def main():
mode = 'rgb'
root_dir = "./training/"
image_dir = root_dir + "images/"
gt_dir = root_dir + "groundtruth/"
imgs, _ = load_all(image_dir, mode)
gt_imgs, _ = load_all(gt_dir, mode)
count = 0
#If augmented_training does not exist we create all directories.
new_image_dir = './augmented_training/satellite/'
new_gt_dir = './augmented_training/ground_truth/'
if not os.path.exists('./augmented_training/'):
os.makedirs('./augmented_training/')
os.makedirs(new_image_dir)
os.makedirs(new_gt_dir)
for i, (img, gt_img) in enumerate(zip(imgs, gt_imgs)):
tmp = img
gt_tmp = gt_img
for k in range(2):
for j in range(4):
save_image(new_image_dir + 'sat_{}.png'.format(count), tmp)
save_image(new_gt_dir + 'gt_{}.png'.format(count), gt_tmp)
tmp = np.rot90(tmp)
gt_tmp = np.rot90(gt_tmp)
count += 1
tmp = np.flip(tmp, 0)
gt_tmp = np.flip(gt_tmp, 0)
def __init__(self, path, show_steps=False, save=False):
self.image = Helpers.load_image(path)
# build the prepocessing pipleine
pipeline = Pipeline([
Helpers.convert_to_grayscale, lambda image: Helpers.blur(image, 5),
Helpers.thresholdify, Helpers.ellipse_morph
])
processed_image = pipeline.process_pipeline(self.image)
# get the contour, crop it out, find the corners and straighten
contour = Helpers.largest_contour(processed_image)
processed_image_cropped = Helpers.cut_out_rect(processed_image,
contour)
corners = Helpers.get_corners(processed_image_cropped)
# apply the same cropping and warping to the original image
image_cropped = Helpers.cut_out_rect(self.image, contour)
straigtened_image = Helpers.warp_perspective(corners, image_cropped)
if show_steps:
Helpers.show(processed_image, 'Preprocessing')
Helpers.show(processed_image_cropped, 'Processed image cropped')
Helpers.show(image_cropped, 'Original image cropped')
Helpers.show(straigtened_image, 'Final image')
self.final = straigtened_image
if save:
Helpers.save_image(
f'{Helpers.IMAGE_DIRECTORY}/extractor_finish.png', self.final)
return None
def main(image_path):
# get the final worksheet from the image
ext = Extractor(image_path, False)
final = ext.final
# get the form code by checking the image's QR code
decoded_qr_code = reader(final)
# extract the cells and student's responses
cells = Cells(final)
# grade the worksheet by using a CNN to OCR the student's responses
grader = Grader(decoded_qr_code)
grader.grade(cells.student_responses)
worksheet = grader.display(final, cells.sorted_contours)
Helpers.save_image(f'{Helpers.IMAGE_DIRECTORY}/graded.png', worksheet)
def run():
if int(arg.images) >= 1:
model = load_generator(z_size, g_conv_dims, img_size)
for k in tqdm(range(int(arg.images)), desc='Generating images'):
seed_torch = int(torch.randint(1, int(10e5), (1, )).numpy())
seed_numpy = int(np.random.randint(1, 10e5, size=1))
np.random.seed(seed_numpy)
torch.manual_seed(seed_torch)
z_latent = create_sample(model, 1, z_size, seed_torch)
image = convert_image(z_latent)
image = np.squeeze(image, axis=0)
save_image(image,
'image_' + str(k),
int(arg.size),
path='../results/')
save_gifs(model, int(arg.size), generate=arg.gif)
else:
print(
'Unable to produce images using {}. Insert a number equal or greater than 1.'
.format(arg.images))
def undistort_img_dir(in_img_dir, M, dist, out_img_dir=None):
"""Undistort test images and apply a watermark to the image if output
directory to save to is provided.
