There is a big event recently. On November 15, the 4K restored version of the classic film The Sea Pianist, which attracted worldwide attention, will hit cinemas across the country. The classic scene of the film eroded by time is only blurred by the preservation of the old film. And the version repaired by 4K technology, it is said to use advanced image super resolution technology, resolution and effect greatly improved, wear video image restore to the film when the real effect, details show incisively and vividly, light and shadow clear, delicate, pleasing to the eye. This image super resolution technology black technology is what is going on, and how to achieve the video super points with artificial intelligence?
What is image super resolution technology?
Super-resolution is a technology that improves the Resolution of the original image by means of hardware or software. Super-resolution reconstruction is the process of obtaining a high Resolution image through a series of low Resolution images. The core idea of super-resolution reconstruction is to exchange temporal bandwidth (obtaining multi-frame image sequence of the same scene) for spatial resolution, and realize the conversion from temporal resolution to spatial resolution. In layman’s terms, super resolution is the transformation of an image from the image on the left to the image on the right. The idea is to provide more pixels to fill out the details of the image. From the point of view of the signal, that is to add more high-frequency components in the original signal.
Research status of super resolution technology
The super-resolution restoration techniques of images can be divided into two categories: one is based on reconstruction and the other is based on learning. In recent years, with the rise of deep learning, image super-resolution technology has emerged a new research direction. SR(Super-resolution) based on deep learning is mainly based on single-sheet low-resolution reconstruction method, namely SISR(Single Image super-resolution).
SISR is an inverse problem. For a low-resolution image, the high-resolution image obtained by different methods will be slightly different. That is, there may be many different high-resolution images corresponding to it. Therefore, it is usually necessary to add a prior information to normalize constraints when solving high-resolution images. In the traditional approach, this prior information can be learned from several instances of low-resolution images in pairs. SR based on deep learning directly learns the end-to-end mapping function from high-resolution image to high-resolution image through neural network.
At present, there are many SR methods based on deep learning. From SRCNN in 2014 to WDSR, the champion of VDSR and NTIRE2018, super-resolution technology based on deep learning keeps innovating network architecture, loss function and learning strategy. Continuous breakthroughs have been made in feature extraction, nonlinear mapping and reconstruction architecture.
In order to obtain high-quality high-resolution images and meet various needs of people in actual production and life, research and development directions of super-resolution image restoration are mainly focused on the following three aspects:
- Improve existing algorithms, and constantly develop new algorithms. The purpose of this method is to improve the ability of super-resolution image restoration, reduce the amount of calculation, speed up the convergence of operation, suitable for different image requirements;
- To develop and seek new degradation imaging models to make them more accurate and comprehensive, and to achieve accurate estimation of point spread function and noise;
- In the restoration of sequences and multiple images, new motion models can be developed and sought for accurate estimation of motion. See here, I believe you have a certain understanding of this new AI technology. Want to learn more about the technology and get some practical experience to prove yourself? This year, there were two AI competitions related to image super-resolution technology. One was the Youku Super Score competition, which ended in September. Recently, the first “National Artificial Intelligence Competition” was launched, in which the AI+4K HDR competition is related to super resolution technology. It can be said that it is a rare opportunity for students to challenge themselves and prove themselves! Using some of the existing AI algorithm models, and looking at previous matches, you can quickly get started.
National Artificial Intelligence Competition
The National Artificial Intelligence Competition (hereinafter referred to as the Competition) was established by shenzhen Municipal People’s Government in August 2019. Based on an international perspective, the competition will create an atmosphere for artificial intelligence innovation and creation, and promote the integrated development of innovation elements such as industry, academia, capital and talent. The competition is sponsored by the People’s Government of Shenzhen Municipality, jointly undertaken by AITISA (hereinafter referred to as the “Alliance”), a new generation of ARTIFICIAL Intelligence Industry Technology Innovation Strategic Alliance established under the guidance of Shenzhen Science and Technology Innovation Commission, Pengcheng Laboratory and the Ministry of Science and Technology, and co-organized by Tencent Technology and others.
The competition includes preliminary, semi-final and final. The competition tasks are as follows:
Task of the preliminary contest: video quadruple overscore, requiring reconstruction of 540p SDR video with random noise into 4K SDR video after denoising;
Task of the second round: video quad hyperscore + image quality enhancement, it is required to rebuild the 540P SDR video with overexposure/underexposure content and untoned color into 4K SDR video with high quality and toned color;
Final task: Video quad hyperscore +HDR, asking to rebuild a low-quality 540P SDR video into a high-quality, colored 4K HDR video.
Competition schedule
Competition time: 2019-10-17 to 2020-1-15
2019-10-17 (12:00:00 AM) to 2019-11-29 (12:00:00 AM)
• Login to the official website of the competition on PC, login with wechat or Github account, and complete personal information registration and registration
• The preliminary and semi-final competitions are online, and the venue is not limited
2019-10-28 (12:00:00 AM) to 2019-11-30 (12:00:00 AM)
2019-12-2 (12:00:01 AM) to 2019-12-6(12:00:01 AM)
Each team can download data (including training set and test set, etc.) on the competition platform, conduct algorithm design, model training and debugging locally, and submit evaluation results online to participate in ranking.
The top 150 teams submit their results for manual review, and the top 100 teams that pass the review advance to the semi-finals.
2019-12-9 (12:00:00 AM) to 2020-1-3 (12:00:00 AM)
Review Stage: 2020-1-6 (12:00:00 AM) to 2020-1-10 (12:00:00 AM)
The winners of the preliminary rounds advance to the semi-finals. The training set and test set will be open during the semi-final, and each contestant will submit the results for scoring, ranking, code verification, etc.
Finals: 2020-1-11 to 2020-1-15
The winners of the second round will advance to the final, which will be held in Shenzhen. The final will be conducted by on-site technical competition and defense. Detailed data and related arrangements will be announced after the second round. Want to challenge yourself, improve your skills and win a big bonus? Open the contest’s official website and compete with the great gods from all over the country.
The game’s official website: www.kesci.com/home/compet…