def _get_rand_transform_matrix(self, image_size, d, batch_size):
M = np.zeros((batch_size, 8))
for i in range(batch_size):
tl_top = random.uniform(-d, d) # Top left corner, top
tl_left = random.uniform(-d, d) # Top left corner, left
bl_bottom = random.uniform(-d, d) # Bot left corner, bot
bl_left = random.uniform(-d, d) # Bot left corner, left
tr_top = random.uniform(-d, d) # Top right corner, top
tr_right = random.uniform(-d, d) # Top right corner, right
br_bottom = random.uniform(-d, d) # Bot right corner, bot
br_right = random.uniform(-d, d) # Bot right corner, right
transform = ProjectiveTransform()
if i == 0:
transform.estimate(
np.array(((0, 0), (0, image_size),
(image_size, image_size), (image_size, 0))),
np.array(((0, 0), (0, image_size),
(image_size, image_size), (image_size, 0))))
else:
transform.estimate(
np.array(
((tl_left, tl_top), (bl_left, image_size - bl_bottom),
(image_size - br_right, image_size - br_bottom),
(image_size - tr_right, tr_top))),
np.array(((0, 0), (0, image_size),
(image_size, image_size), (image_size, 0))))
M[i] = transform.params.flatten()[:8]
return M
def apply_projection_transform(self, Xb, batch_size, image_size):
d = image_size * 0.3 * self.intensity
for i in np.random.choice(batch_size,
int(batch_size * self.p),
replace=False):
tl_top = random.uniform(-d, d)
tl_left = random.uniform(-d, d)
bl_bottom = random.uniform(-d, d)
bl_left = random.uniform(-d, d)
tr_top = random.uniform(-d, d)
tr_right = random.uniform(-d, d)
br_bottom = random.uniform(-d, d)
br_right = random.uniform(-d, d)
transform = ProjectiveTransform()
transform.estimate(
np.array(((tl_left, tl_top), (bl_left, image_size - bl_bottom),
(image_size - br_right, image_size - br_bottom),
(image_size - tr_right, tr_top))),
np.array(((0, 0), (0, image_size), (image_size, image_size),
(image_size, 0))))
Xb[i] = warp(Xb[i],
transform,
output_shape=(image_size, image_size),
order=1,
mode='edge')
return Xb
def apply_projection_transform(X, intensity=0.75, depth=1):
no_samples=X.shape[0]
image_size=X.shape[1]
no_channels=X.shape[3]
d = image_size * 0.3 * intensity
indices_project = np.random.choice(
X.shape[0], math.ceil(X.shape[0]*depth*0.5), replace=False)
X_=[]
for i in indices_project:
tl_top = uniform(-d, d) # Top left corner, top margin
tl_left = uniform(-d, d) # Top left corner, left margin
bl_bottom = uniform(-d, d) # Bottom left corner, bottom margin
bl_left = uniform(-d, d) # Bottom left corner, left margin
tr_top = uniform(-d, d) # Top right corner, top margin
tr_right =uniform(-d, d) # Top right corner, right margin
br_bottom =uniform(-d, d) # Bottom right corner, bottom margin
br_right = uniform(-d, d) # Bottom right corner, right margin
transform = ProjectiveTransform()
transform.estimate(np.array((
(tl_left, tl_top),
(bl_left, image_size - bl_bottom),
(image_size - br_right, image_size - br_bottom),
(image_size - tr_right, tr_top)
)), np.array((
(0, 0),
(0, image_size),
(image_size, image_size),
(image_size, 0)
)))
X_.append(warp(X[i], transform, output_shape=(image_size, image_size), order = 1, mode = 'edge'))
X_.append(warp(X[i], transform.inverse, output_shape=(image_size, image_size), order = 1, mode = 'edge'))
return np.asarray(X_)
def apply_warp_augmentation_on_an_image(self, image):
image_size = image.shape[0]
#aff_tform = AffineTransform(scale=(1, 1/1.2), rotation=1, shear=0.7, translation=(210, 50))
aff_tform = AffineTransform(scale=(1, 1 / 1.2))
image = warp(image,
aff_tform,
output_shape=(image_size, image_size),
order=1,
mode='edge')
x = image_size * 0.2
# top_right - x,y
# bottom_left - x,y
# top_right - x,y
# bottom_right - x,y
tl_x, tl_y, bl_x, bl_y, tr_x, tr_y, br_x, br_y = np.random.uniform(
-x, x, size=8)
src = np.array([[tl_y, tl_x], [bl_y, image_size - bl_x],
[image_size - br_y, image_size - br_x],
[image_size - tr_y, tr_x]])
dst = np.array([[0, 0], [0, image_size], [image_size, image_size],
[image_size, 0]])
proj_tform = ProjectiveTransform()
proj_tform.estimate(src, dst)
image = warp(image,
proj_tform,
output_shape=(image_size, image_size),
order=1,
mode='edge')
return image
def apply_projection_transform(X, intensity):
image_size = X.shape[1]
d = image_size * 0.3 * intensity
tl_top = random.uniform(-d, d)
tl_left = random.uniform(-d, d)
bl_bottom = random.uniform(-d, d)
bl_left = random.uniform(-d, d)
tr_top = random.uniform(-d, d)
tr_right = random.uniform(-d, d)
br_bottom = random.uniform(-d, d)
br_right = random.uniform(-d, d)
transform = ProjectiveTransform()
