Collaborative Discrepancy Optimization for Reliable Image Anomaly Localization

IEEE Transactions on Industrial Informatics 2023

🔥 News

2025.03: 🎉🎉 Check out our updated version of CDO! You can find significant improvements versus our original version! The current version also supports both semi-supervised and unsupervised anomaly detection.
2023.03: 🎉🎉 We published a new paper related to point cloud anomaly detection, Complementary Pseudo Multimodal Feature for Point Cloud Anomaly Detection.
🔗 Paper | 🔗 Code

Abstract

Most unsupervised image anomaly localization methods suffer from overgeneralization due to the high generalization abilities of convolutional neural networks, leading to unreliable predictions. To mitigate this, we propose Collaborative Discrepancy Optimization (CDO), which optimizes normal and abnormal feature distributions with synthetic anomalies. CDO introduces a margin optimization module and an overlap optimization module to maximize the margin and minimize the overlap between the discrepancy distributions (DDs) of normal and abnormal samples. Experiments on MVTec2D and MVTec3D demonstrate that CDO effectively mitigates overgeneralization, achieving excellent anomaly localization performance with real-time computation efficiency.

BibTex Citation

If you find our paper or code useful, please cite us using the following BibTex:

@ARTICLE{10034849,
  author={Cao, Yunkang and Xu, Xiaohao and Liu, Zhaoge and Shen, Weiming},
  journal={IEEE Transactions on Industrial Informatics}, 
  title={Collaborative Discrepancy Optimization for Reliable Image Anomaly Localization}, 
  year={2023},
  volume={},
  number={},
  pages={1-10},
  doi={10.1109/TII.2023.3241579}}

Quick Start

Create a new conda environment and install the required packages:

conda create -n cdo_env python=3.9.12
conda activate cdo_env
pip install -r requirements.txt

Data Preparation

We support the MVTec AD dataset and VisA dataset for anomaly localization in factory settings. Unzip the files to the following directories:

datasets/
├── mvtec_anomaly_detection/
│   ├── bottle/
│   ├── cable/
│   ├── ...
│   └── meta.json
└── VisA_20220922/
    ├── candle/
    ├── capsules/
    ├── ...
    └── meta.json

To generate the required JSON files for training, run the following scripts:

python ./data/gen_metadata/mvtec.py
python ./data/gen_metadata/visa.py

Training Models

Execute Different Training Tasks

Modify the self.train_data.mode and self.test_data.mode parameters in ./configs/benchmark/cdo/cdo_256_100e.py to select different training tasks. The available tasks and their corresponding metadata files are:

EXPERIMENTAL_SETUP = {
    'zero_shot': 'meta_zero_shot.json',
    'few_shot1': 'meta_few_shot1.json',
    'few_shot2': 'meta_few_shot2.json',
    'few_shot4': 'meta_few_shot4.json',
    'few_shot8': 'meta_few_shot8.json',
    'unsupervised': 'meta_unsupervised.json',
    'semi1': 'meta_semi1.json',
    'semi5': 'meta_semi5.json',
    'semi10': 'meta_semi10.json',
}

For example, to run unsupervised training:

self.train_data.mode = 'unsupervised'
self.test_data.mode = 'unsupervised'

Run Experiment

python run_dataset.py

CDO Architecture

CDO Framework

Reference CDO Quantitative Results

Quantitative Results

The original code has been further refined and modified. The following are the quantitative results of the current version on the MVTec Dataset and VisA Dataset.

