Alibaba DAMO Academy Wins at the First “Malanshan Cup” International Competition for Audio and Video Algorithm Optimization
Catch the replay of the Apsara Conference 2020 at this link!
By DAMO Academy
The awards ceremony and summit forum of the first “Malanshan Cup” International Competition for Audio and Video Algorithm Optimization were hosted by the Hunan Internet Information Office and the Hunan Science and Technology Association. It was held on September 8, 2020. The event was conducted by the China Society for Industrial and Applied Mathematics and the China Federation of Internet Societies. It was sponsored by the China (Changsha) Malanshan Video Cultural and Creative Industry Park and Mangotv. The event gathered excellent algorithm elites from all over the world. A large number of talents from universities, research institutes, and Internet enterprises registered for the competition. Among the 1,294 competition teams, 34 were from Peking University, 25 were from Tsinghua University, 37 were from MIT, and other top universities.
At the forum, Ren Peiran, a Senior Algorithm Expert from Alibaba DAMO Academy, was invited to give a presentation about “Intelligent Vision Editing and Enhancement”. During the report, Ren Peiran introduced DAMO Academy’s technology development and product solutions in intelligent video embedding and enhancement for ultra-high definition videos.
After about a month, the ManGoGoGo team, formed by Zeng Hui, Yang Xi, and Xiang Fei from Alibaba DAMO Academy, ranked first in the preliminary and rematch. They won a title with a score of 88.503 for their excellent performance in recovering image quality.
The solution designed by the ManGoGoGo team involved many clever algorithm optimizations for problems, such as noise and compression damages generated during the video collection and post-production stages. For example, the team employed the inter-frame feature of video images to restore video details, solved the problem caused by the imbalance between high resolution and limited video memory through model designs, and combined different advantages of several complementary algorithm models to restore a low-quality video to a high-definition video. With multiple composite optimization solutions, they achieved the best recovery results and won first place, providing innovative ideas that can be used for reference for algorithm model optimization. The winning solution was optimized in two aspects. The first used two patches of 512 x 1024 and 512 x 512 for training with enough input image to make full use of the complementary inter-frame information. The second used the complementary design in model structures, with one focusing on the attention mechanism, and another to obtain more receptive field information in its deeper layer. Through different designs, more complementary integration effects can be obtained, which is the experience accumulated from previous competitions.
Members of ManGoGoGo are from the artificial intelligence center of the DAMO Academy. Zeng Hui, the team leader, published several papers in the computer vision and image processing fields. These papers were included by top conferences, such as CVPR and ICCV, and top journals, such as TPAMI and TIP. Under his leadership, the ManGoGoGo team won the title of the AI + 4K HDR track at the first National Artificial Intelligence Competition.
ManGoGoGo has won several titles in video quality enhancement and restoration competitions recently. Its technology foundation stems from Alibaba’s continuous R&D investments in this field. In recent years, China has been promoting the development strategy for 4K ultra high definition. The State Administration of Radio and Television just issued the “Implementation Guide for 4K Ultra High Definition TV Program Production Technology (2020)” in June. We can foresee an increase in the demand for image quality enhancement in the fields of radio, television, and pan-entertainment videos. In addition, image quality restoration is often used to restore historical images. It is not only helpful for the inheritance of film and television culture, but it is also a typical scenario of the media industry empowered by visual AI technology, which has great commercial potential.
Based on DAMO Academy’s cutting-edge AI algorithm capabilities, such as super-resolution imaging, image de-noising, frame insertion, HDR color enhancement, subtitle rebirth, and scratch removal, the solution includes many typical video enhancement scenarios, such as 4K Ultra HD enhancement, general-purpose image quality enhancement, film repairing, and the upcoming live video enhancement.
Learn more about the latest updates on research and innovation from Alibaba DAMO Academy by catching the replay of the Apsara Conference 2020 at this link!