Methodological Guideline for Anomaly Detection Models

This document presents several out-of-distribution data-driven detection methods, based on neural networks,
particularly adapted to the Safran Visual Industrial Control and Renault Welding use cases. Among these
methods, there is a new design method for improving the detection of out-of-distribution data by a neural
network using ensemble methods, and two reconstruction based methods using Normalizing Flows or Vision
Transformers. One reconstruction based method with Normalizing Flows from the DEEL project is also
presented. This document contains also a small state on the art on reconstruction based method by Diffusion
Models.