transform.estimate(
np.array(((tl_left, tl_top), (bl_left, image_size - bl_bottom),
(image_size - br_right,
image_size - br_bottom), (image_size - tr_right, tr_top))),
np.array(((0, 0), (0, image_size), (image_size, image_size),
(image_size, 0))))
X = warp(X,
transform,
output_shape=(image_size, image_size),
order=1,
mode='edge')
return X
def projection_transform(image, max_warp=0.8, height=32, width=32):
#Warp Location
d = height * 0.3 * np.random.uniform(0, max_warp)
#Warp co-ordinates
tl_top = np.random.uniform(-d, d) # Top left corner, top margin
tl_left = np.random.uniform(-d, d) # Top left corner, left margin
bl_bottom = np.random.uniform(-d, d) # Bottom left corner, bottom margin
bl_left = np.random.uniform(-d, d) # Bottom left corner, left margin
tr_top = np.random.uniform(-d, d) # Top right corner, top margin
tr_right = np.random.uniform(-d, d) # Top right corner, right margin
br_bottom = np.random.uniform(-d, d) # Bottom right corner, bottom margin
br_right = np.random.uniform(-d, d) # Bottom right corner, right margin
##Apply Projection
transform = ProjectiveTransform()
transform.estimate(
np.array(((tl_left, tl_top), (bl_left, height - bl_bottom),
(height - br_right, height - br_bottom), (height - tr_right,
tr_top))),
np.array(((0, 0), (0, height), (height, height), (height, 0))))
output_image = warp(image,
transform,
output_shape=(height, width),
order=1,
mode='edge')
return output_image
def coord(contours, r, g, b, box_img):
t = ProjectiveTransform() # Initiate Projective Transform
# Use the Image and World Reference Points to Generate Source and Destination NumPy arrays for the transform
src = np.asarray([i_bl, i_tl, i_tr, i_br])
dst = np.asarray([w_bl, w_tl, w_tr, w_br])
t.estimate(src, dst) # Prepare the transform model
i_nos = 0 # Count number of objects of the color
objects = [] # Start with an empty array for image coordinates
for k_coord in contours: # Check each contour for a valid object
y_k, x_k, w_k, h_k = cv2.boundingRect(k_coord)
# Bound the individual contour with a rectangle
if w_k > 40 and h_k > 40: # Only select objects bigger than a threshold - to filter out disturbances
cv2.rectangle(box_img, (y_k, x_k), (y_k + w_k, x_k + h_k),
(b, g, r), 2)
# Draw a box on the image - to be subsequently printed on screen
objects.append([x_k + h_k / 2, y_k + w_k / 2])
# Add the object to the array (image coordinates)
i_nos += 1 # Increase the count of number of objects
if i_nos == 0: # If there are no valid objects in the image
world = [] # return empty array for world coordinates
else: # If there are valid objects in the image
world = t(
objects
) # Carry out Projective Transform to populate the world coordinates array
return i_nos, objects, world # return the no. of objects, image coordinates and world coordinates
def randomPerspective(im):
'''
wrapper of Projective (or perspective) transform, from 4 random points selected from 4 corners of the image within a defined region.
'''
region = 1 / 4
A = pl.array([[0, 0], [0, im.shape[0]], [im.shape[1], im.shape[0]],
[im.shape[1], 0]])
B = pl.array([
[
int(randRange(0, im.shape[1] * region)),
int(randRange(0, im.shape[0] * region))
],
[
int(randRange(0, im.shape[1] * region)),
int(randRange(im.shape[0] * (1 - region), im.shape[0]))
],
[
int(randRange(im.shape[1] * (1 - region), im.shape[1])),
int(randRange(im.shape[0] * (1 - region), im.shape[0]))
],
[
int(randRange(im.shape[1] * (1 - region), im.shape[1])),
int(randRange(0, im.shape[0] * region))
],
])
pt = ProjectiveTransform()
pt.estimate(A, B)
return warp(im, pt, output_shape=im.shape[:2])
def random_transform(cls, img):
image_size = img.shape[0]
d = image_size * 0.2
tl_top, tl_left, bl_bottom, bl_left, tr_top, tr_right, br_bottom, br_right = np.random.uniform(
-d, d, size=8) # Bottom right corner, right margin
aft = AffineTransform(scale=(1, 1 / 1.2))
img = warp(img,
aft,
output_shape=(image_size, image_size),
order=1,
mode='edge')
transform = ProjectiveTransform()
transform.estimate(
np.array(((tl_left, tl_top), (bl_left, image_size - bl_bottom),
(image_size - br_right, image_size - br_bottom),
(image_size - tr_right, tr_top))),
np.array(((0, 0), (0, image_size), (image_size, image_size),
(image_size, 0))))
img = warp(img,
transform,
output_shape=(image_size, image_size),
order=1,
mode='edge')
return img
def get_sprites(image, ctrs, debug=False):
"""
This function computes a projective transform from the source (mnist image) to
the destination (contour) and extracts the warped sprite.
