nagatoyuki0943/memseg-tootouch

Unofficial Re-implementation of MemSeg for Anomaly Detection

MemSeg

Unofficial Re-implementation for MemSeg: A semi-supervised method for image surface defect detection using differences and commonalities

Environments

• Docker image: nvcr.io/nvidia/pytorch:20.12-py3

anomalib
opencv
einops
timm
wandb

Process

1. Anomaly Simulation Strategy

• notebook
• Describable Textures Dataset(DTD) [ download ]

2. Model Process

• notebook

Train

python main.py --yaml_config ./configs/bottle.yaml

Train custom data

configs/custom.yaml

EXP_NAME: MemSeg
SEED: 42

DATASET:
  datadir: './datasets'
  texture_source_dir: './datasets/dtd/images'
  target: 'custom'
  resize: !!python/tuple
    - 256 # height
    - 256 # width
  structure_grid_size: 8
  transparency_range:
    - 0.15 # under bound
    - 1. # upper bound
  perlin_scale: 6
  min_perlin_scale: 0
  perlin_noise_threshold: 0.5

将自定义数据放到 datasets/custom文件夹下

格式如下

└─datasets
   └─custom
      ├─test
      │  ├─bad/  <- 这一个级别内存放图片
      │  ├─good/
      │  └─.../
      └─train
          ├─good/
          └─.../

train

python main.py --yaml_config ./configs/custom.yaml

export

Onnx is exported by default during training

You can export custom format by using export.py

Demo

voila "[demo] model inference.ipynb" --port ${port} --Voila.ip ${ip}

Results

TBD

	target	AUROC-image	AUROC-pixel	AUPRO-pixel
0	leather	100	93.93	90.44
1	wood	99.12	92.71	84.96
2	carpet	91.33	91.32	78.34
3	capsule	95.77	88.55	81.56
4	cable	92.41	81.77	64.45
5	metal_nut	99.9	71.13	79.92
6	tile	100	98.1	95.41
7	grid	96.57	76.78	59.63
8	bottle	99.92	95	89.95
9	zipper	97.58	93.76	83.94
10	transistor	97.71	71.78	66.86
11	hazelnut	95.29	91.73	87.83
12	pill	83.69	91.91	72.62
	Average	96.1	87.57	79.69

Citation

@article{DBLP:journals/corr/abs-2205-00908,
  author    = {Minghui Yang and
               Peng Wu and
               Jing Liu and
               Hui Feng},
  title     = {MemSeg: {A} semi-supervised method for image surface defect detection
               using differences and commonalities},
  journal   = {CoRR},
  volume    = {abs/2205.00908},
  year      = {2022},
  url       = {https://doi.org/10.48550/arXiv.2205.00908},
  doi       = {10.48550/arXiv.2205.00908},
  eprinttype = {arXiv},
  eprint    = {2205.00908},
  timestamp = {Tue, 03 May 2022 15:52:06 +0200},
  biburl    = {https://dblp.org/rec/journals/corr/abs-2205-00908.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}