Introduction

Anomaly-detection (AD): is a task where the goal is to find anomalous samples at inference time while during training, only positive (good) samples are given.

Denoising Diffusion Model (DDM): DDMs are models trained to recover images from noisy versions of the same image; they have recently been proven useful for many tasks (with a focus on generative models).

Our Objective

Goal: In this project we will attempt to develop a POC for detecting anomalies in images based on the ability or inability of a DDM to reconstruct them.

Example: Inspection of a product in a factory may take images of all products on the product line. The goal may be to find scratched or damaged products, while during training no such samples were given.

High Level Methodology

• Add random gaussian noise to an image.

• Denoise the noised image using a DDM and reconstruct the image.

• Calculate the “difference” between the original image and the reconstructed image and use the result to determine an anomality score.

Setup and Execution

Download the 256x256 class unconditional model from here: 256x256_diffusion_uncond.pt

Place it in the path models/256x256x_diffusion_uncond.pt (relative path from the root directory of the repository).

Execution

Run all the cells of main_experiment.ipynb.
You can see the output in the following paths:

• output/ - The root folder of the outputs. Everything will be in here.
• output/results.csv - A csv table containing the columns category | category_type | img_auc | pixel_auc with the final results per category. Will have the results of the last execution for each class. Can optimize and set overwrite = False in main_experiment.ipynb in the proper places and that will make sure that only categories that are not already present in the csv will be evaluated.
• output/<category_name>/lightning_logs/ - Can find many output files here. Will have a version per evaluation performed. Each version will have a sample sub-directory inside of it that will included visualizations.