"""
# We make sure that we work on a local copy of the image
img = image.copy()
# We loop through the sprites
sprts = []
for contour in ctrs:
# We compute the projective transform
source_points = np.array([[28, 28], [0, 28], [0, 0], [28, 0]])
destination_points = np.array(contour)
transform = ProjectiveTransform()
transform.estimate(source_points, destination_points)
# We transform the image
warped = warp(img, transform, output_shape=(28, 28))
if debug:
_, axis = plt.subplots(nrows=2, figsize=(8, 3))
axis[0].imshow(img, cmap='gray')
axis[0].plot(destination_points[:, 0], destination_points[:, 1], '.r')
axis[0].set_axis_off()
axis[1].imshow(warped, cmap='gray')
axis[0].set_axis_off()
plt.tight_layout()
plt.show()
sprts.append(warped)
return sprts
def apply_projection_transform_(X, intensity):
image_size = X.shape[1]
d = image_size * 0.3 * intensity
for i in range(X.shape[0]):
tl_top = random.uniform(-d, d) # Top left corner, top margin
tl_left = random.uniform(-d, d) # Top left corner, left margin
bl_bottom = random.uniform(-d, d) # Bottom left corner, bottom margin
bl_left = random.uniform(-d, d) # Bottom left corner, left margin
tr_top = random.uniform(-d, d) # Top right corner, top margin
tr_right = random.uniform(-d, d) # Top right corner, right margin
br_bottom = random.uniform(-d, d) # Bottom right corner, bottom margin
br_right = random.uniform(-d, d) # Bottom right corner, right margin
transform = ProjectiveTransform()
transform.estimate(np.array((
(tl_left, tl_top),
(bl_left, image_size - bl_bottom),
(image_size - br_right, image_size - br_bottom),
(image_size - tr_right, tr_top)
)), np.array((
(0, 0),
(0, image_size),
(image_size, image_size),
(image_size, 0)
)))
X[i] = warp(X[i], transform, output_shape=(image_size, image_size), order = 1, mode = 'edge')
return X
def normalize_image(image, image_colored, rescale_param=0.5):
image = image.copy()
image_colored = image_colored.copy()
image_scaled = rescale(image, rescale_param)
edges = canny(image_scaled)
selem = disk(1)
edges = dilation(edges, selem)
edges = (edges).astype(np.uint8)
img, ext_contours, hierarchy = cv2.findContours(edges.copy(),
cv2.RETR_EXTERNAL,
cv2.CHAIN_APPROX_SIMPLE)
contour = max(ext_contours, key=cv2.contourArea)
contour = contour.squeeze()
epsilon = 0.05 * cv2.arcLength(contour, True)
corners = cv2.approxPolyDP(contour, epsilon, True).squeeze()
corners = (perspective.order_points(corners))
corners = corners / rescale_param
size_square = min(image_scaled.shape)
tform = ProjectiveTransform()
tform.estimate(np.array([[0, 0], [1020, 0], [1020, 720], [0, 720]]),
corners)
image_warped = warp(image_colored, tform)[:720, :1020]
img = img_as_ubyte(image_warped)
img = adjusting_brightness(img[30:-5, 15:-15], a=1.7, b=2)
return img, tform
def apply_projection_transform(Xb, batch_size, image_size):
"""
Applies projection transform to a random subset of images. Projection margins are randomised in a range
depending on the size of the image. Range itself is subject to scaling depending on augmentation intensity.
"""
d = image_size * 0.3 * intensity
for i in np.random.choice(batch_size, int(batch_size * p), replace = False):
tl_top = random.uniform(-d, d) # Top left corner, top margin
tl_left = random.uniform(-d, d) # Top left corner, left margin
bl_bottom = random.uniform(-d, d) # Bottom left corner, bottom margin
bl_left = random.uniform(-d, d) # Bottom left corner, left margin
tr_top = random.uniform(-d, d) # Top right corner, top margin
tr_right = random.uniform(-d, d) # Top right corner, right margin
br_bottom = random.uniform(-d, d) # Bottom right corner, bottom margin
br_right = random.uniform(-d, d) # Bottom right corner, right margin
transform = ProjectiveTransform()
transform.estimate(np.array((
(tl_left, tl_top),
(bl_left, image_size - bl_bottom),
(image_size - br_right, image_size - br_bottom),
(image_size - tr_right, tr_top)
)), np.array((
(0, 0),
(0, image_size),
(image_size, image_size),
(image_size, 0)
)))
Xb[i] = warp(Xb[i], transform, output_shape=(image_size, image_size), order = 1, mode = 'edge')
return Xb
def warp_image_by_corner_points_projection(corner_points, image):
"""Given corner points of a Sudoku, warps original selection to a square image.
