def test_throws_when_intensity_range_out_of_range():
with testing.raises(ValueError):
random_shapes((1000, 1234), max_shapes=1, multichannel=False,
intensity_range=(0, 256))
with testing.raises(ValueError):
random_shapes((2, 2), max_shapes=1,
intensity_range=((-1, 255),))
def gen_image(imgType, noise=True):
if imgType == 'grayscale':
img, _ = random_shapes((512, 512),
max_shapes=30,
intensity_range=(0, 255),
min_size=50,
multichannel=False,
max_size=np.random.randint(50, 150),
allow_overlap=False)
else:
img, _ = random_shapes((512, 512),
max_shapes=30,
intensity_range=(0, 255),
min_size=50,
multichannel=False,
max_size=np.random.randint(50, 150),
allow_overlap=False)
img[img == 255] = 0
mask = np.zeros(img.shape, dtype=np.uint8)
mask[:, :] = img
mask[mask > 0] = 255
if noise:
nsImg = random_noise(img)
return nsImg, mask
def test_generate_triangle_throws_when_size_too_small():
with testing.raises(ValueError):
random_shapes((128, 64),
max_shapes=1,
min_size=1,
max_size=1,
shape='triangle')
def test_throws_when_intensity_range_out_of_range():
with testing.raises(ValueError):
random_shapes((1000, 1234), max_shapes=1, multichannel=False,
intensity_range=(0, 256))
with testing.raises(ValueError):
random_shapes((2, 2), max_shapes=1,
intensity_range=((-1, 255),))
def test_generate_circle_throws_when_size_too_small():
with testing.raises(ValueError):
random_shapes((64, 128),
max_shapes=1,
min_size=1,
max_size=1,
shape='circle')
def test_throws_when_intensity_range_out_of_range():
with testing.raises(ValueError):
with expected_warnings(["`multichannel` is a deprecated argument"]):
random_shapes((1000, 1234),
max_shapes=1,
multichannel=False,
intensity_range=(0, 256))
with testing.raises(ValueError):
random_shapes((2, 2), max_shapes=1, intensity_range=((-1, 255), ))
def test_generates_gray_images_with_correct_shape_deprecated_multichannel():
with expected_warnings(["`multichannel` is a deprecated argument"]):
image, _ = random_shapes((4567, 123),
min_shapes=3,
max_shapes=20,
multichannel=False)
assert image.shape == (4567, 123)
# repeat prior test, but check for positional multichannel warning
with expected_warnings(["Providing the `multichannel` argument"]):
image, _ = random_shapes((4567, 123), 20, 3, 2, None, False)
assert image.shape == (4567, 123)
def random_shape_maze(width,
height,
max_shapes,
max_size,
allow_overlap,
shape=None,
seed=None):
x, _ = random_shapes([height, width],
max_shapes,
max_size=max_size,
multichannel=False,
shape=shape,
allow_overlap=allow_overlap,
random_seed=seed)
x[x == 255] = 0
x[np.nonzero(x)] = 1
# wall
x[0, :] = 1
x[-1, :] = 1
x[:, 0] = 1
x[:, -1] = 1
return np.asarray(x, dtype=np.float32)
def test_returns_empty_labels_and_white_image_when_cannot_fit_shape():
