/computer-vision-anomaly-detection

Computer Vision 이상치 탐지 알고리즘 경진대회

Primary LanguageJupyter Notebook

Computer Vision 이상치 탐지 알고리즘 경진대회

https://dacon.io/competitions/official/235894/overview/description

1. Goal

1.1 Problem

Classification on MVtec AD Dataset

1.2 Metric

Macro f1 score

2. Related work

2.1 Baseline: 0.666

baseline.ipynb

  1. Preprocessing
    1. Resize to (512 x 512)
    2. Data augmentation: flip left-right or up-down
  2. Training
    1. Model: efficientnet_b0
    2. Optimizer: Adam (lr=1e-3)
    3. Loss: CEE
    4. Gradient scaler
  3. Evaluation
    1. Training metric: 0.96759
    2. Test metric: 0.66579

2.2 SOTA baseline

  1. FastFlow (Detection AUROC: 99.4)
    1. Rank 1
    2. Unsupervised anomaly detection

3. Proposed idea

3.1 Proposed 1: 0.614

proposed1.ipynb

  1. Early stopping 적용

3.2 Proposed 2: 0.701

proposed2.ipynb

  1. TensorFlow porting
  2. Validation
    1. validation_on: optimal epochs 결정 -> 더 좋은 성능
    2. validation_off: train_full로 학습

3.3 Proposed 3: 0.745

proposed3.ipynb

  1. Sample weight 적용

3.4 Proposed 4: 0.644

proposed4.ipynb

  1. Model: efficientnet_b0
  2. 2-level classification
    1. Classification(class)
      model1: supervised
    2. Classification(label)
      model2: supervised

3.5 Proposed 5:

proposed5.ipynb

  1. 3-level classification
    1. Classification(class)
      model1: supervised (efficientnet_b0)
    2. Anomaly detection
      model2: unsupervised (PatchCore)
    3. Classification(label)
      model3: supervised (efficientnet_b0)