def upscale(self, img_path, save_intermediate=False, return_image=False, suffix="scaled",
patch_size=8, mode="patch", verbose=True):
"""
Standard method to upscale an image.
:param img_path: path to the image
:param save_intermediate: saves the intermediate upscaled image (bilinear upscale)
:param return_image: returns a image of shape (height, width, channels).
:param suffix: suffix of upscaled image
:param patch_size: size of each patch grid
:param verbose: whether to print messages
:param mode: mode of upscaling. Can be "patch" or "fast"
"""
import os
from scipy.misc import imread, imresize, imsave
# Destination path
path = os.path.splitext(img_path)
filename = path[0] + "_" + suffix + "(%dx)" % (self.scale_factor) + path[1]
# Read image
scale_factor = int(self.scale_factor)
true_img = imread(img_path, mode='RGB')
init_width, init_height = true_img.shape[0], true_img.shape[1]
if verbose: print("Old Size : ", true_img.shape)
if verbose: print("New Size : (%d, %d, 3)" % (init_height * scale_factor, init_width * scale_factor))
img_height, img_width = 0, 0
if mode == "patch" and self.type_true_upscaling:
# Overriding mode for True Upscaling models
mode = 'fast'
print("Patch mode does not work with True Upscaling models yet. Defaulting to mode='fast'")
if mode == 'patch':
# Create patches
if self.type_requires_divisible_shape:
if patch_size % 4 != 0:
print("Deep Denoise requires patch size which is multiple of 4.\nSetting patch_size = 8.")
patch_size = 8
images = img_utils.make_patches(true_img, scale_factor, patch_size, verbose)
nb_images = images.shape[0]
img_width, img_height = images.shape[1], images.shape[2]
print("Number of patches = %d, Patch Shape = (%d, %d)" % (nb_images, img_height, img_width))
else:
# Use full image for super resolution
img_width, img_height = self.__match_autoencoder_size(img_height, img_width, init_height,
init_width, scale_factor)
images = imresize(true_img, (img_width, img_height))
images = np.expand_dims(images, axis=0)
print("Image is reshaped to : (%d, %d, %d)" % (images.shape[1], images.shape[2], images.shape[3]))
# Save intermediate bilinear scaled image is needed for comparison.
intermediate_img = None
if save_intermediate:
if verbose: print("Saving intermediate image.")
fn = path[0] + "_intermediate_" + path[1]
intermediate_img = imresize(true_img, (init_width * scale_factor, init_height * scale_factor))
imsave(fn, intermediate_img)
# Transpose and Process images
if K.image_dim_ordering() == "th":
img_conv = images.transpose((0, 3, 1, 2)).astype(np.float32) / 255.
else:
img_conv = images.astype(np.float32) / 255.
model = self.create_model(img_height, img_width, load_weights=True)
if verbose: print("Model loaded.")
# Create prediction for image patches
result = model.predict(img_conv, batch_size=128, verbose=verbose)
if verbose: print("De-processing images.")
# Deprocess patches
if K.image_dim_ordering() == "th":
result = result.transpose((0, 2, 3, 1)).astype(np.float32) * 255.
else:
result = result.astype(np.float32) * 255.
# Output shape is (original_width * scale, original_height * scale, nb_channels)
if mode == 'patch':
out_shape = (init_width * scale_factor, init_height * scale_factor, 3)
result = img_utils.combine_patches(result, out_shape, scale_factor)
else:
result = result[0, :, :, :] # Access the 3 Dimensional image vector
result = np.clip(result, 0, 255).astype('uint8')
result = cv2.pyrUp(result)
result = cv2.medianBlur(result, 3)
result = cv2.pyrDown(result)
if verbose: print("\nCompleted De-processing image.")
if return_image:
# Return the image without saving. Useful for testing images.
return result
if verbose: print("Saving image.")
imsave(filename, result)
def upscale(self,
img_path,
scale_factor=2,
save_intermediate=False,
return_image=False,
suffix="scaled",
patch_size=8,
mode="patch",
verbose=True,
evaluate=True):
"""
Standard method to upscale an image.
