https://github.com/ordinary9843/aoi-defect-detection

Unsupervised electronic component defect detection · PatchCore (CVPR 2022) · WideResNet50-2 · AUROC 0.9487 on MVTec AD transistor
https://github.com/ordinary9843/aoi-defect-detection

anomaly-detection aoi computer-vision defect-detection docker fastapi mvtec patchcore semiconductor

Last synced: 3 days ago
JSON representation

Unsupervised electronic component defect detection · PatchCore (CVPR 2022) · WideResNet50-2 · AUROC 0.9487 on MVTec AD transistor

• Host: GitHub
• URL: https://github.com/ordinary9843/aoi-defect-detection
• Owner: ordinary9843
• License: mit
• Created: 2026-05-21T17:09:49.000Z (about 1 month ago)
• Default Branch: master
• Last Pushed: 2026-05-22T08:24:07.000Z (about 1 month ago)
• Last Synced: 2026-05-22T12:41:16.642Z (about 1 month ago)
• Topics: anomaly-detection, aoi, computer-vision, defect-detection, docker, fastapi, mvtec, patchcore, semiconductor
• Language: HTML
• Size: 425 KB
• Stars: 0
• Watchers: 0
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

          
# AOI Defect Detection

Unsupervised electronic component defect detection — **no defect labels required**.

Built on [PatchCore (CVPR 2022)](https://openaccess.thecvf.com/content/CVPR2022/html/Roth_Towards_Total_Recall_in_Industrial_Anomaly_Detection_CVPR_2022_paper.html) with WideResNet50-2 backbone, evaluated on the MVTec AD benchmark.

---

## Demo

| Normal component | Defective component |

|---|---|

| ![PASS](assets/demo-pass.png) | ![DEFECT](assets/demo-defected.png) |

**Benchmark — AUROC 0.9487 with ROC curve**

![Benchmark](assets/demo-benchmark.png)

---

## Results

| Metric | Value |

|---|---|

| **AUROC** | **0.9487** |

| Normal pass-through rate | 98.3% |

**Backbone ablation** — transistor defects are geometric (bent leads, misplacement), requiring wider feature representations:

| Backbone | AUROC |

|---|---|

| ResNet18 | 0.8750 |

| **WideResNet50-2** | **0.9487** |

Dataset: MVTec AD — transistor (213 training images, 100 test images)

---

## How It Works

![Architecture](assets/architecture.png)

- **No GPU training** — uses pretrained ImageNet features

- **Calibration separated from build** — threshold set on held-out normal images

- **Heatmap output** — patch-level anomaly localization upsampled to original resolution

---

## Production Architecture

*Cloud services: AWS*

![Production Architecture](assets/production.png)

- **SQS** — decouples camera trigger from inference; handles burst traffic from multiple stations

- **ECS** — containerized inference workers scale horizontally without managing EC2 instances

- **S3** — stores defect images and memory bank `.pkl`; workers pull latest model on startup

- **RDS** — persists per-image pass/fail results and anomaly scores for traceability

- **CloudWatch** — monitors defect rate trend and anomaly score distribution drift

---

## Quick Start

```bash

git clone https://github.com/ordinary9843/aoi-defect-detection.git

cd aoi-defect-detection

docker-compose up --build

```

Open `http://localhost:8000`

Download `transistor.tar.xz` from [MVTec AD](https://www.mvtec.com/company/research/datasets/mvtec-ad) → extract to `data/transistor/`

| Step | Input path |

|---|---|

| 1 — Build | `data/transistor/train/good` |

| 2 — Calibrate | `data/transistor/test/good` |

| 3 — Detect | any image |

| 4 — Benchmark | `data/transistor` |

---

## Performance

| Environment | Latency |

|---|---|

| CPU — Docker | ~2000 ms |

| GPU — RTX 3060 (Docker) | ~400–700 ms |

| Production GPU server | < 100 ms |

*Tested on i5-14400F, 16 GB RAM, RTX 3060 — GPU auto-detected via CUDA*

---

## Limitations

| Issue | Cause | Fix |

|---|---|---|

| `damaged_case` not detected | Subtle defect; patch features insufficient | Higher resolution / pixel-level segmentation |

| Memory bank grows with data | Full patch storage, O(N × 784 × 1536) | Coreset subsampling (PatchCore §3.3) |

| Single global threshold | One 99th-percentile value | Per-category calibration |

| CPU inference too slow for production | GPU requires NVIDIA Container Toolkit on host | Install [NVIDIA Container Toolkit](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html) |

---

## Stack

| Layer | Technology |

|---|---|

| Algorithm | PatchCore (CVPR 2022), WideResNet50-2 |

| Backend | FastAPI, Python 3.11, SSE streaming |

| Inference | scikit-learn kNN, OpenCV |

| Deployment | Docker, docker-compose |

| Frontend | Vanilla JS, HTML canvas |

---

**References**

- Roth et al., *Towards Total Recall in Industrial Anomaly Detection*, CVPR 2022

- Bergmann et al., *MVTec AD*, CVPR 2019