Reading Chuang/Shenzhen Commercial Daily reporter Yuan Siru
On September 14, the "Independent Innovation Lecture Hall" was held in Shenzhen. The theme of this issue was "Image Restoration and Enhancement Based on Deep Learning". Business representatives and experts gathered Gather to discuss the latest developments in the industry.
It is reported that the event is hosted by Shenzhen Science and Technology Association and Advanced Technology Research Institute of Chinese Academy of Sciences, and undertaken by Shenzhen Robotics Association and Shenzhen New Generation Information and Communication Industry Cluster.
Computer Vision is the frontier and focus of the field of artificial intelligence. Since the creation of deep learning algorithms, various tasks of computer vision have achieved unprecedented breakthroughs. Image restoration and enhancement are classic underlying computer vision problems, which are widely used in 4k/8k video, AI photography, ultra-clear image quality, AR/VR and other fields.
This report takes image restoration and enhancement tasks as the starting point, reviews the development history of deep learning algorithms, explains the advantages and bottlenecks of the algorithms, and displays the latest research results. The theme of this issue of the salon is "Image Restoration and Enhancement Based on Deep Learning". The event invited Associate Researcher Dong Chao from the Institute of Digital Technology, Shenzhen Advanced Institute of the Chinese Academy of Sciences to give a report entitled "Development and Prospects of Image Super-resolution Technology". At the same time, Dr. Wang Xintao, a senior researcher at Shenzhen Tencent PCG, and Dr. Jiawei Zhang, a senior researcher at SenseTime at , were invited to give presentations and exchanges on the degradation models and practical applications of image restoration.
At the meeting, Dong Chao, associate researcher of Shenzhen Advanced Institute of Chinese Academy of Sciences, shared the development and prospects of image super-resolution technology. He said that since 2014, super-resolution technology can be divided into six development lines, including super-division network, visual super-division, real scene, adjustable restoration, interpretability, and super-division competition.
Zhang Jiawei, a researcher at Shangtang Technology, said that as deep neural networks are used more and more widely in computer vision, more and more deep neural network structures are also used to solve the problem of degraded image restoration.But they often ignore the degradation model. "The combination of deep neural networks and traditional degraded image restoration algorithms can better restore images," he said.
Tencent researcher Wang Xintao shared that in recent years, image restoration, especially image super-resolution technology has developed rapidly, various indicators have been continuously improved, and visual effects have gradually improved. However, due to the complex and diverse degradation processes in the real world, there is still a large gap between them and practical applications. "From the perspective of practical applications, we tried to do some exploration work on image restoration and enhancement. The first is face restoration GFPGAN, which uses the priori of face generation against the network to assist in the restoration and detail generation of actual low-quality faces. It can effectively solve most face scenes. The other is the general image restoration Real-ESRGAN, which explores the use of pure synthesis to solve actual image restoration problems." He said.
Review: Tan Lugang
.
