Pipeline Anomaly Detection in Industry Manufacturing on UI

Overview

Anomalib is a deep learning library that aims to collect state-of-the-art anomaly detection algorithms for benchmarking on both public and private datasets. With my pull request #729, thank @ashwinvaidya17 and @jpcbertoldo for the suggestion.

In this project, I create a pipeline for training and inference with all dataset formats MVTec (custom) on UI. Everyone can train with all Anomalib models supported and quickly demo. And my project is an end-to-end example in the anomaly detection task.

For more details, everyone can visit and use my project at the link below.
Link my project: https://github.com/vnk8071/anomaly-detection-in-industry-manufacturing/tree/master/anomalib_contribute

Anomalib Version Note (Legacy vs Current)

• Anomalib Contributor: anomalib_contribute folder.
• New Commit: 70fabcd
• Deprecated README: https://github.com/openvinotoolkit/anomalib#1-web-based-pipeline-for-training-and-inference

This project was built against the legacy Anomalib API and pins anomalib==0.3.6.

• Legacy pins: anomalib_contribute/requirements.txt, deprecated/requirements.txt

Anomalib has since moved forward significantly (newer releases, updated APIs/CLI, docs). For the maintained upstream project, use:

• Upstream repo (current): https://github.com/open-edge-platform/anomalib
• Docs (current): https://anomalib.readthedocs.io/en/latest/

Historically, Anomalib was hosted under openvinotoolkit/anomalib; it is now maintained at open-edge-platform/anomalib.

If you install a newer Anomalib version, expect code/config changes may be required.

Project Layout

• anomalib_contribute/: implementation built on legacy Anomalib
• deprecated/: legacy snapshot (see folder for details)

Dataset

MVTec Anomaly Detection (MVTec AD)

MVTec AD is a dataset for benchmarking anomaly detection methods with a focus on industrial inspection. It contains over 5000 high-resolution images divided into fifteen different object and texture categories. Each category comprises a set of defect-free training images and a test set of images with various kinds of defects as well as images without defects.

Dataset link: https://www.mvtec.com/company/research/datasets/mvtec-ad

Installation

conda create -n anomaly-detection python=3.10
conda activate anomaly-detection
pip install -r requirements.txt

Flow

Custom dataset

For each new dataset, the data consist of three folders:

• train, which contains the (defect-free) training images
• test, which contains the test images
• ground_truth, which contains the pixel-precise annotations of anomalous regions

Hardware

GPU available: True (mps), used: True
TPU available: False, using: 0 TPU cores

┏━━━┳━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━┳━━━━━━━━┳━━━━━━━┳━━━━━━━┓
┃   ┃ Name           ┃ Type           ┃ Params ┃ Mode  ┃ FLOPs ┃
┡━━━╇━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━╇━━━━━━━━╇━━━━━━━╇━━━━━━━┩
│ 0 │ pre_processor  │ PreProcessor   │      0 │ train │     0 │
│ 1 │ post_processor │ PostProcessor  │      0 │ train │     0 │
│ 2 │ evaluator      │ Evaluator      │      0 │ train │     0 │
│ 3 │ model          │ PatchcoreModel │  2.8 M │ train │     0 │
└───┴────────────────┴────────────────┴────────┴───────┴───────┘
Trainable params: 2.8 M
Non-trainable params: 0
Total params: 2.8 M
Total estimated model params size (MB): 11
Modules in train mode: 19
Modules in eval mode: 69
Total FLOPs: 0

Models

Train

python train.py --config "configs/patchcore_grid.yaml" --model "patchcore"

Evaluation (Coming soon)

Inference

python script_inference.py --config "configs/patchcore_hazelnut.yaml" --weight "models/patchcore_hazelnut.ckpt" --image "samples/007_hazelnut.png"

Or (simple default):

python script_inference.py

App

Gradio

python demo.py

Open local URL: http://127.0.0.1:7860

Sample:

FastAPI

python app.py

Open local URL: http://127.0.0.1:8000

Homepage:

Train:

Inference:

Container

docker build -t anomaly:v1 .
docker run anomaly:v1

Or:

docker-compose up