Args:
in_img_dir (str): path to directory containing images to undistort
M (numpy.array): camera matrix (output from cv2.calibrateCamera())
dist (numpy.array): distortion coefficients (output from
cv2.calibrateCamera())
out_img_dir (str): (OPTIONAL) if specified, undistorted images will be
saved to this directory.
References:
cv2.putText
- http://docs.opencv.org/2.4/modules/core/doc/drawing_functions.html#cv2.putText
OpenCV fonts
- https://codeyarns.com/2015/03/11/fonts-in-opencv/
Transparent overlays with OpenCV
- http://www.pyimagesearch.com/2016/03/07/transparent-overlays-with-opencv/
"""
for path in glob.iglob(os.path.join(in_img_dir, '*.jpg')):
img = cv2.imread(path)
img = undistort_img(img, M, dist)
# save and show image if requested
if out_img_dir:
# save locals
file_name = os.path.split(path)[-1]
img = watermark(img, 'UNDISTORTED')
# save
save_image(img, os.path.join(out_img_dir, file_name))
def hybrid_img_generation(img_one_path, img_two_path):
# Setup
# Read images and convert to floating point format
image1 = load_image(img_one_path)
image2 = load_image(img_two_path)
image1, image2 = equalize_image_sizes(image1, image2)
# display the dog and cat images
plt.figure(figsize=(3, 3))
plt.imshow((image1 * 255).astype(np.uint8))
plt.figure(figsize=(3, 3))
plt.imshow((image2 * 255).astype(np.uint8))
# For your write up, there are several additional test cases in 'data'.
# Feel free to make your own, too (you'll need to align the images in a
# photo editor such as Photoshop).
# The hybrid images will differ depending on which image you
# assign as image1 (which will provide the low frequencies) and which image
# you asign as image2 (which will provide the high frequencies)
## Hybrid Image Construction ##
# cutoff_frequency is the standard deviation, in pixels, of the Gaussian#
# blur that will remove high frequencies. You may tune this per image pair
# to achieve better results.
cutoff_frequency = 7
low_frequencies, high_frequencies, hybrid_image = gen_hybrid_image(
image1, image2, cutoff_frequency)
## Visualize and save outputs ##
plt.figure()
plt.imshow((low_frequencies * 255).astype(np.uint8))
plt.figure()
plt.imshow(((high_frequencies + 0.5) * 255).astype(np.uint8))
vis = vis_hybrid_image(hybrid_image)
plt.figure(figsize=(20, 20))
plt.imshow(vis)
save_image('../results/low_frequencies.jpg', low_frequencies)
outHigh = np.clip(high_frequencies + 0.5, 0.0, 1.0)
save_image('../results/high_frequencies.jpg', outHigh)
save_image('../results/hybrid_image.jpg', hybrid_image)
save_image('../results/hybrid_image_scales.jpg', vis)
def create_item():
''' Save a new Item in the database '''
owner_id = session.get('user_id')
if not owner_id:
return ("You must be logged in to be able to create items", 401)
form = NewItemForm()
if form.validate_on_submit():
file = request.files.get(form.image_file.name)
saved_path = save_image(file) if file else None
new_item = Item(name=form.data["name"],
category_id=form.data["category_id"],
description=form.data["description"],
image_file=saved_path,
owner_id=owner_id)
db.session.add(new_item)
db.session.commit()
return "ok"
return render_template('edit_item.html', form=form, action="/items")
def update_item(id):
''' Updates an Item in the database '''
user_id = session.get('user_id')
if not user_id:
return ("You must be logged in to be able to edit items", 401)
item = Item.query.filter_by(id=id).first()
if not item:
return "Not Found", 404
if item.owner_id != user_id:
return ("This item doesn't belong to you", 401)
form = NewItemForm(obj=item)
if form.validate_on_submit():
file = request.files.get(form.image_file.name)
saved_path = save_image(file) if file else None
if saved_path:
item.image_file = saved_path
item.name = form.data["name"]
item.description = form.data["description"]
db.session.commit()
return "ok"
return render_template('edit_item.html', form=form, action=item.url)
def filter_test(img_path):
resultsDir = '..' + os.sep + 'results'
if not os.path.exists(resultsDir):
os.mkdir(resultsDir)
test_image = load_image(img_path)
test_image = rescale(test_image, 0.7, mode='reflect', multichannel=True)
'''
Identity filter
This filter should do nothing regardless of the padding method you use.
