Prof. Zhenshen Qu
Harbin Institute of Technology, China
Biography: Prof. Zhenshen Qu received the B.S., M.S., and Ph.D. degrees in Control Science and Engineering from the Harbin Institute of Technology (HIT), in 1995,1998, and 2003, respectively. He is currently a Professor of Control Science and Engineering with the HIT. His current research interests include machine vision, autonomous systems, and swarm intelligence. Prof. Qu is the principle investigator or main participator of National Grand Science Project, National Key Project, NSFC project, and collaborative project with CETC. He published 56 papers and 21 patents, and received first and second grade awards for Science and Technology Advancement.
Speech Title: Industrial Defect Detection in Deep Learning and Large-scale Model Era
Abstract: The rapid advancement of deep learning and the availability of large-scale pre-trained models have significantly impacted the field of industrial defect detection using machine vision. We present an overview of the current state of the art in this domain, highlighting the key challenges, recent developments, and future prospects. Traditionally, the defect detection task has been carried out by human inspectors, but the increasing complexity and volume of manufacturing operations have made this approach less efficient and flexible. The emergence of deep learning has revolutionized the field of machine vision. However, it requires large and diverse datasets of labeled defect samples, which are scarce in real industrial settings. We examine the recent advancements in deep learning-based detection methods, such as anomaly detection, transfer learning, data augmentation and GANs, that aim to address the challenge of limited defect data. The integration of large-scale pre-trained visual models, is also explored as a means to leverage the feature representations learned from diverse datasets and accelerate the development of industrial defect detection solutions. We highlight the potential of these models to be adapted or integrated to specific domains, reducing the reliance on large-scale data collection. The introduced methods are demonstrated with real defect defection applications in electronic, pharmaceutical and automotive industry