# SimpleNet: A Simple Network for Image Anomaly Detection and Localization

**Authors:** Zhikang Liu, Yiming Zhou, Yuansheng Xu, Zilei Wang  
**arXiv:** [2303.15140](https://arxiv.org/abs/2303.15140)  
**Date:** 2023-03-27  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.15140v2/

## Key points

- Authors propose SimpleNet, a network for detecting and localizing anomalies in images
- SimpleNet comprises four components: Feature Extractor, Feature Adapter, Anomaly Feature Generator, and Anomaly Discriminator
- Approach based on transforming pre-trained features, generating synthetic anomalies in feature space, and using a simple discriminator
- SimpleNet outperforms previous methods quantitatively and qualitatively
- Achieves anomaly detection AUROC of 99.6% on MVTec AD benchmark dataset
- Reduces error by 55.5% compared to next best performing model
- Offers faster processing times with a frame rate of 77 FPS on NVIDIA GeForce RTX 3080ti GPU
- Implementation code available on authors' GitHub repository at https://github.com/DonaldRR/SimpleNet
- Research accepted for presentation at CVPR 2023 conference

## Summary

Authors Zhikang Liu, Yiming Zhou, Yuansheng Xu, and Zilei Wang have proposed a simple and application-friendly network called SimpleNet for detecting and localizing anomalies in images. SimpleNet comprises four components: a pre-trained Feature Extractor that generates local features, a shallow Feature Adapter that transfers these local features to the target domain, a simple Anomaly Feature Generator that adds Gaussian noise to normal features to create synthetic anomaly features, and a binary Anomaly Discriminator that distinguishes between anomaly and normal features. During inference, the Anomaly Feature Generator is discarded. The authors base their approach on three intuitions. First, transforming pre-trained features into target-oriented features helps avoid domain bias. Second, generating synthetic anomalies in feature space is more effective as defects may not exhibit much commonality in the image space. Third, a simple discriminator is more efficient and practical. Despite its simplicity, SimpleNet outperforms previous methods both quantitatively and qualitatively. On the MVTec AD benchmark dataset, SimpleNet achieves an anomaly detection AUROC (Area Under the Receiver Operating Characteristic) of 99.6%, reducing the error by 55.5% compared to the next best performing model. Additionally, SimpleNet demonstrates significant improvements in performance on the One-Class Novelty Detection task. Notably, SimpleNet also offers faster processing times than existing methods with a high frame rate of 77 frames per second (FPS) on an NVIDIA GeForce RTX 3080ti GPU. For further details and implementation code of SimpleNet interested readers can refer to the authors' GitHub repository at https://github.com/DonaldRR/SimpleNet . This research has been accepted for presentation at CVPR 2023 conference.