'''
identity_filter = np.asarray([[0, 0, 0], [0, 1, 0], [0, 0, 0]],
dtype=np.float32)
identity_image = my_imfilter(test_image, identity_filter)
plt.imshow(identity_image)
plt.show()
done = save_image('../results/identity_image.jpg', identity_image)
'''
Small blur with a box filter
This filter should remove some high frequencies.
'''
blur_filter = np.ones((3, 3), dtype=np.float32)
# making the filter sum to 1
blur_filter /= np.sum(blur_filter, dtype=np.float32)
blur_image = my_imfilter(test_image, blur_filter)
plt.imshow(blur_image)
plt.show()
done = save_image(resultsDir + os.sep + 'blur_image.jpg', blur_image)
'''
Large blur
This blur would be slow to do directly, so we instead use the fact that Gaussian blurs are separable and blur sequentially in each direction.
'''
# generate a 1x(2k+1) gaussian kernel with mean=0 and sigma = s, see https://stackoverflow.com/questions/17190649/how-to-obtain-a-gaussian-filter-in-python
s, k = 10, 12
large_1d_blur_filter = np.asarray([
exp(-z * z / (2 * s * s)) / sqrt(2 * pi * s * s)
for z in range(-k, k + 1)
],
dtype=np.float32)
large_1d_blur_filter = large_1d_blur_filter.reshape(-1, 1)
large_blur_image = my_imfilter(test_image, large_1d_blur_filter)
# notice the T operator which transposes the filter
large_blur_image = my_imfilter(large_blur_image, large_1d_blur_filter.T)
plt.imshow(large_blur_image)
plt.show()
done = save_image(resultsDir + os.sep + 'large_blur_image.jpg',
large_blur_image)
# Slow (naive) version of large blur
# import time
# large_blur_filter = np.dot(large_1d_blur_filter, large_1d_blur_filter.T)
# t = time.time()
# large_blur_image = my_imfilter(test_image, large_blur_filter);
# t = time.time() - t
# print('{:f} seconds'.format(t))
##
'''
Oriented filter (Sobel operator)
'''
sobel_filter = np.asarray(
[[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]],
dtype=np.float32) # should respond to horizontal gradients
sobel_image = my_imfilter(test_image, sobel_filter)
# 0.5 added because the output image is centered around zero otherwise and mostly black
sobel_image = np.clip(sobel_image + 0.5, 0.0, 1.0)
plt.imshow(sobel_image)
plt.show()
done = save_image(resultsDir + os.sep + 'sobel_image.jpg', sobel_image)
'''
High pass filter (discrete Laplacian)
'''
laplacian_filter = np.asarray([[0, 1, 0], [1, -4, 1], [0, 1, 0]],
dtype=np.float32)
laplacian_image = my_imfilter(test_image, laplacian_filter)
# added because the output image is centered around zero otherwise and mostly black
laplacian_image = np.clip(laplacian_image + 0.5, 0.0, 1.0)
plt.figure()
plt.imshow(laplacian_image)
plt.show()
done = save_image(resultsDir + os.sep + 'laplacian_image.jpg',
laplacian_image)
# High pass "filter" alternative
high_pass_image = test_image - blur_image
high_pass_image = np.clip(high_pass_image + 0.5, 0.0, 1.0)
plt.figure()
plt.imshow(high_pass_image)
plt.show()
done = save_image(resultsDir + os.sep + 'high_pass_image.jpg',
high_pass_image)
plt.imshow((image2*255).astype(np.uint8))
