[Paper Review] Attention Modules Improve Image-Level Anomaly Detection for Industrial Inspection: A DifferNet Case Study

André Luiz Buarque Vieira-e-Silva, Francisco Simões+4 more|arXiv (Cornell University)|Nov 5, 2023

Industrial Vision Systems and Defect Detection4 citations

TL;DR

This paper proposes AttentDifferNet, an attention-augmented variant of the DifferNet anomaly detection model that improves image-level anomaly detection in industrial inspection by integrating channel and spatial attention modules into the feature extractor. It achieves state-of-the-art performance on the in-the-wild InsPLAD-fault dataset, with a 1.77±0.25 percentage point average AUROC improvement over DifferNet across three datasets.

ABSTRACT

Within (semi-)automated visual industrial inspection, learning-based approaches for assessing visual defects, including deep neural networks, enable the processing of otherwise small defect patterns in pixel size on high-resolution imagery. The emergence of these often rarely occurring defect patterns explains the general need for labeled data corpora. To alleviate this issue and advance the current state of the art in unsupervised visual inspection, this work proposes a DifferNet-based solution enhanced with attention modules: AttentDifferNet. It improves image-level detection and classification capabilities on three visual anomaly detection datasets for industrial inspection: InsPLAD-fault, MVTec AD, and Semiconductor Wafer. In comparison to the state of the art, AttentDifferNet achieves improved results, which are, in turn, highlighted throughout our quali-quantitative study. Our quantitative evaluation shows an average improvement - compared to DifferNet - of 1.77 +/- 0.25 percentage points in overall AUROC considering all three datasets, reaching SOTA results in InsPLAD-fault, an industrial inspection in-the-wild dataset. As our variants to AttentDifferNet show great prospects in the context of currently investigated approaches, a baseline is formulated, emphasizing the importance of attention for industrial anomaly detection both in the wild and in controlled environments.

Motivation & Objective

• Address the challenge of limited labeled defect data in industrial visual inspection by advancing unsupervised anomaly detection methods.
• Improve the performance of state-of-the-art anomaly detection models in uncontrolled, in-the-wild industrial inspection scenarios where variability in lighting, scale, and background is high.
• Investigate whether attention mechanisms can enhance feature representation in normalizing flow-based anomaly detection models like DifferNet.
• Establish a new baseline for attention-enhanced unsupervised anomaly detection by demonstrating consistent improvements across diverse industrial datasets.

Proposed method

• Integrate channel and spatial attention modules (specifically SENet blocks) into the backbone feature extractor of DifferNet to refine feature representations.
• Utilize normalizing flows to model the latent distribution of normal samples, enabling anomaly scoring based on likelihood estimation.
• Apply Grad-CAM to visualize and analyze attention-driven feature activation patterns, comparing attention-enhanced and baseline models.
• Train and evaluate AttentDifferNet on three datasets: InsPLAD-fault (in-the-wild), MVTec AD (controlled), and Semiconductor Wafer (controlled), using AUROC as the primary metric.
• Perform ablation studies by systematically removing attention blocks (AB1, AB2, AB3) to evaluate their individual and combined impact on performance.
• Compare AttentDifferNet with standard DifferNet, FastFlow, and RD++ to assess generalizability of the attention integration approach.

Experimental results

Research questions

• RQ1 Can attention modules improve the performance of DifferNet in image-level anomaly detection for industrial inspection, especially in uncontrolled, in-the-wild environments?
• RQ2 How does the integration of attention mechanisms affect feature representation and anomaly localization in normalizing flow-based models?
• RQ3 Does AttentDifferNet achieve state-of-the-art performance on the InsPLAD-fault dataset, which represents real-world industrial inspection challenges?
• RQ4 What is the impact of attention block placement and configuration on model performance, particularly in terms of AUROC improvement?
• RQ5 Can the attention-augmented approach generalize beyond DifferNet, improving other state-of-the-art unsupervised anomaly detection methods?

Key findings

• AttentDifferNet achieves a 1.77±0.25 percentage point average AUROC improvement over DifferNet across the InsPLAD-fault, MVTec AD, and Semiconductor Wafer datasets.
• On the InsPLAD-fault dataset, AttentDifferNet achieves state-of-the-art performance, outperforming existing methods in an in-the-wild industrial inspection setting.
• The ablation study shows that using all three attention blocks (AB1, AB2, AB3) yields the highest AUROC of 86.57% on the Glass Insulator category, significantly outperforming the baseline DifferNet (82.81%).
• Attention blocks placed in later layers (e.g., AB3) have a greater impact on performance, suggesting deeper feature refinement is more effective.
• Qualitative analysis using Grad-CAM shows that AttentDifferNet focuses more precisely on the object of interest in in-the-wild settings and on the defect itself in controlled environments.
• The attention-augmented versions of FastFlow and RD++ also show consistent performance gains, indicating the approach generalizes beyond DifferNet.