摘要
针对真实场景下高低分辨率图像对难以获取和真实低分辨率图像的退化过程复杂未知的问题,提出一种基于生成对抗网络和元学习的刑侦图像超分辨率算法。利用非成对训练样本训练生成对抗网络得到伪成对图像,使超分辨率网络以监督的方式学习伪成对训练样本之间的映射。从大量外部训练图像获取先验信息,并利用图像内部的重复相似性对网络进行训练,凭借元学习策略使网络快速收敛。在标准数据集和真实刑侦图像上的实验结果表明,所提算法能恢复出更多的纹理信息,适应于真实刑侦场景的需要。
A super-resolution algorithm for criminal investigation images based on generative adversarial networks and meta-learning is proposed for the problems of difficult access to high-and-low resolution image pairs in real scenes and the complex unknown degradation process of real low-resolution images.The proposed method utilizes unpaired training samples to train the generative adversarial network(GAN)to obtain pseudo-paired images,and then a mapping between pseudo-paired training samples is learned by the super resolution(SR)network in a supervised manner.Obtain prior information from extensive external training datasets,and fine-tune the network by the internal repetition of information within the images.The network converges quickly with the meta-learning strategy.Experimental results on standard datasets and real video criminal investigation images show that the proposed method can largely recover texture information and is adapted to the real criminal investigation scenarios.
作者
徐健
李新婷
邓聪
牛丽娇
XU Jian;LI Xinting;DENG Cong;NIU Lijiao(School of Communications and Information Engineering,Xi'an University of Posts and Telecommunications,Xi'an 710121,China;Key Laboratory of Electronic Information Application Technology for Scene Investigation Ministry of Public Security,Xi'an University of Posts and Telecommunications,Xi'an 710121,China)
出处
《西安邮电大学学报》
2022年第2期62-71,共10页
Journal of Xi’an University of Posts and Telecommunications
基金
国家自然科学基金项目(61601362,62071380,41874173)
陕西省社会科学基金项目(2021ZX15)
西安邮电大学研究生创新基金项目(CXJJLZ202001)。
关键词
图像超分辨率
生成对抗网络
元学习
刑侦图像
image super-resolution
generative adversarial network
meta-learning
criminal investigation image
作者简介
徐健(1981-),女,博士,教授,从事图像清晰化及识别算法研究。E-mail:Xujian_paper@126.com;李新婷(1996-),女,硕士研究生,研究方向为图像处理。E-mail:836348926@qq.com。