# For your write up, there are several additional test cases in 'data'.
# Feel free to make your own, too (you'll need to align the images in a
# photo editor such as Photoshop).
# The hybrid images will differ depending on which image you
# assign as image1 (which will provide the low frequencies) and which image
# you asign as image2 (which will provide the high frequencies)
## Hybrid Image Construction ##
# cutoff_frequency is the standard deviation, in pixels, of the Gaussian#
# blur that will remove high frequencies. You may tune this per image pair
# to achieve better results.
cutoff_frequency = 500
ksize = (19,19)
low_frequencies, high_frequencies, hybrid_image = gen_hybrid_image(image1, image2, cutoff_frequency, ksize)
## Visualize and save outputs ##
plt.figure()
plt.imshow((low_frequencies*255).astype(np.uint8))
plt.figure()
plt.imshow(((high_frequencies+0.5)*255).astype(np.uint8))
vis = vis_hybrid_image(hybrid_image)
plt.figure(figsize=(20, 20))
plt.imshow(vis)
save_image('../results/low_frequencies.jpg', np.clip((low_frequencies),0,1))
save_image('../results/high_frequencies.jpg', np.clip((high_frequencies+0.5),0,1))
save_image('../results/hybrid_image.jpg', hybrid_image)
save_image('../results/hybrid_image_scales.jpg', vis)
diff = time.time() - start_time
if diff >= 1:
print fps
start_time = time.time()
fps = 0
fps += 1
jpg = bytes[a:b + 2]
bytes = bytes[b + 2:]
img = cv2.imdecode(np.fromstring(jpg, dtype=np.uint8),
cv2.CV_LOAD_IMAGE_COLOR)
# save original capture
helpers.save_image(img)
# process data
img, thresh, response = helpers.detectTriangles(img)
helpers.update_thing_shadow(thing_id, response)
# display
cv2.imshow('result', img)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
# When everything done, release the capture
cv2.destroyAllWindows()
img_stretched = h.contrast_stretching(img)
img2_stretched = h.contrast_stretching(img2)
h.show_images([img, img_stretched, img2, img2_stretched],
["bad_kid", "bad_kid stretched", "dark", "dark stretched"])
############# Gamma correction testing #############
gamma = 1.5
img_corr = h.gamma_corr(img, gamma)
img2_corr = h.gamma_corr(img2, gamma)
h.show_images([img, img_corr, img2, img2_corr],
["bad_kid", "bad_kid gamma", "dark", "dark gamma"])
############# Histogram equalization testing #############
img_eq = h.hist_eq(img)
img2_eq = h.hist_eq(img2)
h.show_images([img, img_eq/255, img2, img2_eq/255],
["bad_kid", "bad_kid equalized", "dark", "dark equalized"])
h.save_image('./results/salt_pepper_noise.jpg', noise)
h.save_image('./results/median_filtered.jpg', filtered_img)
h.save_image('./results/negative_transform.jpg', inv)
h.save_image('./results/contrast_stretched.jpg', img_stretched)
h.save_image('./results/contrast_stretched2.jpg', img2_stretched)
h.save_image('./results/gamma_corrected.jpg', img_corr)
h.save_image('./results/gamma_corrected2.jpg', img2_corr)
h.save_image('./results/hist_equalized.jpg', np.clip((img_eq/255),0,1))
h.save_image('./results/hist_equalized2.jpg', np.clip((img2_eq/255),0,1))
}
pickle.dump(calibration_data, open( "calibration_data.p", "wb" ))
# Let's load the camera calibration parameters to each chessboard as additional detail
# If we don't do this, we won't be able to get an undistorted image from that instance
for chessboard in chessboards:
chessboard.load_undistort_params(camera_matrix = matrix, distortion = distortion_coef)
# Save each image to respective files
for chessboard in chessboards:
if chessboard.has_corners:
save_image(chessboard.image_with_corners(), "corners", chessboard.i)
if chessboard.can_undistort:
save_image(chessboard.undistorted_image(), "undistortedboard", chessboard.i)
# Visualization
raw_images, images_with_corners, undistorted_images = [], [], []
for chessboard in chessboards:
raw_images.append(chessboard.image())
if chessboard.has_corners:
images_with_corners.append(chessboard.image_with_corners())
if not os.path.exists(resultsDir):
os.mkdir(resultsDir)
image_title = 'mona_lisa.jpg'
image_path = dataDir + os.sep + image_title
test_image = load_image(image_path)
test_image = rescale(test_image, 0.7, multichannel=True, mode='reflect')
'''
Identity filter
This filter should do nothing regardless of the padding method you use.
