def get_train_data():
images, labels = load_data()
imgs, lbls = [], []
for i, (img, lbl) in enumerate(zip(images, labels)):
if i in [2, 24, 41, 85, 75, 6, 7]:
continue
imgs += [img]
lbls += [lbl]
cropped_images1, cropped_masks1, _ = ImageCutter.image_and_mask_cutter(imgs, lbls,
window_h=256, window_w=256, step_x=50, step_y=50, classes_to_get=[40, 70, 80])
cropped_images2, cropped_masks2, _ = ImageCutter.image_and_mask_cutter(
imgs, lbls, window_h=256, window_w=256, step_x=40, step_y=40, classes_to_get=[70])
cropped_images3, cropped_masks3, _ = ImageCutter.image_and_mask_cutter(
imgs, lbls, window_h=256, window_w=256, step_x=60, step_y=60, classes_to_get=[60])
cropped_images4, cropped_masks4, _ = ImageCutter.image_and_mask_cutter(
imgs, lbls, window_h=256, window_w=256, step_x=60, step_y=60, classes_to_get=[10])
cropped_images = cropped_images1 + cropped_images2 + cropped_images3 + cropped_images4
cropped_masks = cropped_masks1 + cropped_masks2 + cropped_masks3 + cropped_masks4
return cropped_images, cropped_masks
def image_and_mask_cutter(images,
masks,
window_h,
window_w,
step_x,
step_y,
classes_to_get,
use_all_px=True):
"""
Crops `images` and `masks` using sliding window with resize.
Parameters
----------
images : list
List of input images.
masks : list
List of input masks.
window_h : int
Output image height.
window_w : int
Output image width.
step_x : int
Sliding window step by OX.
step_y : int
Sliding window step by OX.
scale_factor : float
Scale factor, must be in range (0, 1). After each 'sliding window step' the original images
are resized to (previous_width * scale_factor, previous_height * scale_factor).
postprocessing : func
Post processing function, using on cropped image (may be function what calculate num positives pixels).
use_all_px : bool
If True, all pixels of image would be in output lists.
Returns
-------
Three list:
1. cropped images
2. cropped masks
3. additional list (result of post processing)
"""
assert (len(images) > 0)
assert (len(images) == len(masks))
assert (window_h > 0 and window_w > 0 and step_x > 0 and step_y > 0)
cropped_images = []
cropped_masks = []
additional_list = []
dx = 0
dy = 0
for index, (img, mask) in enumerate(zip(images, masks)):
print(index)
assert (img.shape[:2] == mask.shape[:2])
current_height, current_width = img.shape[:2]
for dy in range(int((current_height - window_h) / step_y)):
for dx in range(int((current_width - window_w) / step_x)):
crop_img, crop_mask = ImageCutter.crop_img_and_mask(
img, mask, dy * step_y, dy * step_y + window_h,
dx * step_x, dx * step_x + window_w)
if ImageCutter.has_class(crop_mask, classes_to_get):
cropped_images.append(crop_img)
cropped_masks.append(crop_mask)
if use_all_px:
overlap_y = dy * step_y + window_h != current_height
overlap_x = dx * step_x + window_w != current_width
if overlap_y:
for dx in range(int((current_width - window_w) / step_x)):
crop_img, crop_mask = ImageCutter.crop_img_and_mask(
img, mask, current_height - window_h,
current_height, dx * step_x,
dx * step_x + window_w)
if ImageCutter.has_class(crop_mask, classes_to_get):
cropped_images.append(crop_img)
cropped_masks.append(crop_mask)
if overlap_x:
for dy in range(int((current_height - window_h) / step_y)):
crop_img, crop_mask = ImageCutter.crop_img_and_mask(
img, mask, dy * step_y, dy * step_y + window_h,
current_width - window_w, current_width)
if ImageCutter.has_class(crop_mask, classes_to_get):
cropped_images.append(crop_img)
cropped_masks.append(crop_mask)
if overlap_x and overlap_y:
crop_img, crop_mask = ImageCutter.crop_img_and_mask(
img, mask, current_height - window_h, current_height,
current_width - window_w, current_width)
if ImageCutter.has_class(crop_mask, classes_to_get):
cropped_images.append(crop_img)
cropped_masks.append(crop_mask)
return cropped_images, cropped_masks, additional_list
uniq = np.unique(label)
if i in [25, 41, 14, 19, 43, 2, 5]:
Xtest.append(images[i])
Ytest.append(label)
continue
else:
Xtrain.append(images[i])
Ytrain.append(label)
for image, label in zip(copy(Xtrain), copy(Ytrain)):
uniq = np.unique(label)
if 40 in uniq or 70 in uniq or 80 in uniq:
Ytrain += [label] * 10
Xtrain += [image] * 10
Xtrain, Ytrain, _ = ImageCutter.image_and_mask_cutter(Xtrain, Ytrain, 256, 256, 128, 128, 0.05)
Xtest, Ytest, _ = ImageCutter.image_and_mask_cutter(Xtest, Ytest, 256, 256, 128, 128, 0.05)
Xtrain = np.asarray(Xtrain).astype(np.float32) / 255
Xtrain = [i for i in Xtrain]
Ytrain = np.asarray(Ytrain).astype(np.uint8) // 10
Ytrain = [i for i in Ytrain]
Xtest = np.asarray(Xtest).astype(np.float32) / 255
Xtest = [i for i in Xtest]
Ytest = np.asarray(Ytest).astype(np.uint8) // 10
Ytest = [i for i in Ytest]
data = Data(Xtrain, Ytrain)
data = AffineAugment(num_matrices=1, noise_type='gaussian', keep_old_data=True)(data)
data = ElasticAugment(num_maps=1, border_mode='reflect_101', keep_old_data=True)(data)
Xtest = []
Ytest = []
Xtrain = []
Ytrain = []
for i, label in enumerate(labels):
uniq = np.unique(label)
if i in [7, 11, 75, 99]:
Xtest.append(images[i])
Ytest.append(label)
continue
else:
Xtrain.append(images[i])
Ytrain.append(label)
Xtrain, Ytrain, _ = ImageCutter.image_and_mask_cutter(
Xtrain, Ytrain, 384, 384, 128, 128, 0.005)
Xtest, Ytest, _ = ImageCutter.image_and_mask_cutter(
Xtest, Ytest, 384, 384, 128, 128, 0.005)
Xtrain = np.asarray(Xtrain).astype(np.float32) / 255
Xtrain = [i for i in Xtrain]
Ytrain = np.asarray(Ytrain).astype(np.uint8) // 10
Ytrain = [i for i in Ytrain]
Xtest = np.asarray(Xtest).astype(np.float32) / 255
Xtest = [i for i in Xtest]
Ytest = np.asarray(Ytest).astype(np.uint8) // 10
Ytest = [i for i in Ytest]
data = Data(Xtrain, Ytrain)
data = AffineAugment(num_matrices=5,