:param corner_points:
:type: corner_points: list
:param image:
:type image:
:return:
:rtype:
"""
# Clarify by storing in named variables.
top_left, top_right, bottom_left, bottom_right = np.array(corner_points)
top_edge = np.linalg.norm(top_right - top_left)
bottom_edge = np.linalg.norm(bottom_right - bottom_left)
left_edge = np.linalg.norm(top_left - bottom_left)
right_edge = np.linalg.norm(top_right - bottom_right)
L = int(np.ceil(max([top_edge, bottom_edge, left_edge, right_edge])))
src = np.array([top_left, top_right, bottom_left, bottom_right])
dst = np.array([[0, 0], [L - 1, 0], [0, L - 1], [L - 1, L - 1]])
tr = ProjectiveTransform()
tr.estimate(dst, src)
warped_image = warp(image, tr, output_shape=(L, L))
out = resize(warped_image, (500, 500))
return out
def robustEstimate(ptsA, ptsB): """ Perform robust estimation on the given correspondences using RANSAC. Args: ---- ptsA: A 2 x N matrix of points. ptsB: A 2 x N matrix of points. Returns: ------- The number of inliers within the points. """ src, dst, N = [], [], ptsA.shape[1] for i in xrange(N): src.append((ptsA[0, i], ptsA[1, i])) dst.append((ptsB[0, i], ptsB[1, i])) src, dst = np.asarray(src), np.asarray(dst) model = ProjectiveTransform() model.estimate(src, dst) model_robust, inliers = ransac((src, dst), ProjectiveTransform, min_samples=3, residual_threshold=2, max_trials=100) return inliers
def projection_transform(self, x):
image_size = x.shape[1]
change = image_size * 0.3 * self.intensity
x_return = np.empty(x.shape, dtype=x.dtype)
indices = np.random.choice(x.shape[0], int(x.shape[0] * self.ratio),
replace=False)
for i in indices:
changes = []
for _ in range(8):
changes.append(random.uniform(-change, change))
transform = ProjectiveTransform()
transform.estimate(np.array(
(
(changes[0], changes[1]), # top left
(changes[2], image_size - changes[3]), # bottom left
(image_size - changes[4], changes[5]), # top right
(image_size - changes[6], image_size - changes[7]) # bottom right
)), np.array(
(
(0, 0),
(0, image_size),
(image_size, 0),
(image_size, image_size)
))
)
x_return[i] = warp(x[i], transform,
output_shape=(image_size, image_size),
order=1, mode="edge")
return x_return
def source_to_projection(src: List, verbose: bool = False):
# for point in src:
# print(round(point[0],3), round(point[1], 3))
src = np.asarray(src) # bottom_left, top_left, top_right, bottom_right
dst = np.asarray([[0, 0], [0, 1], [1, 1], [1, 0]])
pt = ProjectiveTransform()
pt.estimate(src, dst)
return pt
def projection(self, tile: HipsTile) -> ProjectiveTransform:
"""Estimate projective transformation on a HiPS tile."""
corners = tile.meta.skycoord_corners.to_pixel(self.geometry.wcs)
src = np.array(corners).T.reshape((4, 2))
dst = tile_corner_pixel_coordinates(tile.meta.width)
pt = ProjectiveTransform()
pt.estimate(src, dst)
return pt
def _make_transform_to_crop(self, points, center_points):
dst_points = self._get_points_to_transform(points, center_points)
src_points = array([
(0, 0),
(0, self.HEIGHT),
(self.WIDTH, self.HEIGHT),
(self.WIDTH, 0),
])
tform = ProjectiveTransform()
tform.estimate(src_points, dst_points)
return tform
def correct_warping(self):
self.warpcorrected_keypoints = []
coordinates = np.array(self.coordinates[:4])
target_coordinates = np.asarray([[0, 0], [0, 1], [1, 1], [1, 0]])
t = ProjectiveTransform()
t.estimate(coordinates,target_coordinates)
for frame_count, keypoint_t in enumerate(self.keypoints):
if np.size(keypoint_t) != 0:
corrected_kp_t = t(keypoint_t[:,:2])
self.warpcorrected_keypoints.append(corrected_kp_t)
else:
self.warpcorrected_keypoints.append(np.array([]))
return self.warpcorrected_keypoints
def test_projective_estimation():
# exact solution
tform = estimate_transform('projective', SRC[:4, :], DST[:4, :])
assert_array_almost_equal(tform(SRC[:4, :]), DST[:4, :])
# over-determined
tform2 = estimate_transform('projective', SRC, DST)
assert_array_almost_equal(tform2.inverse(tform2(SRC)), SRC)
# via estimate method
tform3 = ProjectiveTransform()
tform3.estimate(SRC, DST)
assert_array_almost_equal(tform3._matrix, tform2._matrix)
def transform_img(x, y, path, cla):
"""
Projective transform of all images in a class
"""
#Scale as in LeCun
scaler = MinMaxScaler(feature_range=(-0.1, 1.175))
all_img = os.listdir(path)
#List structure so I can .append
aux = x.tolist()
for img in all_img:
if int(img[0:2]) == cla:
image_path = path + '/' + img
#prepare parameters for randomization
intensity = 0.75
image_read = cv2.imread(image_path, 0) #read in greyscale
resize = cv2.resize(image_read, (32, 32),
interpolation=cv2.INTER_CUBIC)
image_shape = resize.shape
image_size = image_shape[0]
d = image_size * 0.3 * intensity
#With these 8 parameters we can perform a transofrmation of the image in such a way
#that the image is different enough from the original but not too different, since
#we should be able to still recognize the class in the transformed image.
