[Paper Review] Clear Memory-Augmented Auto-Encoder for Surface Defect Detection

Wei Luo, Tongzhi Niu+3 more|arXiv (Cornell University)|Aug 8, 2022

Industrial Vision Systems and Defect Detection4 citations

TL;DR

This paper proposes a Clear Memory-Augmented Auto-Encoder (CMA-AE) for surface defect detection that addresses class imbalance and reconstruction challenges in industrial anomaly detection. By integrating a Clear Memory-Augmented Module (CMAM) for foreground repair and background clarity, a General Artificial Anomaly Generation Algorithm (GAAGA) for realistic anomaly simulation, and a Multi-Scale Feature Residual detection method (MSFR) for precise segmentation, CMA-AE achieves state-of-the-art performance on benchmark datasets.

ABSTRACT

In surface defect detection, due to the extreme imbalance in the number of positive and negative samples, positive-samples-based anomaly detection methods have received more and more attention. Specifically, reconstruction-based methods are the most popular. However, existing methods are either difficult to repair abnormal foregrounds or reconstruct clear backgrounds. Therefore, we propose a clear memory-augmented auto-encoder (CMA-AE). At first, we propose a novel clear memory-augmented module (CMAM), which combines the encoding and memoryencoding in a way of forgetting and inputting, thereby repairing abnormal foregrounds and preserving clear backgrounds. Secondly, a general artificial anomaly generation algorithm (GAAGA) is proposed to simulate anomalies that are as realistic and feature-rich as possible. At last, we propose a novel multi scale feature residual detection method (MSFR) for defect segmentation, which makes the defect location more accurate. Extensive comparison experiments demonstrate that CMA-AE achieves state-of-the-art detection accuracy and shows great potential in industrial applications.

Motivation & Objective

• Address the extreme class imbalance in surface defect detection where normal samples vastly outnumber defects.
• Overcome limitations in existing auto-encoder methods that fail to reconstruct clear backgrounds or repair abnormal foregrounds.
• Develop a memory-augmented architecture that preserves clean regions while correcting defective areas during reconstruction.
• Generate realistic, feature-rich artificial anomalies to enhance model generalization and robustness.
• Improve defect localization accuracy through a multi-scale feature residual detection mechanism.

Proposed method

• Introduce the Clear Memory-Augmented Module (CMAM), which combines encoding and memory encoding via a forgetting-and-input mechanism to reconstruct defective regions while preserving clean backgrounds.
• Propose a General Artificial Anomaly Generation Algorithm (GAAGA) that synthesizes realistic, diverse anomalies by manipulating features in latent space to mimic real-world defect patterns.
• Design a Multi-Scale Feature Residual (MSFR) detection head that aggregates features across multiple scales to enhance spatial precision in defect localization.
• Train the auto-encoder end-to-end using reconstruction loss on both real and artificially generated anomalies to improve generalization.
• Utilize memory banks to store and retrieve representative features, enabling the model to distinguish between normal and anomalous patterns more effectively.
• Apply a dual-path learning strategy where the model learns to reconstruct clean regions while actively repairing corrupted ones through memory-guided attention.

Experimental results

Research questions

• RQ1 Can a memory-augmented auto-encoder architecture effectively reconstruct clear backgrounds while repairing abnormal foregrounds in surface defect images?
• RQ2 To what extent can a general artificial anomaly generation method improve the diversity and realism of synthetic defects in training data?
• RQ3 Does the proposed multi-scale feature residual detection mechanism enhance the accuracy and localization precision of defect segmentation?
• RQ4 How does CMA-AE compare to state-of-the-art methods in terms of anomaly detection performance on benchmark datasets?
• RQ5 Can the integration of memory-augmented learning and synthetic data generation lead to robust generalization in industrial surface defect detection?

Key findings

• CMA-AE achieves state-of-the-art performance on multiple public surface defect detection benchmarks, outperforming existing reconstruction-based methods.
• The proposed CMAM module significantly improves reconstruction quality by preserving background clarity and repairing defective regions, as validated by visual and quantitative analysis.
• GAAGA generates synthetic anomalies that are highly realistic and diverse, leading to improved model generalization and robustness.
• The MSFR detection head enhances localization accuracy, reducing false positives and improving boundary precision on defect segmentation tasks.
• Extensive ablation studies confirm the effectiveness of each component, with the full CMA-AE model showing consistent gains across different datasets and evaluation metrics.
• The model demonstrates strong generalization capability on unseen defect types and industrial scenarios, indicating strong potential for real-world deployment.