A small code to create a new dataset with smaller patches for semantic segmentation dataset. Another code to read the data from the new dataset.

image_io.py

"""
This function can be used to read the images from the dataset created by the patchify_dataset.py.
Example Usage:
X_test, y_test = get_images("ModMonuSeg/Test/", (im_width, im_height), gt_extension='png')
"""
def get_images(parent_dir, im_shape, img_folder="TissueImages/", gt_folder="GroundTruth/", gt_extension=None):
  tissue_dir = parent_dir + img_folder
  gt_dir = parent_dir + gt_folder

  im_width, im_height = im_shape


  ids = next(os.walk(tissue_dir))[2]
  print("No. of images = ", len(ids))

  X = np.zeros((len(ids), im_height, im_width, 3), dtype=np.float32)
  y = np.zeros((len(ids), im_height, im_width, 1), dtype=np.float32)

  # tqdm is used to display the progress bar
  for n, id_ in tqdm_notebook(enumerate(ids), total=len(ids)):
    # Load images
    img = load_img(tissue_dir+id_)
    x_img = img_to_array(img)
    x_img = resize(x_img, (im_width, im_height, 3), mode = 'constant', preserve_range = True)
    # Load masks
    mask_id_ = id_.split('.')[0]
    mask_id_ = '-'.join(id_.split('-')[:-1]) + '_bin_mask-'+ id_.split('-')[-1]
    if gt_extension:
      mask_id_ = mask_id_.split('.')[0] + '.' + gt_extension
    mask = img_to_array(load_img(gt_dir+mask_id_, grayscale=True))
    mask = resize(mask, (im_width, im_height, 1), mode = 'constant', preserve_range = True)
    # Save images
    X[n] = x_img/255.0
    y[n] = mask/255.0

  return X, y

patchify_dataset.py

"""
Used to patchify the dataset of Semantic Segmentation
Set the directory for the current dataset, be sure to change the parameters.
Give new Dataset Dir, also specify the sub directories.
Patch size and Stride can be changed.
"""
import os
from skimage.io import imread, imsave
from skimage.util.shape import view_as_windows

# Parameters to change.
curr_dataset_dir = "MonuSeg/"
new_dataset_dir = "ModMonuSeg/"

sub_dirs = ["Training/TissueImages/", "Training/GroundTruth/",
             "Test/TissueImages/", "Test/GroundTruth/"]
patch_size = {'rgb': (256,256, 3), 'grayscale': (256,256)}
stride = {'rgb': (256//2, 256//2, 3), 'grayscale': (256//2, 256//2)}

# Create the new dataset directory
os.makedirs(new_dataset_dir, exist_ok=True)

for sub_dir in sub_dirs:
    new_dir_ = new_dataset_dir + sub_dir
    # Create sub_dir in the new dataset directory
    os.makedirs(new_dir_, exist_ok=True)
    dir_ = curr_dataset_dir + sub_dir
    ids = next(os.walk(dir_))[2]

    for id_ in ids:
        img = imread(dir_ + id_)
        img_type = 'rgb' if img.ndim == 3 else 'grayscale'
        new_imgs = view_as_windows(img, patch_size[img_type], stride[img_type])
        new_imgs = new_imgs.reshape(-1, 256, 256, 3) if img.ndim == 3 else new_imgs.reshape(-1, 256, 256)
        for i, patch in enumerate(new_imgs):
            file_name, extension = id_.split('.')
            new_file_name = file_name + f'-PCH{i}.' + extension
            imsave(new_dir_ + new_file_name, patch)