tl_top = random.uniform(-d, d) # Top left corner, top margin
tl_left = random.uniform(-d, d) # Top left corner, left margin
bl_bottom = random.uniform(-d,
d) # Bottom left corner, bottom margin
bl_left = random.uniform(-d, d) # Bottom left corner, left margin
tr_top = random.uniform(-d, d) # Top right corner, top margin
tr_right = random.uniform(-d, d) # Top right corner, right margin
br_bottom = random.uniform(-d,
d) # Bottom right corner, bottom margin
br_right = random.uniform(-d,
d) # Bottom right corner, right margin
transform = ProjectiveTransform()
transform.estimate(
np.array(((tl_left, tl_top), (bl_left, image_size - bl_bottom),
(image_size - br_right, image_size - br_bottom),
(image_size - tr_right, tr_top))),
np.array(((0, 0), (0, image_size), (image_size, image_size),
(image_size, 0))))
warped = warp(image_read,
transform,
output_shape=(image_size, image_size),
order=1,
mode='edge')
X_new = scaler.fit_transform(warped)
warped = np.reshape(X_new, (32, 32, 1))
aux.append(warped)
y.append(cla)
return np.array(aux), y
def randomPerspective(im,mask):
region = 1/4
A = pl.array([[0, 0], [0, im.shape[0]], [im.shape[1], im.shape[0]], [im.shape[1], 0]])
B = pl.array([[int(randRange(0, im.shape[1] * region)), int(randRange(0, im.shape[0] * region))],
[int(randRange(0, im.shape[1] * region)), int(randRange(im.shape[0] * (1-region), im.shape[0]))],
[int(randRange(im.shape[1] * (1-region), im.shape[1])), int(randRange(im.shape[0] * (1-region), im.shape[0]))],
[int(randRange(im.shape[1] * (1-region), im.shape[1])), int(randRange(0, im.shape[0] * region))],
])
pt = ProjectiveTransform()
pt.estimate(A, B)
return warp(im, pt, output_shape=im.shape[:2]), warp(mask, pt, output_shape=im.shape[:2])
def projective_transform(image, original_corners, final_corners):
'''
Correct perspective warping
'''
original_corners = np.array(original_corners)
final_corners = np.array(final_corners)
w = np.max(final_corners[:, 0])
h = np.max(final_corners[:, 1])
tform = ProjectiveTransform()
tform.estimate(final_corners, original_corners)
warped = warp(image, tform, output_shape=(h, w))
return warped
def prep_image(self):
"""Takes the solved coordinate system and makes a piecewise \
transform on the origin image to the target image"""
transform = ProjectiveTransform()
self.coord_solver.coordinates = self.coord_solver.min_coords.copy()
self.new_image = np.zeros(self.coord_solver.image.shape)
coords = np.array([self.coord_solver.coordinates[x:x+2, y:y+2, :].reshape([4, 2]) for x in \
range(self.coord_solver.coordinates.shape[0]) for y in range(self.coord_solver.coordinates.shape[1]) \
if (self.coord_solver.coordinates[x:x+2, y:y+2, :].shape == (2, 2, 2))])
canonical_coords = np.indices((self.coord_solver.width, self.coord_solver.height)).T.astype('float32')
flattened_canonical = np.array([canonical_coords[x:x+2, y:y+2, :].reshape([4, 2]) for x in \
range(canonical_coords.shape[0]-1) for y in range(canonical_coords.shape[1]-1)])
mesh_size = self.coord_solver.mesh_factor
print "needs %s calcs" % coords.shape[0]
for k in range(coords.shape[0]):
src = mesh_size*coords[k, :, :]
canon_coord = mesh_size*flattened_canonical[k, :, :]
des = mesh_size*flattened_canonical[k, :, :]
if not transform.estimate(src, des):
raise Exception("estimate failed at %s" % str(k))
area_in_question_x = canon_coord[0, 0].astype(int)
area_in_question_y = canon_coord[0, 1].astype(int)
scaled_area = tf.warp(self.coord_solver.image, transform)
self.new_image[area_in_question_y:area_in_question_y+mesh_size, \
area_in_question_x:area_in_question_x+mesh_size] += scaled_area[area_in_question_y:\
area_in_question_y+mesh_size, area_in_question_x:area_in_question_x+mesh_size]
def prep_image(self):
"""Takes the solved coordinate system and makes a piecewise \
transform on the origin image to the target image"""
transform = ProjectiveTransform()
self.coord_solver.coordinates = self.coord_solver.min_coords.copy()
self.new_image = np.zeros(self.coord_solver.image.shape)
coords = np.array([self.coord_solver.coordinates[x:x+2, y:y+2, :].reshape([4, 2]) for x in \
range(self.coord_solver.coordinates.shape[0]) for y in range(self.coord_solver.coordinates.shape[1]) \
if (self.coord_solver.coordinates[x:x+2, y:y+2, :].shape == (2, 2, 2))])
canonical_coords = np.indices((self.coord_solver.width, self.coord_solver.height)).T.astype('float32')
flattened_canonical = np.array([canonical_coords[x:x+2, y:y+2, :].reshape([4, 2]) for x in \
