Dog Cat Segmentation

This work has been done thanks to the U-Net: Convolutional Networks for Biomedical Image Segmentation (Ronneberger et al. 2015)

Requirements

• Python 3.8
• TensorFlow 2.2.0
• mathplotlib
• opencv
• numpy

The required libraries can also be installed with pip:

pip install -r requirements.txt

The dataset used for traning can be downloaded with:

./dataset.sh

Training

To train the model run the following command:

python train.py --lr float --epochs int \
                --train_steps int --val_steps int \ 
                --checkpoint_dir path --weight_path path

where:

• --lr: Learning rate. Default=0.001
• --epochs: Number of training epochs. Default=20
• --train_steps: Number of training steps per epoch. Default=100
• --val_steps: Number of validation steps. Default=50
• --checkpoint_dir: Path to write checkpoint. Default="./model/checkpoint"
• --weight_path: path to save model weight. Default="./model/weights/unet.h5"

Results

• Images from validation set:

Original Image	True Mask	Predicted Mask

• Random images from the internet:

Original Image	Predicted Mask