# The circle will never fit this.
with expected_warnings(['Could not fit']):
image, labels = random_shapes(
(10000, 10000), max_shapes=1, min_size=10000, shape='circle')
assert len(labels) == 0
assert (image == 255).all()
def WorldG():
image2, _ = random_shapes((128, 128),
shape='circle',
min_shapes=10,
max_shapes=15,
min_size=20,
max_size=25,
intensity_range=((100, 254), ))
image2 = np.where(image2 == 255, 0, image2)
x_1 = np.random.randint(0, 120)
y_1 = np.random.randint(0, 120)
x_2 = np.random.randint(0, 120)
y_2 = np.random.randint(0, 120)
while distance(x_1, x_2, y_1, y_2) < 50:
x_1 = np.random.randint(0, 120)
y_1 = np.random.randint(0, 120)
x_2 = np.random.randint(0, 120)
y_2 = np.random.randint(0, 120)
cv2.circle(image2, (x_1, y_1), 5, (255, 255, 255), thickness=cv2.FILLED)
cv2.circle(image2, (x_2, y_2), 5, (255, 255, 255), thickness=cv2.FILLED)
cv2.imwrite('test.png', image2)
#plt.imshow(image2)
#plt.title('my picture1')
#plt.show()
return image2
def test_generates_white_image_when_min_pixel_intensity_255():
image, labels = random_shapes((128, 128),
max_shapes=3,
min_pixel_intensity=255,
random_seed=42)
assert len(labels) > 0
assert (image == 255).all()
def flow_random_shapes(self, ):
while True:
images = np.empty(self.im_shape)
kspaces = np.empty(self.im_shape)
for i in range(self.batch_size):
image, _ = random_shapes(
(self.size, self.size),
max_shapes=self.n_shapes,
multichannel=False,
allow_overlap=True,
max_size=self.size,
)
image = image.astype('float32')
image /= 255
kspace = fft(image)
images[i, ..., 0] = image
kspaces[i, ..., 0] = kspace
mask = gen_mask(kspaces[0, ..., 0], accel_factor=self.af)
fourier_mask = np.repeat(mask.astype(np.float), self.size, axis=0)
mask_batch = np.repeat(fourier_mask[None, ...],
len(kspaces),
axis=0)[..., None]
kspaces *= mask_batch
mask_batch = mask_batch[..., 0]
yield (kspaces, mask_batch), images
def random_mask(H, W, mask_size=200):
"""
mask: ranges in [0,1]
"""
masked_image = np.zeros([H, W, 3])
mask = random_shapes(
[H, W],
max_shapes=1,
min_shapes=1,
max_size=mask_size,
min_size=mask_size / 2,
multichannel=False,
intensity_range=[0, 0],
)[0]
mask = np.stack((mask, mask, mask), axis=-1)
random_state = np.random.RandomState(None)
distortion_range = 50
alpha = np.random.rand() * 6
forward_dx = (
gaussian_filter((random_state.rand(H, W) * 2 - 1), distortion_range, mode="constant", cval=0) * alpha * 1000
)
forward_dy = (
gaussian_filter((random_state.rand(H, W) * 2 - 1), distortion_range, mode="constant", cval=0) * alpha * 1000
)
mask = forward_mapping(mask, forward_dy, forward_dx, maxIter=3, precision=1e-3) / 255
mask = 1 - gaussian(mask, sigma=0, preserve_range=True, multichannel=False, anti_aliasing=True)
mask = mask[:, :, 0]
return mask
def test_returns_empty_labels_and_white_image_when_cannot_fit_shape():
# The circle will never fit this.
with expected_warnings(['Could not fit']):
image, labels = random_shapes(
(10000, 10000), max_shapes=1, min_size=10000, shape='circle')
assert len(labels) == 0
assert (image == 255).all()
def generate_rect_mask(mesh, min_shapes=1, max_shapes=1, skimage_labels=False, multichannel=False, intensity_range=((1, 1)), units=False, shape='rectangle',**kargs):
image, labels = random_shapes(mesh.shape, min_shapes=min_shapes, shape='rectangle', max_shapes=max_shapes, multichannel=multichannel, intensity_range=intensity_range, **kargs)
idx = image == 255
image[idx] = 0
if not skimage_labels:
labels = convert_labels(labels, mesh, units=units)
return image, labels
def gen_one_image(h=64, w=64):
shapes = ('rectangle', 'triangle', 'circle')
n_images = 3 #np.random.randint(1, 9)
matrix = np.zeros((h, w, n_images))
image = np.zeros((h, w, 3))
mask = np.full((h, w), -1)
for i in range(n_images):
n_shape = np.random.randint(len(shapes))
tmp_image, _ = random_shapes(
(h, w),
max_shapes=1,
min_shapes=1,
shape=shapes[n_shape],
# multichannel=True, num_channels=3,
min_size=11,
max_size=22,
allow_overlap=True)
matrix[tmp_image[:, :, 0] < 255, i] = 1
mask[tmp_image[:, :, 0] < 255] = n_shape
tmp_image[tmp_image == 255] = 0
image[tmp_image[:, :, 0] > 0] = tmp_image[tmp_image[:, :, 0] > 0]
image /= 255
image[image == 0] = 1
return image, mask, matrix
def test_random_shapes_is_reproducible_with_seed():
random_seed = 42
labels = []
for _ in range(5):
_, l = random_shapes((128, 128), max_shapes=5, random_seed=random_seed)
labels.append(l)
assert all(other == labels[0] for other in labels[1:])
def _apply_shapes(self, data: np.ndarray) -> np.ndarray:
channels, height, width = get_image_dims(data)
