U-Net Segmentation with Tensorboard

This is a simple implementation of the U-Net arhitecture and a project to utilize segmentation using Tensorboard.

A Korean translation of U-Net Paper: http://bit.ly/UNet_Paper_Translation

This tutorial depends on the following libraries:

• Tensorflow == 1.14
• Keras == 2.3.1

My computing resources are as follows:

• CPU: Intel i7-8700k
• GPU: GTX 1080ti
• RAM: 64GB

How to Run

python main.py

Tensorboard

tensorboard --logdir=./logs --host localhost

---

Overview

Data

The original dataset is from ISBI challenge, and I've downloaded it and done the pre-processing.

You can find it in folder data/membrane.

Model Architecture

Training Detail

1. Data Augmentation

10 times more images were used from the original number.

Method	Value
Rotation Range	0.2
Width Shift Range	0.05
Height Shift Range	0.05
Shear range	0.05
Zoom range	0.05
Horizontal Flip	True
Fill Mode	reflect

Augmentation images examples

2. Hyperparameters

• epochs : 50
• batch size : 5
• Learning rate : 0.0001

3. Optimizer, Loss function and Metric

• Adam
• Binary Cross Entropy
• Accuracy

Results

Model performance was approximately 91% accuracy for validation data when 50 epochs were trained.

Tensorboard

Reference

• Paper: U-Net: Convolutional Networks for Biomedical Image Segmentation
• Code: https://github.com/zhixuhao/unet