Table 1: Quantitative Results of Unsupervised Experiment on MVTec Dataset

| | mAUROC_sp_max | mAP_sp_max | mF1_max_sp_max | mAUROC_px | mAP_px | mF1_max_px | mAUPRO_px | |:--------------:|:-----------------:|:--------------:|:------------------:|:-------------:|:----------:|:--------------:|:-------------:| | carpet | 100.00 | 100.00 | 100.00 | 98.81 | 57.30 | 58.92 | 95.95 | | grid | 100.00 | 100.00 | 100.00 | 99.24 | 45.88 | 49.85 | 97.64 | | leather | 100.00 | 100.00 | 100.00 | 99.16 | 43.34 | 43.58 | 98.35 | | tile | 100.00 | 100.00 | 100.00 | 96.59 | 71.68 | 64.09 | 88.93 | | wood | 99.47 | 99.84 | 98.33 | 95.84 | 51.63 | 53.69 | 92.06 | | bottle | 100.00 | 100.00 | 100.00 | 98.95 | 80.95 | 77.05 | 96.65 | | cable | 99.57 | 99.74 | 97.83 | 97.51 | 61.10 | 62.99 | 92.05 | | capsule | 97.57 | 99.48 | 97.72 | 98.78 | 41.99 | 47.50 | 96.14 | | hazelnut | 100.00 | 100.00 | 100.00 | 99.11 | 65.72 | 65.56 | 95.94 | | metal_nut | 100.00 | 100.00 | 100.00 | 97.80 | 77.84 | 82.10 | 94.99 | | pill | 99.15 | 99.86 | 98.56 | 99.05 | 79.95 | 75.37 | 96.36 | | screw | 95.57 | 98.48 | 94.96 | 99.31 | 37.04 | 40.50 | 96.54 | | toothbrush | 96.67 | 98.73 | 95.24 | 98.99 | 52.60 | 58.28 | 91.37 | | transistor | 99.83 | 99.76 | 97.56 | 95.61 | 59.92 | 55.13 | 90.60 | | zipper | 99.55 | 99.88 | 98.74 | 98.45 | 56.23 | 60.51 | 95.21 | | average | 99.13 | 99.68 | 98.44 | 98.12 | 57.46 | 58.77 | 94.30 |

Table 2: Quantitative Results of Unsupervised Experiment on VisA Dataset

| | mAUROC_sp_max | mAP_sp_max | mF1_max_sp_max | mAUROC_px | mAP_px | mF1_max_px | mAUPRO_px | |:--------------:|:-----------------:|:--------------:|:------------------:|:-------------:|:----------:|:--------------:|:-------------:| | pcb1 | 96.07 | 94.43 | 93.78 | 99.77 | 86.12 | 78.36 | 95.83 | | pcb2 | 98.02 | 97.97 | 95.38 | 99.04 | 17.81 | 25.88 | 91.62 | | pcb3 | 97.33 | 97.28 | 92.23 | 99.27 | 31.25 | 30.53 | 93.76 | | pcb4 | 99.66 | 99.64 | 98.49 | 99.22 | 49.87 | 53.25 | 95.07 | | macaroni1 | 93.83 | 90.93 | 86.91 | 99.78 | 19.02 | 23.76 | 98.29 | | macaroni2 | 92.62 | 91.75 | 84.04 | 99.73 | 12.66 | 21.89 | 99.33 | | capsules | 93.80 | 96.33 | 91.08 | 99.28 | 63.91 | 60.77 | 95.72 | | candle | 98.06 | 98.03 | 94.63 | 99.42 | 23.62 | 34.59 | 96.92 | | cashew | 97.36 | 98.75 | 94.42 | 98.75 | 51.07 | 52.52 | 95.54 | | chewinggum | 99.43 | 99.68 | 98.99 | 99.14 | 57.44 | 56.03 | 90.52 | | fryum | 97.48 | 98.79 | 94.12 | 96.94 | 44.90 | 50.58 | 93.29 | | pipe_fryum | 99.12 | 99.56 | 97.00 | 99.08 | 52.60 | 56.90 | 95.37 | | average | 96.54 | 96.20 | 92.88 | 99.07 | 40.34 | 43.65 | 94.80 |

Table 3: Quantitative Results of semi10 Experiment on MVTec Dataset

CDO

Install / Use

README