About

Helper package with multiple U-Net implementations in Keras as well as useful utility tools helpful when working with image segmentation tasks

Features:

• U-Net models implemented in Keras
  • Vanilla U-Net implementation based on the original paper
  • Customizable U-Net
  • U-Net optimized for satellite images based on DeepSense.AI Kaggle competition entry
• Utility functions:
  • Plotting images and masks with overlay
  • Plotting images masks and predictions with overlay (prediction on top of original image)
  • Plotting training history for metrics and losses
  • Cropping smaller patches out of bigger image (e.g. satellite imagery) using sliding window technique (also with overlap if needed)
  • Plotting smaller patches to visualize the cropped big image
  • Reconstructing smaller patches back to a big image
  • Data augmentation helper function
• Notebooks (examples):
  • Training custom U-Net for whale tails segmentation
  • Semantic segmentation for satellite images
  • Semantic segmentation for medical images ISBI challenge 2015

Installation:

pip install git+https://github.com/karolzak/keras-unet

or

pip install keras-unet

Usage examples:

• U-Net implementations in Keras:
  • Vanilla U-Net
  • Customizable U-Net
  • U-Net for satellite images
• Utils:
  • Plot training history
  • Plot images and segmentation masks
  • Get smaller patches/crops from bigger image
  • Plot small patches into single big image
  • Reconstruct a bigger image from smaller patches/crops

Vanilla U-Net

Model scheme can be viewed here

from keras_unet.models import vanilla_unet

model = vanilla_unet(input_shape=(512, 512, 3))

[back to usage examples]

Customizable U-Net

Model scheme can be viewed here

from keras_unet.models import custom_unet

model = custom_unet(
    input_shape=(512, 512, 3),
    use_batch_norm=False,
    num_classes=1,
    filters=64,
    dropout=0.2,
    output_activation='sigmoid')

[back to usage examples]

U-Net for satellite images

Model scheme can be viewed here

from keras_unet.models import satellite_unet

model = satellite_unet(input_shape=(512, 512, 3))

[back to usage examples]

Plot training history

history = model.fit_generator(...)

from keras_unet.utils import plot_segm_history

plot_segm_history(
    history, # required - keras training history object
    metrics=['iou', 'val_iou'], # optional - metrics names to plot
    losses=['loss', 'val_loss']) # optional - loss names to plot

Output:

[back to usage examples]

Plot images and segmentation masks

from keras_unet.utils import plot_imgs

plot_imgs(
    org_imgs=x_val, # required - original images
    mask_imgs=y_val, # required - ground truth masks
    pred_imgs=y_pred, # optional - predicted masks
    nm_img_to_plot=9) # optional - number of images to plot

Output:

[back to usage examples]

Get smaller patches/crops from bigger image

from PIL import Image
import numpy as np
from keras_unet.utils import get_patches

x = np.array(Image.open("../docs/sat_image_1.jpg"))
print("x shape: ", str(x.shape))

x_crops = get_patches(
    img_arr=x, # required - array of images to be cropped
    size=100, # default is 256
    stride=100) # default is 256

print("x_crops shape: ", str(x_crops.shape))

Output:

x shape:  (1000, 1000, 3)   
x_crops shape:  (100, 100, 100, 3)

[back to usage examples]

Plot small patches into single big image

from keras_unet.utils import plot_patches
   
print("x_crops shape: ", str(x_crops.shape))         
plot_patches(
    img_arr=x_crops, # required - array of cropped out images
    org_img_size=(1000, 1000), # required - original size of the image
    stride=100) # use only if stride is different from patch size

Output:

x_crops shape:  (100, 100, 100, 3)

[back to usage examples]

Reconstruct a bigger image from smaller patches/crops

import matplotlib.pyplot as plt
from keras_unet.utils import reconstruct_from_patches

print("x_crops shape: ", str(x_crops.shape))

x_reconstructed = reconstruct_from_patches(
    img_arr=x_crops, # required - array of cropped out images
    org_img_size=(1000, 1000), # required - original size of the image
    stride=100) # use only if stride is different from patch size

print("x_reconstructed shape: ", str(x_reconstructed.shape))

plt.figure(figsize=(10,10))
plt.imshow(x_reconstructed[0])
plt.show()

Output:

x_crops shape:  (100, 100, 100, 3)
x_reconstructed shape:  (1, 1000, 1000, 3)

[back to usage examples]