:param img_path: path to the image
:param scale_factor: scale factor can be any value, usually is an integer
:param save_intermediate: saves the intermediate upscaled image (bilinear upscale)
:param return_image: returns a image of shape (height, width, channels).
:param suffix: suffix of upscaled image
:param patch_size: size of each patch grid
:param verbose: whether to print messages
:param evaluate: evaluate the upscaled image on the original image.
"""
import os
from scipy.misc import imread, imresize, imsave
# Destination path
path = os.path.splitext(img_path)
filename = path[0] + "_" + suffix + "(%dx)" % (scale_factor) + path[1]
# Read image
scale_factor = int(scale_factor)
true_img = imread(img_path, mode='RGB')
init_width, init_height = true_img.shape[0], true_img.shape[1]
if verbose: print("Old Size : ", true_img.shape)
if verbose:
print("New Size : (%d, %d, 3)" %
(init_height * scale_factor, init_width * scale_factor))
img_height, img_width = 0, 0
if mode == 'patch':
# Create patches
if self.model_name in self.denoise_models:
if patch_size % 4 != 0:
print(
"Deep Denoise requires patch size which is multiple of 4.\nSetting patch_size = 8."
)
patch_size = 8
images = img_utils.make_patches(true_img, scale_factor, patch_size,
verbose)
nb_images = images.shape[0]
img_width, img_height = images.shape[1], images.shape[2]
print("Number of patches = %d, Patch Shape = (%d, %d)" %
(nb_images, img_height, img_width))
else:
# Use full image for super resolution
img_width, img_height = self.__match_denoise_size(
denoise_models, img_height, img_width, init_height, init_width,
scale_factor)
images = imresize(true_img, (img_width, img_height))
images = np.expand_dims(images, axis=0)
print("Image is reshaped to : (%d, %d, %d)" %
(images.shape[1], images.shape[2], images.shape[3]))
# Save intermediate bilinear scaled image is needed for comparison.
intermediate_img = None
if save_intermediate:
if verbose: print("Saving intermediate image.")
fn = path[0] + "_intermediate_" + path[1]
intermediate_img = imresize(
true_img,
(init_width * scale_factor, init_height * scale_factor))
imsave(fn, intermediate_img)
# Transpose and Process images
if K.image_dim_ordering() == "th":
img_conv = images.transpose((0, 3, 1, 2)).astype(np.float32) / 255.
else:
img_conv = images.astype(np.float32) / 255.
model = self.create_model(img_height, img_width, load_weights=True)
if verbose: print("Model loaded.")
# Create prediction for image patches
result = model.predict(img_conv, batch_size=128, verbose=verbose)
if verbose: print("De-processing images.")
# Deprocess patches
if K.image_dim_ordering() == "th":
result = result.transpose((0, 2, 3, 1)).astype(np.float32) * 255.
else:
result = result.astype(np.float32) * 255.
# Output shape is (original_width * scale, original_height * scale, nb_channels)
if mode == 'patch':
out_shape = (init_width * scale_factor, init_height * scale_factor,
3)
result = img_utils.combine_patches(result, out_shape, scale_factor)
else:
result = result[
0, :, :, :] # Access the 3 Dimensional image vector
result = np.clip(result, 0, 255).astype('uint8')
if verbose: print("\nCompleted De-processing image.")
if return_image:
# Return the image without saving. Useful for testing images.
return result
if verbose: print("Saving image.")
imsave(filename, result)
if evaluate:
if verbose: print("Evaluating results.")