'''
identity_filter = np.asarray([[0, 0, 0], [0, 1, 0], [0, 0, 0]],
dtype=np.float32)
identity_image = my_imfilter(test_image, identity_filter)
plt.imshow(identity_image)
done = save_image('../results/identity_image.jpg', identity_image)
'''
Small blur with a box filter
This filter should remove some high frequencies.
'''
blur_filter = np.ones((3, 3), dtype=np.float32)
blur_filter /= np.sum(blur_filter,
dtype=np.float32) # making the filter sum to 1
blur_image = my_imfilter(test_image, blur_filter)
plt.imshow(blur_image)
done = save_image(resultsDir + os.sep + 'blur_image.jpg', blur_image)
'''
Large blur
This blur would be slow to do directly, so we instead use the fact that Gaussian blurs are separable and blur sequentially in each direction.
'''
# generate a gaussian kernel with any parameters of your choice. you may only in this case use a function
# Validation dataset
dataset_val = helpers.PeopleDataset()
val_type = "val" if args.year in '2017' else "minival"
dataset_val.load_coco(args.dataset, val_type, year=args.year, auto_download=args.download, class_ids=[1])
dataset_val.prepare()
# Image Augmentation
# Right/Left flip 50% of the time
augmentation = imgaug.augmenters.Fliplr(0.5)
print("Training network heads")
model.train(dataset_train, dataset_val,
learning_rate=config.LEARNING_RATE,
epochs=10,
layers='heads',
augmentation=augmentation)
else:
model = modellib.MaskRCNN(mode="inference", config=config,
model_dir=args.logs)
model_path = model.find_last()
print("Loading weights ", model_path)
model.load_weights(model_path, by_name=True)
if args.image:
helpers.save_image(args.image, lambda image: model.detect([image], verbose=0))
elif args.video:
helpers.save_video(args.video, lambda image: model.detect([image], verbose=0)[0])
else:
print("You must provide either --image or --video commandline parameter when using inference mode")
