 0.3545 for the improved Pix2Pix-VGG19 algorithm compared to Pix2Pix,and is increased by 1.5707 compared to SRGAN,within a 9×9 threshold.