On February 8th, the results of the Earth AI (Earth Observation Artificial Intelligence Technology) competition at The Dubai World Expo were announced live online, and the Holiday Coder (Holiday Programmer) team of Fudan University won the championship!
The members of the Holiday Coder team are all from the team of academicians Jin Yaqiu of the Key Laboratory of Electromagnetic Wave Information Science of the Ministry of Education of Fudan University, namely Pu Xinyang and Song Jingduo, 2021 master's students, and Jia Hecheng, a 2019 doctoral student.
Fudan University Holiday Coder team: Pu Xinyang, Song Jingduo, Jia Hecheng
The Earth AI Competition at The Dubai Expo was attended by 9 university teams from China, Italy, the United Arab Emirates and other countries. The competition was held from January 28 to February 4 for a week. This is also the first artificial intelligence technology competition held at the World Expo.
New Year's Events,
Fudan students fought hard for 7 days to win the championship of the international competition of artificial intelligence
Chinese New Year's Eve night, families sit around the table, eat Chinese New Year's Eve meals, watch the Spring Festival Gala, and enjoy the fun. At eleven o'clock, Three students from Fudan Holiday Coder team, Pu Xinyang, Song Jingduo and Jia Hecheng, sat in front of the computer and held an online meeting exclusive to them.
The Earth AI competition at the Dubai World Expo coincided with the Chinese New Year, and the time was tight, and the team members could only celebrate the festival, visit relatives and friends, and conduct online meetings and voice exchanges after 10 p.m., often discussing until 1 or 2 a.m. They connected Xi'an in Shaanxi, Fuxin in Liaoning and Dongyang in Zhejiang to "brainstorm" for a common goal.
Captain Pu Xinyang said the reason why he still participated during the Spring Festival: "Because such a world-class competition can be encountered, and its remote sensing image recognition competition task based on artificial intelligence technology is very closely related to our usual research topics, and there is more experience and competition accumulation." During the week-long competition, it is also a rare and valuable opportunity to improve your practical experience and code ability. ”
The Earth AI competition revolves around image recognition and interpretation of artificial intelligence technology. The organizers provided 2100 pictures, and the target scene slices of these remote sensing images were classified into 21 categories, such as agricultural supplies, airplanes, beaches, buildings, jungles, houses, etc., with 100 pictures in each category. In the first six days of the training stage, the team automatically classified the training set of 2100 slices through artificial intelligence technology, and continuously optimized the algorithm model.
Pu Xinyang trained the model for the competition
On February 4, the day of the final, the organizers released the test data set for the final, including 2100 pictures and 140 pictures, without classification tags. In the last 2 hours, the teams used the algorithm to classify and identify the images and upload the identification results to the organizers.
The organizers announced that the Fudan team won the championship
In the end, the accuracy of the fudan Hollyday Coder team's indicators rode the dust and won the crown!
"One of the most important factors for our model to achieve the highest results is the accumulation of artificial intelligence technology in the field in the early stage of the laboratory." Captain Pu Xinyang said. The dataset open to this competition has only 100 samples per category, while the general deep learning technology requires at least thousands of samples per category to achieve high recognition accuracy. In addition, the classification of fine requirements are very high, such as residential areas need to be subdivided into low, medium and high densities, which increase the difficulty of accurate classification.
The laboratory has long studied the physical mechanism of remote sensing, deeply integrated the physical mechanism with the neural network, and developed an interpretable and strong generalized artificial intelligence technology. On the basis of the accumulation of previous technologies, in view of the difficulties such as small sample size and fine classification in this competition, the strategies of data enhancement, data augmentation and transfer learning were further adopted, which finally solved this problem, so that the trained model achieved extremely high accuracy in the test set.
At the time of the competition held in the New Year of Nongyin, it just came to an end on the opening day of the 2022 Beijing Winter Olympic Games, and on the days when the women's football team won the championship and the Olympic gold victory was frequently reported, Fudan students added a happy news of winning the international event.
"The significance of this competition is not only an exercise for us, but also gives us the opportunity to get to know and get to know the professors and scholars who are also committed to the research of artificial intelligence remote sensing images in many universities around the world, which promotes us to broaden our horizons in scientific research and inspires us to explore and develop in a wider field." Pu Xinyang said.
Among the members of the champion team,
Some have only one side, and some have never met
"Actually, I haven't met with Master Pu Xinyang yet, and I've only met once with Master Song Jingduo." Jai Hecheng said.
After learning about the Earth AI competition, Pu Xinyang and Song Jingduo, who had just entered the university last year, quickly decided to participate, looking forward to testing the image recognition technology they learned in practice.