# Wrap intensity_range in another tuple for color images. User might have provided nested tuple already though,
# so don't wrap those again.
intensity_range = self.intensity_range
if channels > 1 and not isinstance(intensity_range[0], tuple):
intensity_range = (intensity_range, )
shapes, _ = random_shapes(image_shape=(height, width),
max_shapes=self.max_shapes,
max_size=self.max_size,
num_channels=channels,
intensity_range=intensity_range,
allow_overlap=True)
alpha = np.random.uniform(self.transparency_range[0],
self.transparency_range[1])
# Convert shapes to range [0,1] if image is floating point
normalized_shapes = shapes / 255.0 if np.issubdtype(
data.dtype, np.floating) else shapes
blend = normalized_shapes * alpha + data * (1.0 - alpha)
# Combine images without whitening non-shape regions
overlay = np.where(shapes < 255, blend, data)
return overlay.astype(data.dtype)
def test_generates_white_image_when_intensity_range_255():
image, labels = random_shapes((128, 128),
max_shapes=3,
intensity_range=((255, 255), ),
random_seed=42)
assert len(labels) > 0
assert (image == 255).all()
def make_image():
i = np.random.randint(0,255,(200,256,4),dtype='uint8')
i[:,:,3] = 0
shape = random_shapes((200, 256), max_shapes=4,allow_overlap=True,
intensity_range=((100, 255)),num_channels=1)
shape = shape[0].reshape((200,256))
i[:,:,3][shape[:]<255] =255
return i
def generate_shapes(img_size):
img_array, labels = random_shapes(img_size,
min_shapes=2,
max_shapes=10,
multichannel=False,
allow_overlap=True)
img = Image.fromarray(img_array)
return img
def test_random_shapes_is_reproducible_with_seed():
random_seed = 42
labels = []
for _ in range(5):
_, label = random_shapes((128, 128), max_shapes=5,
random_seed=random_seed)
labels.append(label)
assert all(other == labels[0] for other in labels[1:])
def create_simple_object_detection_dataset(path,
n_samples=100,
n_objects_max=3,
n_objects_min=1,
size=(150, 150),
min_size=0.2):
(path / 'images').mkdir(parents=True, exist_ok=True)
(path / 'class_images').mkdir(parents=True, exist_ok=True)
min_dimension = size[0]
if (size[1] < size[0]):
min_dimension = size[1]
# create class labels
cname = ['red', 'green', 'blue']
color = [(255, 0, 0), (0, 255, 0), (0, 0, 255)]
for clr, name in zip(color, cname):
img_name = f'{name}'
img = np.ones((50, 50, 3), dtype=np.uint8)
draw_rectangle(img, start=(0, 0), dimensions=(50, 50), color=clr)
save_img(path / 'class_images', img_name, img)
type_shapes = ['rectangle', 'circle', 'ellipse']
# create images + annotations
annotations = {}
images = {}
for i in range(n_samples):
labels = []
bboxs = []
img_name = f'img_{i}'
image, shapes = random_shapes(size,
n_objects_max,
multichannel=True,
shape=type_shapes[randint(0, 2)],
min_shapes=n_objects_min,
min_size=min_size * min_dimension)
for shape in shapes:
shape_name = shape[0]
rr_0, rr_1 = shape[1][0]
cc_0, cc_1 = shape[1][1]
middle_x = int((rr_0 + rr_1) / 2)
middle_y = int((cc_0 + cc_1) / 2)
label = (image[middle_x, middle_y].tolist(), shape_name)
bbox = (int(cc_0), int(rr_0), int(cc_1), int(rr_1))
labels.append(label)
bboxs.append(bbox)
img_file = img_name + ".jpg"
images[img_file] = image
save_img(path / 'images', img_name, image)
annotations[img_file] = {'labels': labels, 'bboxs': bboxs}
