Automated Industrial Surface Defect Detection and Localization Using Deep Anomaly Detection

Members

Problem and Motivation

Surface defect detection is a critical quality control task in manufacturing. Manual inspection is slow, subjective and error-prone under high-throughput conditions. Automated detection using computer vision can reduce waste, lower costs and improve consistency. However, defective samples are rare in practice, making supervised classification impractical because there are not enough labeled examples of every possible defect type.

This project explores unsupervised anomaly detection, where models are trained exclusively on defect-free images and learn to flag anything that deviates from normality at test time. We compare a reconstruction-based baseline with a state-of-the-art feature-embedding method and add a quantitative explainability layer so the system reports not only whether an image is anomalous but why.

Goals

1. Detect whether a test image contains a defect (image-level binary classification)
2. Localize the defect at pixel level (produce an anomaly heatmap)
3. Explain why the model flagged the image, with quantitative evaluation of explanation quality

We evaluate on three categories from the MVTec Anomaly Detection (MVTec AD) dataset: Bottle (rigid object), Hazelnut (organic object) and Carpet (texture), covering structurally different defect scenarios.

Pipeline Overview

flowchart TD
    A["1. Download MVTec AD Dataset
3 categories: bottle, hazelnut, carpet"]
    B["2. Exploratory Data Analysis
notebooks/eda.ipynb"]
    C["3. Preprocessing
Resize 256x256, Normalize, Mask Binarization"]
    D["4. Train Autoencoder Baseline
MSE loss, Adam, Early Stopping"]
    E["5. Build PatchCore Memory Bank
WideResNet50 backbone, Coreset Subsampling"]
    F["6. Evaluate Both Methods
Image AUROC, Pixel AUROC, AUPRO"]
    G["7. Qualitative Visualization
2 correct detections + 1 failure case"]
    H["8. Explainability Analysis"]
    H1["Grad-CAM
Autoencoder"]
    H2["Occlusion Sensitivity
PatchCore"]
    I["9. XAI Evaluation
Pointing Game, Top-5% IoU"]
    J["10. Backbone Ablation
ResNet18 vs WideResNet50"]
    K["Final Results
metrics.csv, ablation.csv, explainability.csv"]

    A --> B
    A --> C
    C --> D
    C --> E
    D --> F
    E --> F
    F --> G
    D --> H
    E --> H
    H --> H1
    H --> H2
    H1 --> I
    H2 --> I
    E --> J
    F --> K
    G --> K
    I --> K
    J --> K

    style A fill:#4a90d9,color:#fff,stroke:#4a90d9
    style D fill:#e67e22,color:#fff,stroke:#e67e22
    style E fill:#27ae60,color:#fff,stroke:#27ae60
    style F fill:#8e44ad,color:#fff,stroke:#8e44ad
    style H fill:#c0392b,color:#fff,stroke:#c0392b
    style K fill:#2c3e50,color:#fff,stroke:#2c3e50
    style H1 fill:#7c3aed,color:#fff,stroke:#7c3aed
    style H2 fill:#0d9488,color:#fff,stroke:#0d9488
        

Methods

Explainability

Each method is explained with a different attribution technique, chosen based on architectural constraints:

Both methods are scored against ground truth masks using Pointing Game accuracy (does the peak attribution fall inside the defect?) and Top-5% IoU (overlap of the hottest 5% of attribution pixels with the defect).

Results

Detection and Localization

PatchCore achieves near-perfect scores on all categories. The autoencoder fails on carpet (Image AUROC 0.319, worse than random) because reconstruction error cannot distinguish defects from normal texture variation.

Backbone Ablation

WideResNet50 provides marginal accuracy gains (≤1.3 points) at 25-35% higher inference cost. ResNet18 is preferable under latency constraints.

Explainability Metrics

Both methods operate at 16x16 spatial resolution, limiting tight localization. The autoencoder's 0.0 Pointing Game on carpet mirrors its detection failure: an explanation cannot point at a defect the model is not finding.

Technologies Used

References

1. Bergmann, P., Fauser, M., Sattlegger, D., & Steger, C. (2021). The MVTec Anomaly Detection Dataset: A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection. IJCV, 129, 1038-1059.
2. Roth, K., Pemula, L., Zepeda, J., Scholkopf, B., Brox, T., & Gehler, P. (2022). Towards Total Recall in Industrial Anomaly Detection. CVPR.
3. Selvaraju, R. R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., & Batra, D. (2017). Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization. ICCV.
4. Zeiler, M. D., & Fergus, R. (2014). Visualizing and Understanding Convolutional Networks. ECCV.
5. Zhang, J., Bargal, S. A., Lin, Z., Brandt, J., Shen, X., & Sclaroff, S. (2018). Top-Down Neural Attention by Excitation Backprop. IJCV.