.基于改进CycleGAN的视频监控人脸超分辨率恢复算法

针对有监督超分辨率算法训练过程需要大量成对图像、处理真实低分辨率图像视觉恢复效果差等问题,提出了一种基于改进CycleGAN的半监督算法Cycle-SRNet.首先,利用退化模型获得与真实低分辨率人脸相似的图像,用于训练网络参数;其次,通过重建模型恢复出具有真实效果的高分辨率人脸图像;最后引入感知损失函数保持人脸结构相似性,以更好地恢复面部特征.实验结果表明,该算法不需要成对的图像进行网络训练,在视觉效果上能够将模糊的视频监控低分辨率人脸图像恢复成清晰可辨的人脸图像,在FID、PSNR和SSIM指标上超越了SRCNN、SRGAN、CinCGAN等方法.

Improved video surveillance face super-resolution recovery algorithm based on CycleGAN

Traditional super-resolution algorithms usually require paired-image for networks training, which have poor visual recovery effects when processing real-world low-resolution images such as video surveillance faces. In this paper, an improved semi-supervised algorithm named Cycle-SRNet based on CycleGAN is proposed. Cycle-SRNet is consist of two models, the degradation model generate images similar to real-world low-resolution faces for network training, while the reconstruction model restore high resolution face image. The perceptual loss function is introduced to ensure the structural similarity of the face so as to recover the facial features better. Experimental results show that the algorithm can restore fuzzy low-resolution video surveillance face images into recognizable face images, and it exceeds SRCNN, SRGAN and other supervised methods in terms of FID and PSNR indicators.