U-Net

• Tensorflow implementations of U-Net: Convolutional Networks for Biomedical Image Segmentation.
• U-Net 을 구현했습니다.
• 

Requirements

• Python v3.6
• tensorflow v1.4

How to use

• python3 main.py --todo [train or test or img_aug]
• tensorboard --logdir summary

Data Source

• Original data set can be downloaded from isbi challenge cite.
• Image Augmentation : Original data set 이 적어서 rotation, cropping 을 이용해 데이터를 생성했습니다.
  • Original data (30 * (512, 512)) - rotation - central cropping - random cropping -> new data (300 * (256, 256))

과적합 방지

• 데이터셋 300 -> 1500 : 정확도 90% -> 92%
• 배치정규화 : 정확도 92% -> 93%

Result

• 1000 epoch 학습한 후,
  • train data 로 한 결과
    • 
  • test data 로 한 결과
    • 

To be added

• Image Augmentation
• Batch Normalization
• Drop Out

References

• Studies
  • U-Net: Convolutional Networks for Biomedical Image Segmentation
• Github repositories
  • https://github.com/zhixuhao/unet