IEEE.orgIEEE Xplore Digital Library |  IEEE Standards |  IEEE Spectrum |  More Sites

URUCON 2024 - Banner

Keynote

"Defect Detection in Industrial Manufacturing Using Generative Image Modeling"

PhD. Pablo Musé

Full Professor, IIE, Facultad de Ingeniería, UdelaR

Session language: spanish with english subtitles

Abstract

Image anomaly detection is a classic problem that, in recent years, has received increasing interest from the computer vision community, motivated by applications ranging from surveillance and security to medical imaging. One of the most common applications, which we will address in this talk, is the automation of product quality control in an industrial environment. Defects being anomalous events, it is difficult or even impossible to collect and label a sufficient amount of defective samples to achieve a good statistical characterization. For this reason, most efforts in defect detection methods have been focused on learning the distribution of normal samples and detecting defects as unlikely events under the normality assumption. In this work we propose a high-dimensional density learning method based on normalizing flows that, starting from a multi-scale feature extraction of the image using a transformer architecture, allows for an accurate estimation of the likelihood of structures in an image under the normality hypothesis. Combining this estimation with the multiple hypothesis testing methodology, we obtain an automatic image anomaly detection method that surpasses the state of the art in several benchmark databases. Joint work with Matías Tailanián and Álvaro Pardo.

Biography

Pablo Musé received an electrical engineering degree from Universidad de la República, Uruguay (1999), an MSc. in mathematics and statistical learning (2001), and a Ph.D. in applied mathematics (2004), both from École Normale Supérieure Paris-Saclay, France. From 2005 to 2006, he was a Senior Researcher with Cognitech, Pasadena, CA, USA, where he worked on computer vision and image processing applications. In 2006 and 2007, he was a postdoctoral scholar at the Seismological Laboratory, California Institute of Technology, Pasadena, working on remote sensing using optical imaging, radar, and GPS networks. Since 2008, he has been with the Division of Electrical Engineering, Universidad de la República, where he is currently a Full Professor of signal processing. He is also an associate researcher at Centre Borelli of ENS Paris-Saclay and a co-founder and managing partner at Digital Sense. His research interests include machine learning, image restoration and analysis, computational photography, and remote sensing.