FUSegNet / data /utils /runtime_patch.py
introvoyz041's picture
Migrated from GitHub
36ca0d9 verified
Raw
History Blame Contribute Delete
5.2 kB
from jenti.patch import Patch
import numpy as np
import random
def choose_fg_idx(
patch_mask, # an nd-array with size patch_H x patch_W or patch_H x patch_W x patch_Ch
fg_idx:list, # list of foreground indices. e.g. [2, 4, 7, 9]
MAX_ROI:bool=True, # if true and the returned patch is a foreground patch, then it
# returns the patch that has maximum info or region of interest (roi)
):
"""
It is a helper function that picks a foreground index. If MAX_ROI is True,
then it returns the index of that patch that has max info or roi in it. Otherwise,
it returns a randomly chosen foreground index.
Return
--------
It returns a foreground index.
"""
if MAX_ROI: # pick the index of the foreground patch that has maximum roi
max_nonzeros = 0 # Maximum no. of nonzeros. Initially set it to 0.
final_fg_idx: int # index of the patch that has maximum roi
for idx in fg_idx:
x = patch_mask[idx] # get the foreground patch mask
n_nonzero = np.count_nonzero(x) # no. of nonzeros in the patch mask
# Compare with current max no. of nonzeros
if n_nonzero > max_nonzeros:
final_fg_idx = idx # update index if new count is higher than the previous count
max_nonzeros = n_nonzero # update max no. of nonzeros
return final_fg_idx
else: # randomly pick a foreground index
return random.choice(fg_idx)
def runtime_patch(
image, # an nd-array with size H x W or H x W x Ch
mask, # an nd-array with size H x W or H x W x Ch
patch_shape:tuple=(256,256), # patch size
overlap:tuple=(0,0), # overlap between adjacent patches
FG_PROB:float=0.9, # probability of choosing a foreground
MAX_ROI:bool=True, # if true and the returned patch is a foreground patch, then it
# returns the patch that has maximum info or region of interest (roi)
):
"""
This function returns an image patch and the corresponding mask patch. The patch
can be a background patch or a foreground patch.
foreground patch: It contains information or region of interest (roi)
background patch: It does not contain any info or roi
Return
--------
It returns an image patch and the corresponding mask patch.
Size of image/mask patch: (patch_H, patch_W, ch) or (patch_H, patch_W)
"""
patch = Patch(patch_shape, overlap, patch_name='patch2d', csv_output=False)
patch_img, _, _ = patch.patch2d(image)
patch_mask, _, _ = patch.patch2d(mask)
# Separate foreground (fg) and background
fg_idx, bg_idx = [], []
for i,x in enumerate(patch_mask):
if np.sum(x) > 0: fg_idx.append(i) # fg
else: bg_idx.append(i) # background
# If no foreground, then randomly return a background patch
if len(fg_idx) == 0:
# Randomly choose a bg index
final_bg_idx = random.choice(bg_idx)
return patch_img[final_bg_idx], patch_mask[final_bg_idx]
# If no background, then randomly return a foreground patch
if len(bg_idx) == 0:
final_fg_idx = choose_fg_idx(patch_mask, fg_idx, MAX_ROI)
return patch_img[final_fg_idx], patch_mask[final_fg_idx]
# Choose foreground or background based on a probability distribution
fg_flag: bool
fg_flag = True if np.random.uniform(low=0, high=1, size=1) <= FG_PROB else False
if fg_flag: # pick a foreground patch
final_fg_idx = choose_fg_idx(patch_mask, fg_idx, MAX_ROI)
return patch_img[final_fg_idx], patch_mask[final_fg_idx]
else: # pick a background
final_bg_idx = random.choice(bg_idx)
return patch_img[final_bg_idx], patch_mask[final_bg_idx]
# =============================================================================
# # Example
#
# import cv2
# import matplotlib.pyplot as plt
# import os
# import random
#
# # Parameters
# FG_PROB = 0.9 # probability of selecting a foreground image
# MAX_ROI = True # select the patch that has maximum roi
#
# # Directory
# img_dir = r'.\dataset\test\images'
# mask_dir = r'.\dataset\test\labels'
#
# # List of images
# names = os.listdir(img_dir)
#
# name = random.choice(names)
#
# # Read image
# image = cv2.imread(os.path.join(img_dir, name))[:,:,::-1]
# mask = cv2.imread(os.path.join(mask_dir, name), 0)
#
# mask = np.expand_dims(mask, axis=-1)
#
# # Create patches
# patch_shape = [256, 256]
# overlap = [10,10] # overlap between two adjacent patches along both axes
#
# ip, mp = runtime_patch(
# image,
# mask,
# patch_shape=(256,256),
# overlap=(0,0),
# FG_PROB=0.9,
# MAX_ROI=True)
#
#
#
# fig, ax = plt.subplots(2,1, figsize=(15,7))
# ax[0].imshow(ip)
# ax[1].imshow(mp, cmap='gray')
#
# =============================================================================