-
Notifications
You must be signed in to change notification settings - Fork 0
Unet
Biomedical λΆμΌμμ Image Segmentation μ λͺ©μ μΌλ‘ μ μλ End-to-End λ°©μμ Fully-Convolutional Network κΈ°λ° λͺ¨λΈ
- Architecture
- Patch νμ λ°©μ
- Training Method
-
λμ λ²μμ μ΄λ―Έμ§ ν½μ λ‘λΆν° μλ―Έ μ 보λ₯Ό μΆμΆνκ³ , μλ―Έμ 보λ₯Ό κΈ°λ°μΌλ‘ κ° ν½μ λ§λ€ κ°μ²΄λ₯Ό λΆλ₯νλ U λͺ¨μμ μν€ν μ² λ₯Ό μ μ
-
Model Structure μ ν¬κ² 3κ°μ§λ‘ ꡬμ±λμ΄μλ€
- μμΆ κ²½λ‘ (Contracting Path) - μ’μΈ‘
- μ μ§μ μΌλ‘ λμ λ²μμ μ΄λ―Έμ§ ν½μ μ 보며 (Context Information)μ μΆμΆνλ κ²½λ‘
- νμ₯ κ²½λ‘ (Expanding Path) - μ°μΈ‘
- μλ―Έμ 보λ₯Ό ν½μ μμΉμ 보μ κ²°ν©(Localization)νμ¬ κ° ν½μ λ§λ€ μ΄λ€ κ°μ²΄μ μνλμ§λ₯Ό ꡬλΆνλ κ²½λ‘
- μ ν κ΅¬κ° (Bottle Neck) - μ€μ νλ¨
- μμΆ κ²½λ‘ (Contracting Path) μμ νμ₯ κ²½λ‘ (Expanding Path) λ‘ μ νλλ λΆλΆ
- μμΆ κ²½λ‘ (Contracting Path) - μ’μΈ‘
-
context (Semantic Information) μΈμκ³Ό localization κ°μ trade-off λ₯Ό 보μ
- Contracting path μμλ μ΄λ―Έμ§μ Feature map (context μ 보)μ μΆμΆνκ³ , Expanding Path λ Feature map μ λ€μ Upsampling νμ¬ μ νν localization λ₯Ό μννλ€
- Expanding Path κ³Όμ μ§ν μμ, Contracting Path μμ μΆμΆλ Feature map (context μ 보)μ κ°μ Έμμ λΆμ°©νμ¬ Contracting Path μμ μμ±λ Context μ 보μ Expanding Path μ μμΉμ 보λ₯Ό κ²°ν©νλ λ°©μμΌλ‘ trade-off λ₯Ό 보μνλ€
- U-net μ μ λ°μ μΈ μ΄λ―Έμ§λ₯Ό μ΄λν΄κ°λ©° μΌμΌμ΄ νμΈνλ Sliding Window λ°©μμ μ¬μ©νμ§ μλλ€
- Sliding Window λ°©μμ μ¬μ©νλ©΄, μ΄λ―Έ μ¬μ©ν window ꡬμμ λ€μ Sliding Window μμ λ€μ κ²μ¦νκ² λλ€
- μκ°κ³Ό μ°μ° μΈ‘λ©΄μμ λλΉκ° λ°μ
- Overlap tile Method (Strategy)
- U-net μ padding μ μ μ©νμ§ μκ³ Convolution μ°μ°μ μννκΈ°μ,
μΆλ ₯ ν¬κΈ°κ° μ
λ ₯ ν¬κΈ° λ³΄λ€ νμ μλ€
- μΆλ ₯ ν¬κΈ°μ μ λ ₯ ν¬κΈ°λ₯Ό λμΌνκ² μ‘°μ νλ €λ©΄ zero padding κ³Ό κ°μ λ°©μμΌλ‘ μ λ ₯μ κ°μΌ μ΄ν convolution μ°μ°μ μνν΄μΌν¨
- U-net μ μ΄ λμ Overlap tile method λ₯Ό μ¬μ©ν¨
- μ¬μ΄μ¦κ° μ€μ΄λ€κ²λλ©΄μ μ±μμ§μ§μμ κ°μ₯μ리μ missing value λΆλΆμ λ―Έλ¬λ§ κΈ°λ²μ ν΅ν΄ κ²½κ³λΆλΆμ μ΄λ―Έμ§λ₯Ό λμΉμμΌμ κ°μ μ±μ
- μ΄ κΈ°λ²μ ν΅ν΄ μ 보 μμ€μ μ΅μννλ©΄μ context μ 보λ₯Ό μ μ₯ν μ μμλ€κ³ μ μλ μκ°
- Elastic Deformation
-
λ Όλ¬Έμμ μ¬μ©ν Data Augmentation λ°©λ²
- Data Augmentation μμ μ£Όλ‘ μ¬μ©νλ μ νλ³νμ νλ₯ μ μΌλ‘ noise λ₯Ό κ°ν΄μ£Όλ λ°©λ²
-
μΌλ°μ μΈ Data Augmentation (De-colorized, rotate, flip) μΌλ‘ μ νλ³νμ μ·¨νλ©΄ μ΄λ―Έμ§κ° μΌμ νκ² λ³νλ κ·μΉμ±μ λκ²λλ€
-
νλ₯ μ μΌλ‘ noise λ₯Ό κ°ν΄μ£Όλ Elastic Deformation μ μ¬μ©νλ©΄ λ μμ°μ€λ½κ³ νμ€μ κ°κΉμ΄ Data λ₯Ό Augmentation ν μ μκ² λλ€
- λ Biomedical λΆμΌμ μ ν©ν λ°μ΄ν°λ₯Ό μμ±ν μ μκ²λ¨
- Biomedical λΆμΌμμ λ€λ£¨λ μΈν¬ λ±μ μ΄λ―Έμ§ λ°μ΄ν°λ μ΄μμλ μλͺ 체μ΄κΈ°μ λͺ¨μμ΄ λλΆλΆ λ€λ₯΄λ€
- μκ°μ μΈ λͺ¨μμ λ³νλν Elastic Deformation μ΄ μ ννν μ μλ€κ³ ν¨
Reference
-
https://arxiv.org/pdf/1505.04597.pdf
- Original paper
- https://www.youtube.com/watch?v=O_7mR4H9WLk
- Paper Review Calls 011 - U-Net: Convolutional Networks for Biomedical Image Segmentation
- https://joungheekim.github.io/2020/09/28/paper-review/
- https://medium.com/@msmapark2/u-net-%EB%85%BC%EB%AC%B8-%EB%A6%AC%EB%B7%B0-u-net-convolutional-networks-for-biomedical-image-segmentation-456d6901b28a
- https://velog.io/@guide333/U-Net-%EC%A0%95%EB%A6%AC
- https://lmb.informatik.uni-freiburg.de/people/ronneber/u-net/
- https://kuklife.tistory.com/119
- https://m.blog.naver.com/9709193/221979612209
- https://en.wikipedia.org/wiki/U-Net
Human Pose Estimation
CNN Visualization
Image Generation
Multi-modal Learning