# to achieve better results.
cutoff_frequency = 2
low_frequencies, high_frequencies, hybrid_image = gen_hybrid_image_fft(
image1, image2, cutoff_frequency)
# bouns
## Visualize and save outputs ##
plt.figure()
plt.imshow((low_frequencies * 255).astype(np.uint8))
plt.figure()
plt.imshow(((high_frequencies + 0.5) * 255).astype(np.uint8))
vis = vis_hybrid_image(hybrid_image)
plt.figure(figsize=(20, 20))
plt.imshow(vis)
save_image('../results/low_frequencies.jpg', low_frequencies - .1)
save_image('../results/high_frequencies.jpg', high_frequencies)
save_image('../results/hybrid_image.jpg', hybrid_image - .3)
save_image('../results/hybrid_image_scales.jpg', vis - .3)
#bouns
low_frequencies_f, high_frequencies_f, hybrid_image_f = gen_hybrid_image(
image1, image2, cutoff_frequency)
plt.figure()
plt.imshow((low_frequencies_f * 255).astype(np.uint8))
plt.figure()
plt.imshow(((high_frequencies_f + 0.5) * 255).astype(np.uint8))
vis_f = vis_hybrid_image(hybrid_image_f)
plt.figure(figsize=(20, 20))
plt.imshow(vis)
# Load image
lena = imread("../lena.jpg")
# Draw and display an image
plt.imshow(lena)
plt.show()
# Matrix indices are [rows, columns]
# Extract the rows 100 to 200 & all columns
sub_lena = lena[100:200, 0:-1]
show_image(sub_lena)
# Accessing color channels: third index -> [rows, cols, channel]
red_lena = lena[:, :, 0]
show_image(red_lena, cmap="gray")
# Convert to grayscale and use different colormap
gray_lena = rgb2gray(lena)
show_image(gray_lena, colormap="inferno")
# Convert to other colorspace
hsv_lena = convert_colorspace(lena, "RGB", "HSV")
# Only show saturation
sat_lena = hsv_lena[:, :, 1]
show_image(sat_lena, "gray")
# Save image
save_image(sat_lena, "sat_lena.jpg", "gray")
# Load image
lena = imread("../lena.jpg")
# Draw and display an image
plt.imshow(lena)
plt.show()
# Matrix indices are [rows, columns]
# Extract the rows 100 to 200 & all columns
sub_lena = lena[100:200 , 0:-1]
show_image(sub_lena)
# Accessing color channels: third index -> [rows, cols, channel]
red_lena = lena[:,:,0]
show_image(red_lena, cmap="gray")
# Convert to grayscale and use different colormap
gray_lena = rgb2gray(lena)
show_image(gray_lena, colormap="inferno")
# Convert to other colorspace
hsv_lena = convert_colorspace(lena, "RGB", "HSV")
# Only show saturation
sat_lena = hsv_lena[:,:,1]
show_image(sat_lena, "gray")
# Save image
save_image(sat_lena, "sat_lena.jpg", "gray")
plt.imshow((image2 * 255).astype(np.uint8))
# For your write up, there are several additional test cases in 'data'.
# Feel free to make your own, too (you'll need to align the images in a
# photo editor such as Photoshop).
# The hybrid images will differ depending on which image you
# assign as image1 (which will provide the low frequencies) and which image
# you asign as image2 (which will provide the high frequencies)
## Hybrid Image Construction ##
# cutoff_frequency is the standard deviation, in pixels, of the Gaussian#
# blur that will remove high frequencies. You may tune this per image pair
# to achieve better results.
cutoff_frequency = 7
low_frequencies, high_frequencies, hybrid_image = gen_hybrid_image(
image1, image2, cutoff_frequency)
## Visualize and save outputs ##
plt.figure()
plt.imshow((low_frequencies * 255).astype(np.uint8))
plt.figure()
plt.imshow(((high_frequencies + 0.5) * 255).astype(np.uint8))
vis = vis_hybrid_image(hybrid_image)
plt.figure(figsize=(20, 20))
plt.imshow(vis)
save_image('../results/low_frequencies.jpg', low_frequencies)
save_image('../results/high_frequencies.jpg', high_frequencies + 0.5)
save_image('../results/hybrid_image.jpg', hybrid_image)
save_image('../results/hybrid_image_scales.jpg', vis)
def main():
# print command line arguments
if (len(sys.argv) < 2):