range(canonical_coords.shape[0]-1) for y in range(canonical_coords.shape[1]-1)])
mesh_size = self.coord_solver.mesh_factor
print "needs %s calcs" % coords.shape[0]
coord_grid = np.indices(self.coord_solver.image.shape[:-1]).T.astype('float32').reshape(-1,2)
for k in range(coords.shape[0]):
des = mesh_size*coords[k, :, :]
canon_coord = mesh_size*flattened_canonical[k, :, :]
src = mesh_size*flattened_canonical[0, :, :]
if not transform.estimate(des, canon_coord):
raise Exception("estimate failed at %s" % str(k))
area_in_question_x = canon_coord[0, 0].astype(int)
area_in_question_y = canon_coord[0, 1].astype(int)
scaled_area = tf.warp(self.coord_solver.image, transform)
area_path = path.Path([des[0],des[1],des[3],des[2],des[0]])
points_in_area = area_path.contains_points(coord_grid,radius=0.00001).reshape(self.coord_solver.image.shape[:-1])
self.new_image += scaled_area*points_in_area[:,:,np.newaxis]
def normalize_slice_projection(slice_pixel_array, crop_bounds, desired_width,
desired_height):
result = np.zeros((desired_height, desired_width))
#Crop bounds must be converted from (x, y) points to (y, x) points
source_bounds = np.asarray([[0, 0], [0, desired_width],
[desired_height, desired_width],
[desired_height, 0]])
destination_bounds = np.asarray([[crop_bounds[0][1], crop_bounds[0][0]],
[crop_bounds[1][1], crop_bounds[1][0]],
[crop_bounds[2][1], crop_bounds[2][0]],
[crop_bounds[3][1], crop_bounds[3][0]]])
projective_transform = ProjectiveTransform()
if not projective_transform.estimate(source_bounds, destination_bounds):
print("Cannot project from crop bounds to desired image dimensions")
else:
for x in range(0, desired_width):
for y in range(0, desired_height):
normalized_point = [y, x]
transform = projective_transform(normalized_point)
slice_point = transform[0]
value = MathUtil.sample_image_bilinear(slice_pixel_array,
slice_point[1],
slice_point[0])
result[y][x] = value
return result
def preprocess_image(im):
im = resize_im(im, MAX_SIDE_LENGTH)
gray = rgb2gray(im)
# Get the edges of the receipt
top_line, right_line, bottom_line, left_line = get_receipt_edges(gray)
# Intersect to get corners
TR = line_intersection(top_line, right_line)
TL = line_intersection(top_line, left_line)
BR = line_intersection(bottom_line, right_line)
BL = line_intersection(bottom_line, left_line)
# Warp so receipt corners are image corners
transform = ProjectiveTransform()
height = max([BL[1] - TL[1], BR[1] - TR[1]])
width = max([TR[0] - TL[0], BR[1] - BL[1]])
src_pts = np.array([TL, TR, BL, BR])
dest_pts = np.array([[0, 0], [width, 0], [0, height], [width, height]])
success = transform.estimate(src_pts, dest_pts)
warped_im = warp(gray, transform.inverse)[:int(height), :int(width)]
warped_gray = rgb2gray(warped_im)
enhanced_gray = img_as_ubyte(adjust_log(warped_gray))
return enhanced_gray
def increase_data(images, nb_int, labels=None):
image_size = images[0].shape[0]
data_size = images.shape[0]
for i in range(data_size):
for j in range(nb_int):
delta = random.uniform(-20,
20) # scale using augmentation intensity
d = random.uniform(-10, 10)
tl_top = random.uniform(-d, d) # Top left corner, top margin
tl_left = random.uniform(-d, d) # Top left corner, left margin
bl_bottom = random.uniform(-d,
d) # Bottom left corner, bottom margin
bl_left = random.uniform(-d, d) # Bottom left corner, left margin
tr_top = random.uniform(-d, d) # Top right corner, top margin
tr_right = random.uniform(-d, d) # Top right corner, right margin
br_bottom = random.uniform(-d,
d) # Bottom right corner, bottom margin
br_right = random.uniform(-d,
d) # Bottom right corner, right margin
transform = ProjectiveTransform()
transform.estimate(
np.array(((tl_left, tl_top), (bl_left, image_size - bl_bottom),
(image_size - br_right, image_size - br_bottom),
(image_size - tr_right, tr_top))),
np.array(((0, 0), (0, image_size), (image_size, image_size),
(image_size, 0))))
images = np.append(images, [
warp(images[i],
transform,
output_shape=(32, 32),
order=1,
mode='edge')
],
axis=0)
images = np.append(images, [
rotate(images[i], random.uniform(-delta, delta), mode='edge')
],
axis=0)
if labels is None:
pass
else:
#labels = np.append(labels,labels[i])
labels = np.append(labels, labels[i])
if i % 100 == 0:
print(i / data_size * 100, '% de progression')
return (images, labels)
def skew(self):
#return 5 versions of the original image skewed in 5 different random ways,