save_img_annotations(path, annotations)
return (images, annotations)
def generateRandShapes(min_s, max_s, min_size, overlap, sizes=(1000, 1000)):
image, _ = random_shapes(sizes,
min_shapes=min_s,
max_shapes=max_s,
min_size=min_size,
allow_overlap=overlap,
num_trials=100)
return image
def __generate_random_shapes(self, min_shapes, max_shapes):
img, _ = random_shapes(self.shape,
max_shapes,
min_shapes=min_shapes,
multichannel=False,
allow_overlap=True,
random_seed=np.random.randint(2**32 - 1))
# Numpy random usage for random seed unifies random seed which can be set for repeatability
attempt = np.array(img != 255, dtype=bool)
return attempt, np.sum(attempt)
def test_can_generate_one_by_one_rectangle():
image, labels = random_shapes((50, 128),
max_shapes=1,
min_size=1,
max_size=1,
shape='rectangle',
min_pixel_intensity=1)
assert len(labels) == 1
_, bbox = labels[0]
crop = image[bbox[0][0]:bbox[0][1], bbox[1][0]:bbox[1][1]]
assert (crop < 255).sum() == 3 # rgb
def test_can_generate_one_by_one_rectangle():
image, labels = random_shapes((50, 128),
max_shapes=1,
min_size=1,
max_size=1,
shape='rectangle')
assert len(labels) == 1
_, bbox = labels[0]
crop = image[bbox[0][0]:bbox[0][1], bbox[1][0]:bbox[1][1]]
# rgb
assert (np.shape(crop) == (1, 1, 3) and np.any(crop >= 1)
and np.any(crop < 255))
def _create_label_data():
y_true, _ = random_shapes(image_shape=(200, 200),
min_shapes=20,
max_shapes=30,
min_size=10,
multichannel=False)
y_true[y_true == 255] = 0
y_true, _, _ = relabel_sequential(y_true)
y_pred = np.zeros_like(y_true)
y_pred[3:, 3:] = y_true[:-3, :-3]
return y_true, y_pred
def get_results(shapes, camera_images, num):
result = random_shapes((120, 1920),
intensity_range=(0, 1),
max_shapes=1,
min_size=99,
max_size=100,
shape=shapes,
multichannel=False)
result_pad = np.pad(result[0], ((480, 480), (0, 0)),
mode='constant',
constant_values=255)
camera_images.append(result_pad)
append_single(camera_images, int(num - 1))
def generate_shape():
for i in range (500):
image_name = "shape_" + str(i) + ".png"
#create shape image with color
shape, _ = random_shapes((256,256), min_shapes=5, max_shapes=10, min_size=20, allow_overlap=True)
shape_img = Image.fromarray(shape)
shape_img.save("shape_color/"+image_name,"PNG")
#create mask
mask_img = shape_img.convert("L")
mask_array = np.asarray(mask_img)
mask_array = np.where(mask_array < 255, 0,255)
mask_img = Image.fromarray(mask_array.astype(np.uint8))
mask_img.save("shape_mask/"+image_name,"PNG")
def test_can_generate_one_by_one_rectangle():
image, labels = random_shapes(
(50, 128),
max_shapes=1,
min_size=1,
max_size=1,
shape='rectangle')
assert len(labels) == 1
_, bbox = labels[0]
crop = image[bbox[0][0]:bbox[0][1], bbox[1][0]:bbox[1][1]]
# rgb
assert (np.shape(crop) == (1, 1, 3) and np.any(crop >= 1)
and np.any(crop < 255))
def generate_level(self):
level_setup = random_shapes(
image_shape=settings.BRICKS_DIMENSION,
min_shapes=10,
max_shapes=20,
min_size=3,
max_size=10,
allow_overlap=True,
intensity_range=((30, 255),))[0]
for x in range(settings.BRICKS_DIMENSION[0]):
for y in range(settings.BRICKS_DIMENSION[1]):
if all(color != 255 for color in level_setup[x][y]):
position = pygame.Vector2(board_left() + settings.BRICKS_POSITION[0] + x * settings.BRICK_SIZE[0],