print("not enought arguments, put the weight file as argument")
return
weight_path = sys.argv[1]
reg = 1e-6
model = Sequential()
model.add(
Conv2D(32, (3, 3),
padding='same',
input_shape=(window_size, window_size, 3)))
model.add(LeakyReLU(alpha=0.1))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))
model.add(Conv2D(64, (3, 3), padding='same'))
model.add(LeakyReLU(alpha=0.1))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.5))
model.add(Conv2D(128, (3, 3), padding='same'))
model.add(LeakyReLU(alpha=0.1))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.5))
model.add(Flatten())
model.add(Dense(512, kernel_regularizer=l2(reg)))
model.add(LeakyReLU(alpha=0.1))
model.add(Dropout(0.5))
model.add(Dense(1, activation='sigmoid', kernel_regularizer=l2(reg)))
opt = Adam(lr=0.001)
model.compile(loss='binary_crossentropy',
optimizer=opt,
metrics=['accuracy'])
reduce_lr = keras.callbacks.ReduceLROnPlateau(monitor='val_loss',
verbose=1,
factor=0.2,
patience=5,
min_lr=0.0)
model.load_weights(weight_path)
nb_pred = 50
pred_path = './test_set_images/test_'
save_path = pred_dir
if not os.path.exists(pred_dir):
os.makedirs(pred_dir)
for i in range(1, nb_pred + 1):
to_predict_img = load_image(
pred_path + str(i) + '/test_' + str(i) + '.png', mode)
to_predict_windows = windows_from_img(to_predict_img, window_size,
patch_size)
to_predict_windows = np.asarray(to_predict_windows)
pred = model.predict(to_predict_windows, batch_size)
save_image(
save_path + 'pred_' + str(i) + '.png',
prediction_to_img(pred, test_img_size, patch_size, threshold=0.4))
if not os.path.exists(resultsDir):
os.mkdir(resultsDir)
test_image = load_image('../data/cat.bmp')
test_image = rescale(test_image, 0.7, mode='reflect', multichannel=True)
'''
Identity filter
This filter should do nothing regardless of the padding method you use.
'''
identity_filter = np.asarray(
[[0, 0, 0], [0, 1, 0], [0, 0, 0]], dtype=np.float32)
identity_image = my_imfilter(test_image, identity_filter)
plt.imshow(identity_image)
plt.show()
done = save_image('../results/identity_image.jpg', identity_image)
'''
Small blur with a box filter
This filter should remove some high frequencies.
'''
blur_filter = np.ones((3, 3), dtype=np.float32)
# making the filter sum to 1
blur_filter /= np.sum(blur_filter, dtype=np.float32)
blur_image = my_imfilter(test_image, blur_filter)
plt.imshow(blur_image)
plt.show()
done = save_image(resultsDir + os.sep + 'blur_image.jpg', blur_image)
## Hybrid Image Construction ##
# cutoff_frequency is the standard deviation, in pixels, of the Gaussian#
# blur that will remove high frequencies. You may tune this per image pair
# to achieve better results.
cutoff_frequency = 7
low_frequencies, high_frequencies, hybrid_image = gen_hybrid_image(
image1, image2, cutoff_frequency)
## Visualize and save outputs ##
plt.figure()
plt.imshow((low_frequencies * 255).astype(np.uint8))
plt.figure()
plt.imshow(((high_frequencies + 0.5) * 255).astype(np.uint8))
vis = vis_hybrid_image(hybrid_image)
plt.figure(figsize=(20, 20))
plt.imshow(vis)
# save_image('../results/low_frequencies.jpg', low_frequencies)
# save_image('../results/high_frequencies.jpg', high_frequencies+0.5)
# save_image('../results/hybrid_image.jpg', hybrid_image)
# save_image('../results/hybrid_image_scales.jpg', vis)
# 注：经过处理后的图像像素表示形式是0-1之间的浮点数，需要转换成0-255的整型数才能存储为正常的图片，否则所有图片会存储为纯黑色（像素为0）
save_image('../results/low_frequencies.jpg',
(low_frequencies * 255).astype(np.uint8))
save_image('../results/high_frequencies.jpg',
((high_frequencies + 0.5) * 255).astype(np.uint8))
save_image('../results/hybrid_image.jpg',
(hybrid_image * 255).astype(np.uint8))
save_image('../results/hybrid_image_scales.jpg', (vis * 255).astype(np.uint8))