# Diversity-Measurable Anomaly Detection

**Authors:** Wenrui Liu, Hong Chang, Bingpeng Ma, Shiguang Shan, Xilin Chen  
**arXiv:** [2303.05047](https://arxiv.org/abs/2303.05047)  
**Date:** 2023-03-09  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.05047v1/

## Key points

- Limitations of reconstruction-based anomaly detection models
- Previous approaches suffer from shortcut learning and transmission of abnormal information
- Introduction of Diversity-Measurable Anomaly Detection (DMAD) framework
- DMAD aims to enhance reconstruction diversity while avoiding undesired generalization on anomalies
- Pyramid Deformation Module (PDM) models diverse normals and measures severity of anomalies
- Information compression module improves reliability of final anomaly score
- Experimental results demonstrate effectiveness of DMAD in front of contaminated data and anomaly-like normal samples
- Novel approach for anomaly detection that addresses limitations and achieves improved performance in various scenarios

## Summary

The paper discusses the limitations of reconstruction-based anomaly detection models which suppress the generalization ability for anomalies but fail to accurately reconstruct diverse normal patterns. Previous approaches have attempted to address this issue by modeling sample diversity, however they often suffer from shortcut learning leading to the transmission of abnormal information. To overcome this tradeoff problem, the authors propose a framework called Diversity-Measurable Anomaly Detection (DMAD). DMAD aims to enhance reconstruction diversity while avoiding undesired generalization on anomalies by introducing a Pyramid Deformation Module (PDM) that models diverse normals and measures the severity of anomalies by estimating multi-scale deformation fields from reconstructed reference to original input. An information compression module is also integrated into PDM which decouples deformation from prototypical embedding and improves the reliability of the final anomaly score. Experimental results on surveillance videos and industrial images demonstrate the effectiveness of DMAD in front of contaminated data and anomaly-like normal samples. This paper presents a novel approach for anomaly detection that addresses the limitations of existing methods and achieves improved performance in various scenarios.