board_top() + settings.BRICKS_POSITION[1] + y * settings.BRICK_SIZE[1])
self.bricks.add(Brick(self.screen, position, level_setup[x][y]))
def test_generated_shape_for_channel_axis(channel_axis):
shape = (128, 64)
num_channels = 5
image, _ = random_shapes(shape,
num_channels=num_channels,
min_shapes=3,
max_shapes=10,
channel_axis=channel_axis)
if channel_axis is None:
expected_shape = shape
else:
expected_shape = tuple(np.insert(shape, channel_axis, num_channels))
assert image.shape == expected_shape
def random_mask(image, mask, num_mask=20, min_size=5, max_size=128):
"""
"""
raw, _ = random_shapes(image.shape[:2],
num_mask,
min_shapes=2,
min_size=min_size,
max_size=max_size,
multichannel=False,
allow_overlap=True)
mask_raw = 1 - raw / 255.
noise = np.random.random(image.shape)
image[mask_raw > 0] = noise[mask_raw > 0] * random.randint(0, 255)
mask = mask * (mask_raw == 0)
return image, mask
def test_generates_correct_bounding_boxes_for_triangles():
image, labels = random_shapes(
(128, 128),
max_shapes=1,
shape='triangle',
random_seed=42)
assert len(labels) == 1
label, bbox = labels[0]
assert label == 'triangle', label
crop = image[bbox[0][0]:bbox[0][1], bbox[1][0]:bbox[1][1]]
# The crop is filled.
assert (crop >= 0).any() and (crop < 255).any()
# The crop is complete.
image[bbox[0][0]:bbox[0][1], bbox[1][0]:bbox[1][1]] = 255
assert (image == 255).all()
def test_generate_triangle_throws_when_size_too_small():
with testing.raises(ValueError):
random_shapes(
(128, 64), max_shapes=1, min_size=1, max_size=1, shape='triangle')
def test_generate_circle_throws_when_size_too_small():
with testing.raises(ValueError):
random_shapes(
(64, 128), max_shapes=1, min_size=1, max_size=1, shape='circle')
def test_generates_color_images_with_correct_shape():
image, _ = random_shapes((128, 128), max_shapes=10)
assert image.shape == (128, 128, 3)
def test_generates_white_image_when_intensity_range_255():
image, labels = random_shapes((128, 128), max_shapes=3,
intensity_range=((255, 255),),
random_seed=42)
assert len(labels) > 0
assert (image == 255).all()
def test_generates_gray_images_with_correct_shape():
image, _ = random_shapes(
(4567, 123), min_shapes=3, max_shapes=20, multichannel=False)
assert image.shape == (4567, 123)
"""
=============
Random Shapes
=============
Example of generating random shapes with particular properties.
"""
import matplotlib.pyplot as plt
from skimage.draw import random_shapes
# Let's start simple and generate a 128x128 image
# with a single grayscale rectangle.
result = random_shapes((128, 128), max_shapes=1, shape='rectangle',
multichannel=False)
# We get back a tuple consisting of (1) the image with the generated shapes
# and (2) a list of label tuples with the kind of shape (e.g. circle,
# rectangle) and ((r0, r1), (c0, c1)) coordinates.
image, labels = result
print('Image shape: {}\nLabels: {}'.format(image.shape, labels))
# We can visualize the images.
fig, axes = plt.subplots(nrows=2, ncols=3)
ax = axes.ravel()
ax[0].imshow(image, cmap='gray')
ax[0].set_title('Grayscale shape')
# The generated images can be much more complex. For example, let's try many
# shapes of any color. If we want the colors to be particularly light, we can