# Convert initial image into correct format
if intermediate_img is None:
intermediate_img = imresize(
true_img,
(init_width * scale_factor, init_height * scale_factor))
if mode == 'patch':
intermediate_img = img_utils.make_patches(intermediate_img,
scale_factor,
patch_size,
upscale=False)
else:
img_width, img_height = self.__match_denoise_size(
denoise_models, img_height, img_width, init_height,
init_width, scale_factor)
intermediate_img = imresize(true_img, (img_width, img_height))
intermediate_img = np.expand_dims(intermediate_img, axis=0)
if K.image_dim_ordering() == "th":
intermediate_img = intermediate_img.transpose(
(0, 3, 1, 2)).astype(np.float32) / 255.
else:
intermediate_img = intermediate_img.astype(np.float32) / 255.
eval_model = self.create_model(img_height,
img_width,
load_weights=True)
# Evaluating the initial image patches, which gets transformed to the output image, to the input image
error = eval_model.evaluate(img_conv,
intermediate_img,
batch_size=128)
print("\nMean Squared Error of %s : " % (self.model_name),
error[0])
print("Peak Signal to Noise Ratio of %s : " % (self.model_name),
error[1])
def upscale(self, img_path, scale_factor=2, save_intermediate=False, return_image=False, suffix="scaled",
patch_size=8, mode="patch", verbose=True, evaluate=True):
"""
Standard method to upscale an image.
:param img_path: path to the image
:param scale_factor: scale factor can be any value, usually is an integer
:param save_intermediate: saves the intermediate upscaled image (bilinear upscale)
:param return_image: returns a image of shape (height, width, channels).
:param suffix: suffix of upscaled image
:param patch_size: size of each patch grid
:param verbose: whether to print messages
:param evaluate: evaluate the upscaled image on the original image.
"""
import os
from scipy.misc import imread, imresize, imsave
# Destination path
path = os.path.splitext(img_path)
filename = path[0] + "_" + suffix + "(%dx)" % (scale_factor) + path[1]
# Read image
scale_factor = int(scale_factor)
true_img = imread(img_path, mode='RGB')
init_width, init_height = true_img.shape[0], true_img.shape[1]
if verbose: print("Old Size : ", true_img.shape)
if verbose: print("New Size : (%d, %d, 3)" % (init_height * scale_factor, init_width * scale_factor))
img_height, img_width = 0, 0
if mode == 'patch':
# Create patches
if self.model_name in self.denoise_models:
if patch_size % 4 != 0:
print("Deep Denoise requires patch size which is multiple of 4.\nSetting patch_size = 8.")
patch_size = 8
images = img_utils.make_patches(true_img, scale_factor, patch_size, verbose)
nb_images = images.shape[0]
img_width, img_height = images.shape[1], images.shape[2]
print("Number of patches = %d, Patch Shape = (%d, %d)" % (nb_images, img_height, img_width))
else:
# Use full image for super resolution
img_width, img_height = self.__match_denoise_size(denoise_models, img_height, img_width, init_height,
init_width, scale_factor)
images = imresize(true_img, (img_width, img_height))
images = np.expand_dims(images, axis=0)
print("Image is reshaped to : (%d, %d, %d)" % (images.shape[1], images.shape[2], images.shape[3]))
# Save intermediate bilinear scaled image is needed for comparison.
intermediate_img = None
if save_intermediate:
if verbose: print("Saving intermediate image.")
fn = path[0] + "_intermediate_" + path[1]
intermediate_img = imresize(true_img, (init_width * scale_factor, init_height * scale_factor))
imsave(fn, intermediate_img)
# Transpose and Process images
if K.image_dim_ordering() == "th":
img_conv = images.transpose((0, 3, 1, 2)).astype(np.float32) / 255.
else:
img_conv = images.astype(np.float32) / 255.
model = self.create_model(img_height, img_width, load_weights=True)
if verbose: print("Model loaded.")
# Create prediction for image patches
result = model.predict(img_conv, batch_size=128, verbose=verbose)
if verbose: print("De-processing images.")
# Deprocess patches
if K.image_dim_ordering() == "th":
result = result.transpose((0, 2, 3, 1)).astype(np.float32) * 255.
else:
result = result.astype(np.float32) * 255.
# Output shape is (original_width * scale, original_height * scale, nb_channels)
if mode == 'patch':
out_shape = (init_width * scale_factor, init_height * scale_factor, 3)
result = img_utils.combine_patches(result, out_shape, scale_factor)
else:
result = result[0, :, :, :] # Access the 3 Dimensional image vector
result = np.clip(result, 0, 255).astype('uint8')
if verbose: print("\nCompleted De-processing image.")
if return_image:
# Return the image without saving. Useful for testing images.
return result
if verbose: print("Saving image.")
imsave(filename, result)
if evaluate:
if verbose: print("Evaluating results.")