As a doctoral brother, Jia Hecheng was hesitant at first when he received the invitation of his sisters to participate in the team. "After all, during the New Year, I want to spend more time with my family. After thinking about it later, I should still join the team, because the competition is a good exercise opportunity and you can learn a lot together. ”
The team in which Jia Hecheng participated won the 2020 "Tianzhi Cup" Artificial Intelligence Challenge
Jia Hecheng has a lot of experience in participating in the competition, and once teamed up with laboratory students to participate in the "Tianzhi Cup" artificial intelligence challenge in 2020 and won the championship. In the fourth "Zhongke Xingtu Cup" high-score remote sensing image interpretation software competition held in the same year, Jia Hecheng led the team to participate in the international track competition for optical remote sensing image aircraft object detection and recognition, ranking fifth. He has also participated in scientific research projects related to satellite remote sensing in the field for a long time and has rich practical experience.
Jia Hecheng (third from the left) and the team members of the 2020 "Zhongke Xingtu Cup" high-score remote sensing image interpretation software competition took a group photo in the laboratory
In this way, Jia Hecheng and Song Jingduo, who had only met once offline, and Pu Xinyang, who had not yet met, formed a team at the "cloud" end.
The three members of the team have a clear division of labor and are united and cooperative. Captain Pu Xinyang is responsible for the investigation of the current mainstream recognition algorithm, which is optimized based on the swin-transformer (deep self-attention network based on the design of the sliding window operation hierarchy). For this competition, they found that there was very little data open for each category, and data expansion and enhancement were the key tasks, which were responsible for Song Jingduo and Jia Hecheng. The two parts of algorithm tuning and data expansion go hand in hand, and when they are completed, the tuned algorithm is tested with augmented data, and other comparative algorithms are tested, and the final result with high accuracy is obtained.
Combined with some scene classification and target recognition data exposed to in the previous competition, as well as some scene classification datasets published online, Song Jingduo and Jia Hecheng expanded the data set by about seven times, from 2100 to more than 14000 data sets. Data augmentation and enhancement also have implications for their future scientific research work.
With two days left in the game, the algorithm model had been debugged, and Jai Hecheng insisted in the team group: "We need to complete the task of data enhancement so that the game is complete." This inspired their motivation, and the patience of captain Pu Xinyang in debugging the model impressed the teammates.
'Although the game was short, I learned a lot from both of my teammates, especially their perseverance. Song Jingduo said.
Behind the emergence of talents is the power of Academician Jin Yaqiu's example
The three members of the team believe that the champion and mentor Xu Feng's daily scientific research awareness are inseparable, answering questions and guiding ideas. As early as 2019, Professor Xu Feng and Professor Wang Haipeng of the Key Laboratory of Electromagnetic Wave Information Science of the Ministry of Education of Fudan University led the team to win two championships of the first "Tianzhi Cup" Artificial Intelligence Challenge, and won the first "Tianzhi Cup" Artificial Intelligence Challenge for two consecutive times. In a number of domestic and international competitions, there are laboratory students who have won awards.
Behind the emergence of talents is the power of the role model of Jin Yaqiu, an academician of the Chinese Academy of Sciences and a professor in the School of Information Science and Engineering, and the guidance of starlight. It is understood that the laboratory led by Academician Jin Yaqiu is the Shanghai Model Collective and the Key Laboratory of the Ministry of Education, and is also the "Zhong Yang-style" scientific research team of Fudan University in 2018.
What moved young students such as Pu Xinyang, Song Jingduo and Jia Hecheng was that at the weekly laboratory group meeting, the 76-year-old academician Jin Yaqiu was almost never absent. The large group meeting takes a long time, often 2 or 3 hours to start, after each student reports on the progress of scientific research, Professor Jin Yaqiu will give some guidance, effective suggestions for improvement, and share years of accumulated scientific research experience.
"Academician Jin, who is too old, has been involved in the cultivation of each student, and we have great admiration for his scientific research spirit, and also inspire our own scientific research awareness and diligent spirit." Song Jingduo said.
Under the guidance of Academician Jin Yaqiu and Professor Xu Feng and other scholars, the students of the Key Laboratory of Electromagnetic Wave Information Science of the Ministry of Education will march along the starlight, show fudan style in different arenas, and contribute Fudan strength to the national society.
Composition
Fusion Media Center
Text
Hu Huizhong
Image
Respondents provided
Editor-in-charge
Cheng Zhao
Editing
Bian Kaiyue and Qiu Jiexin