#http://scikit-image.org/docs/dev/api/skimage.transform.html#skimage.transform.ProjectiveTransform
outputDirectory=self.initializeDirectory("skewed") #prepare the output directory
for (fileName,image) in zip(self.inputFilenames, self.inputImageList):
(imageLength, imageWidth,_) = image.shape
skewExtend = int(max(imageLength, imageWidth) * 0.20) #Will distord the image by max 20% of max length, width
for nbNewImages in range(5): #generate 5 skewed images
d = skewExtend #distorsion applied is proportionnal to the parameter
#define the distortions to apply
topLeftTopShift = random.uniform(-d, d)
topLeftLeftShift = random.uniform(-d, d)
bottomLeftBottomShift = random.uniform(-d, d)
bottomLeftLeftShift = random.uniform(-d, d)
topRightTopShift = random.uniform(-d, d)
topRightRightShift = random.uniform(-d, d)
bottomRightBottomShift = random.uniform(-d, d)
bottomRightRightShift = random.uniform(-d, d)
#enable the projective transform
transform = ProjectiveTransform()
#tear the image
transform.estimate(np.array((
(topLeftLeftShift , topLeftTopShift),
(bottomLeftLeftShift, imageWidth- bottomLeftBottomShift),
(imageLength - bottomRightRightShift, imageWidth - bottomRightBottomShift),
(imageLength- topRightRightShift, topRightTopShift))),
np.array((
(0,0),
(0, imageWidth),
(imageLength, imageWidth),
(imageLength,0)))
)
#apply the skew
skewed = warp(image, transform, mode='edge')
skewed = skewed*255
outputFilename = os.path.join(outputDirectory,"skewed"+str(nbNewImages)+"_"+str(fileName))
cv2.imwrite (outputFilename, skewed) #store skewed image
if self.DBG:
numberImages= len(os.listdir(outputDirectory))
print("[DEBUG] : Saved {} images in {} ".format(numberImages, outputDirectory))
def make_homography(shifts, horizon_line, resolution):
horizon_line = resolution * horizon_line
left_top_dx, left_top_dy = shifts[0][0]
right_top_dx, right_top_dy = shifts[0][1]
left_hor_dx, left_hor_dy = shifts[1][0]
right_hor_dx, right_hor_dy = shifts[1][1]
points_src = np.array(
[
[0, 0], # left top
[resolution, 0], # right top
[0, horizon_line], # left bottom
[resolution, horizon_line]
],
dtype='float32') # right bottom
points_tgt = np.array(
[
[left_top_dx, left_top_dy], # left top
[resolution + right_top_dx, right_top_dy], # right top
[left_hor_dx, horizon_line + left_hor_dy], # left bottom
[resolution + right_hor_dx, horizon_line + right_hor_dy]
],
dtype='float32') # right bottom
sky_transform = ProjectiveTransform()
sky_transform.estimate(points_src, points_tgt)
points_src = np.array(
[
[0, horizon_line], # left horizon
[resolution, horizon_line], # right horizon
[0, resolution], # left bottom
[resolution, resolution]
],
dtype='float32') # right bottom
points_tgt = np.array(
[
[left_hor_dx, horizon_line - left_hor_dy], # left horizon
[resolution + right_hor_dx, horizon_line - right_hor_dy
], # right horizon
[left_top_dx, resolution - left_top_dy], # left bottom
[resolution + right_top_dx, resolution - right_top_dy]
],
dtype='float32') # right bottom
earth_transform = ProjectiveTransform()
earth_transform.estimate(points_src, points_tgt)
return sky_transform, earth_transform
def perspective_warp(img, grid_corners):
top_left, top_right, bottom_left, bottom_right = grid_corners
top_edge = np.linalg.norm(top_right - top_left)
bottom_edge = np.linalg.norm(bottom_right - bottom_left)
left_edge = np.linalg.norm(top_left - bottom_left)
right_edge = np.linalg.norm(top_right - bottom_right)
L = int(np.ceil(max([top_edge, bottom_edge, left_edge, right_edge])))
src = np.flip(grid_corners, 1) # Flip x and y axes
dst = np.array([[0, 0], [L-1, 0], [0, L-1], [L-1, L-1]])
tf = ProjectiveTransform()
tf.estimate(dst, src)
warped_img = warp(img, tf, output_shape=(L, L))
return warped_img
def apply_projection_transform_extended(X, y, intensity, labels):
X_extended = np.empty([0, X.shape[1], X.shape[2], X.shape[3]],
dtype=X.dtype)
y_extended = np.empty([0], dtype=y.dtype)
image_size = X.shape[1]
d = image_size * 0.3 * intensity
for c in labels:
X_temp = np.empty([0, X.shape[1], X.shape[2], X.shape[3]],
dtype=X.dtype)
X_temp = np.append(X_temp, X[y == c], axis=0)
for i in range(X_temp.shape[0]):
tl_top = random.uniform(-d, d) # Top left corner, top margin
tl_left = random.uniform(-d, d) # Top left corner, left margin
bl_bottom = random.uniform(-d,
d) # Bottom left corner, bottom margin
bl_left = random.uniform(-d, d) # Bottom left corner, left margin
tr_top = random.uniform(-d, d) # Top right corner, top margin
