A generic method for detecting anomalies on manufactured parts

This thesis investigates learning-based methods for industrial optical quality control. It provides autonomous solutions that do not require a computer vision expert to adapt to a new object. In the first part, we add an image normalization first step to the anomaly detection pipeline to facilitate learning. We study image registration methods and develop an alternative for the images of repetitive objects, that is robust to symmetries and repeated patterns. We also propose an unsupervised image normalization neural network that learn to reduce the impact of textural variability. 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 with embeddings provided by standard backbones. We then propose GLAD, a global to local Gaussian Mixture Model anomaly detector. We then demonstrate that recently developed anomaly detectors are not adapted to real case scenarios. 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 adapts to product evolution. In the last chapter of this part, we show that GLAD easily extends 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. First, we explore semantic segmentation backbones and additional multiscale strategies for anomaly detection. In the final chapter, we study how to fine-tune the embedding backbones using self-supervised and contrastive learning.