Cui Qinghua's team proposed a new algorithm for universal image embedding and won the first prize in the Google Universal Image Embedding Research Competition

Computer vision is a science that allows computers to understand "information seen", is a hot field of artificial intelligence, and has important applications
in various visual tasks, including intelligent analysis of medical images.
In recent years, instance-level recognition (ILR) has become a new research hotspot
in the field of computer vision.
The so-called "instance-level recognition" is to identify a specific instance of an object rather than simply identify the computer vision task to which it belongs in the past, which has great application prospects in the field of extensive image search and matching (such as Baidu image recognition, Google image recognition), and this kind of computer vision algorithm also has potential application value
in the field of medical image intelligent analysis.
Among them, general image embedding calculation is the basis of instance-level recognition, and its performance directly affects the recognition effect
.

Recently, Professor Cui Qinghua and doctoral student Shihao Shao of the Department of Medical Bioinformatics, School of Basic Medicine of Peking University, proposed a new algorithm in the field of general image embedding calculation methods, and they proposed CLIP pre-training based on Laion-2B The VIT-H scheme establishes a new training and fine-tuning method, analyzes the model fusion scheme and feasibility under the situation of feature vector space disunity, and the algorithm was selected in the just-concluded Google Universal Image Embedding Competition ) to win the championship
.
This competition is the designated competition of the European Conference on Computer Vision (ECCV), one of the top conferences in the field of computer vision, and is an annual series of instance-level recognition competitions, attracting from the United States, Germany, France, Japan, Australia and other countries 1022 teams participated
.

The main page of this contest

In this competition, the new algorithm proposed by Professor Cui Qinghua and Shao Shihao finally won the first place in the public data and blind test data rankings with scores of 0.
732 and 0.
728, and the details of the method were announced here style="line-height: 150%;font-size: 16px" _istranslated="1">
。 In view of these outstanding results, the two were also invited to attend the European Computer Vision Conference to be held in Tel Aviv, Israel, and at the "Instance-Level Recognition Workshop.
" ECCV)" to give live oral presentations
.
The proposed algorithm has significantly improved the task of instance-level image target recognition and is expected to be applied to the intelligent analysis
of medical images.

Under the guidance of Professor Cui Qinghua, Shao Shihao, an eight-year doctoral student in basic medicine at Peking University in 2017, is committed to proposing new artificial intelligence algorithms to solve problems in the field of computer vision and applying them to medical image intelligent analysis and processing Challenge winner) and gastrointestinal tumor image segmentation (University of Wisconsin UW-Madison GI Tract Image Segmentation Competition Gold Medal) have been successfully applied
.
This work has been supported
by the National Science Foundation for Outstanding Young Scholars, the Innovation Group of the National Natural Science Foundation of China and other funds.

Screenshot of the ranking page of this competition

(School of Basic Medicine, Peking University)