From
Timetable to
Place Amphi Dorothy Hodgkin, ENS Paris-Saclay
Aitor Artola
A generic method for detecting anomalies on manufactured parts
Abstract
This thesis investigates learning-based methods for industrial optical quality control. Its goal is to provide automatic and robust solutions that don't require a computer vision expert to adapt to a new object. In the first part, we propose to add an image normalization first step to the anomaly detection pipeline, that facilitates the subsequent learning steps.
We first study image registration methods and develop an alternative for the images of repetitive objects, that is robust to the symmetries and redundant patterns. We also propose an unsupervised image normalization neural network learning method to reduce the impact of normal variability on false detections.
The second part deals with stochastic background models used to learn the structure of anomaly-free objects and to measure the normality of new objects. We first provide a review of these methods and show that simple background models can achieve state-of-the-art performance when using pre-trained neural networks as image embeddings. We then propose GLAD, a Gaussian Mixture Model (GMM)-based anomaly detector that uses global-to-local training to be data efficient. We then demonstrate that recently developed anomaly detectors are not adapted to real case scenarios due to biases in benchmark datasets. In particular, we point out the fragility of these models under moderate capture variations of the observed objects. We therefore propose an online version of GLAD that is able to adapt to product evolution without retraining from scratch. In the last chapter of this part, we show that GLAD is a generalist anomaly detector, by applying it to video change detection. In the last part, we discuss data embedding. Most solutions, including ours, use semantic classification as pretext task to compute an image embedding. We show that the use of such backbones is not adapted to the detection of small anomalies. We develop two alternatives to solve this problem in the last two chapters. First, we explore semantic segmentation backbones and additional multiscale strategies for anomaly detection. Then, in the final chapter, we study how to fine-tune the embedding backbones using self-supervised and contrastive learning to improve the treatment of normal variability.
Jury
- Charles Kervrann, directeur de recherche, INRIA de l’Université de Rennes, France
- Danijel Skocaj, professor at the University of Ljubljana, Slovenia
- Dorit Merhof, Professor, University of Regensburg, Germany
- Angélique Loesch, ingénieur de Recherche, CEA Saclay - Nano Innov, France
- Pablo Musé, Professor , Universidad de la Republica, Uruguay