tr_right = random.uniform(-d, d) # Top right corner, right margin
br_bottom = random.uniform(-d,
d) # Bottom right corner, bottom margin
br_right = random.uniform(-d,
d) # Bottom right corner, right margin
#http://scikit-image.org/docs/dev/api/skimage.transform.html#skimage.transform.ProjectiveTransform
transform = ProjectiveTransform()
transform.estimate(
np.array(((tl_left, tl_top), (bl_left, image_size - bl_bottom),
(image_size - br_right, image_size - br_bottom),
(image_size - tr_right, tr_top))),
np.array(((0, 0), (0, image_size), (image_size, image_size),
(image_size, 0))))
X_temp[i] = warp(X_temp[i],
transform,
output_shape=(image_size, image_size),
order=1,
mode='edge',
preserve_range=True)
X_extended = np.append(X_extended, X_temp, axis=0)
y_extended = np.append(
y_extended,
np.full((X_extended.shape[0] - y_extended.shape[0]), c, dtype=int))
return X_extended, y_extended
class HomographyWidget(QtGui.QWidget):
def __init__(self, parent=None):
super(HomographyWidget, self).__init__(parent)
self._initUI()
# Initialize the UI
def _initUI(self):
# Widget parameters
self.setMinimumWidth(300)
# Create the figure
self._fig = Figure()
# Canvas configuration
self._canvas = FigureCanvas(self._fig)
self._canvas.setParent(self)
self._canvas.mpl_connect('button_press_event', self._onPick)
# Plot configuration
self._plt = self._fig.add_subplot(111)
self._plt.xaxis.set_visible(False)
self._plt.yaxis.set_visible(False)
# Finalize figure
self._fig.subplots_adjust(wspace=0, hspace=0)
# Reset the variables
self.reset()
# Create the layout
vbox = QtGui.QVBoxLayout()
# Add Canvas to the layout
vbox.addWidget(self._canvas)
# Set the layout
self.setLayout(vbox)
zp = ZoomPan()
figZoom = zp.zoom_factory(self._plt)
figPan = zp.pan_factory(self._plt)
# Reset the variables to original state
def reset(self):
self._image = None
self._render = None
self._points = []
self._lastPoints = []
self._canvas.hide()
# Set an image to the widget
def setImage(self, image):
self._image = image
self._render = image
self._canvas.show()
self._redraw()
# Get the image of the widget
def getImage(self):
pass
def setHomography(self, points):
# Save points
self._lastPoints = points
# Redraw canvas
self._redraw()
# Redraw the image and points
def _redraw(self):
# Clear the canvas
self._plt.clear()
if len(self._points) == 2 and len(self._lastPoints) == 4:
# Get points
src = self._pointsToVector(self._lastPoints)
dest = self._pointsToVector(self._rectangle())
# Compute Transformation
self._projective = ProjectiveTransform()
self._projective.estimate(src, dest)
# Prepare output image
self._render = warp(self._image, self._projective.inverse)
# Plot the image
if self._render is not None:
self._plt.autoscale(True)
self._plt.imshow(self._render)
self._plt.autoscale(False)
# Plot the points
if len(self._points) > 0:
xs = [x for (x, _) in self._rectangle()]
ys = [y for (_, y) in self._rectangle()]
self._plt.plot(xs + [xs[0]], ys + [ys[0]], '-', color='green')
xs = [x for (x, _) in self._points]
ys = [y for (_, y) in self._points]
self._plt.plot(xs + [xs[0]], ys + [ys[0]], 'o', color='blue')
# Draw the canvas
self._canvas.draw()
# Handle click events
def _onPick(self, event):
if event.button == 3:
self._redraw()
elif event.button != 1:
return
# Get point position
x = event.xdata
y = event.ydata
if x is None or y is None:
return
# For each existing points
for px, py in self._points:
# Compute distance to current point
dst = np.sqrt((px - x) ** 2 + (py - y) ** 2)
# If the distance is small remove it
if dst < 10:
self._removePoint(px, py)
self._redraw()
return
# Delegate to add the point
self._addPoint(x, y)
# Redraw the image
self._redraw()
# Add a new point
def _addPoint(self, x, y):
# Count points
n = len(self._points)
# If less than 3 points just add it
if n < 2:
self._points.append((x, y))
return
# Remove an existing point
def _removePoint(self, x, y):
# Remove the point
self._points = list(filter(lambda v: v != (x, y), self._points))
def _rectangle(self):
# Get xs and ys
xs = [x for (x, _) in self._points]
ys = [y for (_, y) in self._points]
# Compute ranges
xmax = max(xs)
xmin = min(xs)
ymax = max(ys)
ymin = min(ys)
# Return rectangle
return [(xmin, ymin), (xmin, ymax), (xmax, ymax), (xmax, ymin)]
def _pointsToVector(self, points):
# Get points values
x1, y1 = points[0]
x2, y2 = points[1]
x3, y3 = points[2]
x4, y4 = points[3]
# Return the vector
return np.array([[x1, y1], [x2, y2], [x3, y3], [x4, y4]])