Skip Ribbon Commands
Skip to main content

NTU EEE PhD Student wins international award


Yang Jianfei, a PhD student supervised by Prof Xie Lihua from NTU EEE, led the team to win the 2018 Pacific Earthquake Engineering Research (PEER) Hub ImageNet Challenge. He is the team leader and other his teammates are from Sun Yat-sen University, Chinese Academy of Sciences and UC Berkeley.

The 2018 PEER Hub ImageNet Challenge is the first image-based structural damage recognition competition, held during August to December, 2018. As a pioneer study, the challenge collects a large image dataset which is relevant to the field of structural engineering, and designs eight difficult detection tasks including scene recognition, damage check, spalling condition check, material type recognition, collapse check, component type recognition, damage level recognition and damage type classification. These vision-based tasks will contribute to the establishment of automated structural health monitoring in the civil engineering. 

 Over 50 teams from both universities and industries participate in the challenge. The universities include Stanford university, UC Berkeley, UCLA, Tsinghua university, Purdue University and so on. Some competitors are from AI companies with much experience in computer vision applications. 

In the challenge, Jianfei and his teammates proposed a multi-task transfer learning framework for these eight tasks, and won the champion with the highest overall score and the first rank of four tasks. The challenge includes eight image classification problems. The traditional way is to build eight independent machine learning models. This is very cumbersome and it ignores the relationship among tasks. The team build a unified multi-task transfer learning framework that uses shared feature extractor and distinctive classifier. The algorithm they coded can learn eight tasks simultaneously, which helps to learn more robust and common patterns suitable for all tasks. Furthermore, they pay attention to the hierarchical relationship among eight tasks, and hence they design a particular training strategy. Finally the framework took the first place in the overall accuracy and four independent tasks. The team’s report was also awarded as the best technical report in the meeting.

Jianfei, the team leader, was invited to attend the annual meeting of PEER 2019 in UCLA, USA. The director of PEER, Prof. Khalid Mosalam, awarded the team personally during the meeting. He praised that the team from NTU won the challenge with undisputed score and thanks for the contribution of promoting the research of interdisciplinary interactions with machine learning and civil engineering.

Click here to find out more information on PHI Challenge Winners. 



Published on 20 March 2019​​​​

Not sure which programme to go for? Use our programme finder
Loading header/footer ...