# TrustMAE: A Noise-Resilient Defect Classification Framework using Memory-Augmented Auto-Encoders with Trust Regions

**Authors:** Daniel Stanley Tan, Yi-Chun Chen, Trista Pei-Chun Chen, Wei-Chao Chen  
**arXiv:** [2012.14629](https://arxiv.org/abs/2012.14629)  
**Date:** 2020-12-29  
**Summary page:** https://www.summarizepaper.com/en/paper/2012.14629v1/

## Key points

- TrustMAE is a framework for product defect classification
- TrustMAE can accept datasets with unlabeled images, unlike other methods
- It is robust against noises or defective images in the training dataset
- Utilizes a memory-augmented auto-encoder with a sparse memory addressing scheme
- Incorporates a trust-region memory updating scheme to prevent noise in memory slots
- Can reconstruct defect-free images and identify defective regions using a perceptual distance network
- Performs competitively on noise-free MVTec datasets and outperforms other baselines at up to 40% noise level
- Achieves the best AUC score on most classes and overall mean AUC compared to other works on defect detection
- Lower performance in certain classes due to difficulties in modeling large variations in normal data caused by rotations or randomly appearing prints
- Does not include a warping module but compares favorably to recent works in defect segmentation using pixel-level AUC evaluation

## Summary

, , , , . In this paper, the authors propose a framework called TrustMAE to address the problem of product defect classification. The existing summary highlights that TrustMAE is capable of accepting datasets with unlabeled images, unlike other methods that rely on difficult-to-collect and laborious-to-label defective images. Additionally, TrustMAE is robust against noises or defective images in the training dataset, which sets it apart from most anomaly detection methods. The framework utilizes a memory-augmented auto-encoder with a sparse memory addressing scheme to prevent over-generalization. It also incorporates a novel trust-region memory updating scheme to keep noises away from the memory slots. This combination allows TrustMAE to reconstruct defect-free images and identify defective regions using a perceptual distance network. To evaluate its performance, TrustMAE is compared against various state-of-the-art baselines using noise-free MVTec datasets. The results show that TrustMAE performs competitively under these conditions. However, what makes it particularly impressive is its effectiveness at a noise level of up to 40%, significantly outperforming other baselines. The authors discuss some additional details about their work and mention that their model achieves the best AUC score on most classes and overall mean AUC when compared to other works on defect detection. However, they acknowledge lower performance in certain classes such as metal nut, pill, screw, and capsule due to difficulties in modeling large variations in normal data caused by rotations or randomly appearing prints. The authors note that while adding a warping module could improve performance by aligning data, it falls outside the scope of their paper. They also compare their method with prior works in defect segmentation using pixel-level AUC evaluation. By imposing a threshold on reconstruction errors, they obtain segmentation maps and demonstrate competitive performance compared to recent works in defect segmentation. Overall, TrustMAE proves to be an effective framework for product defect classification. Its ability to handle unlabeled datasets, robustness against noisy training data, and competitive performance make it a valuable contribution to the field.