# Convert initial image into correct format
if intermediate_img is None:
intermediate_img = imresize(true_img, (init_width * scale_factor, init_height * scale_factor))
if mode == 'patch':
intermediate_img = img_utils.make_patches(intermediate_img, scale_factor, patch_size, upscale=False)
else:
img_width, img_height = self.__match_denoise_size(denoise_models, img_height, img_width, init_height,
init_width, scale_factor)
intermediate_img = imresize(true_img, (img_width, img_height))
intermediate_img = np.expand_dims(intermediate_img, axis=0)
if K.image_dim_ordering() == "th":
intermediate_img = intermediate_img.transpose((0, 3, 1, 2)).astype(np.float32) / 255.
else:
intermediate_img = intermediate_img.astype(np.float32) / 255.
eval_model = self.create_model(img_height, img_width, load_weights=True)
# Evaluating the initial image patches, which gets transformed to the output image, to the input image
error = eval_model.evaluate(img_conv, intermediate_img, batch_size=128)
print("\nMean Squared Error of %s : " % (self.model_name), error[0])
print("Peak Signal to Noise Ratio of %s : " % (self.model_name), error[1])
print("New Size : (%d, %d, 3)" % (init_dim_1 * scaling_factor, init_dim_2 * scaling_factor))
images = make_patches(true_img, scaling_factor, patch_size, verbose)
nb_images = images.shape[0]
img_dim_1, img_dim_2 = images.shape[1], images.shape[2]
print("Number of patches = %d, Patch Shape = (%d, %d)" % (nb_images, img_dim_2, img_dim_1))
if verbose: print("Saving intermediate image...")
fn = path[0] + "_intermediate_" + path[1]
intermediate_img = imresize(true_img, (init_dim_1 * scaling_factor, init_dim_2 * scaling_factor))
imsave(fn, intermediate_img)
img_conv = images.astype(np.float32) / 255.
result = model.predict(img_conv, batch_size=64, verbose=verbose)
if verbose: print("De-processing images...")
result = result.astype(np.float32) * 255.
out_shape = (init_dim_1 * scaling_factor, init_dim_2 * scaling_factor, 3)
result = combine_patches(result, out_shape, scaling_factor)
result = np.clip(result, 0, 255).astype('uint8')
result = cv2.pyrUp(result)
result = cv2.medianBlur(result, 3)
result = cv2.pyrDown(result)
if verbose: print("De-processing image Completed!!!")
if verbose: print("Saving image...")
imsave(filename, result)
def upscale(self, model, img_path, save_intermediate=False, return_image=False, suffix="scaled",
patch_size=8, mode="patch", verbose=True):
"""
Standard method to upscale an image.
:param img_path: path to the image
:param save_intermediate: saves the intermediate upscaled image (bilinear upscale)
:param return_image: returns a image of shape (height, width, channels).
:param suffix: suffix of upscaled image
:param patch_size: size of each patch grid
:param verbose: whether to print messages
:param mode: mode of upscaling. Can be "patch" or "fast"
"""
import os
from scipy.misc import imread, imresize, imsave
# Destination path
path = os.path.splitext(img_path)
filename = "./final1/"+ path[0] + path[1]
# Read image
scale_factor = int(self.scale_factor)
true_img = imread(img_path, mode='RGB')
init_dim_1, init_dim_2 = true_img.shape[0], true_img.shape[1]
if verbose: print("Old Size : ", true_img.shape)
if verbose: print("New Size : (%d, %d, 3)" % (init_dim_1 * scale_factor, init_dim_2 * scale_factor))
img_dim_1, img_dim_2 = 0, 0
# Use full image for super resolution
img_dim_1, img_dim_2 = self.__match_autoencoder_size(img_dim_1, img_dim_2, init_dim_1, init_dim_2,
scale_factor)
images = imresize(true_img, (img_dim_1, img_dim_2))
images = np.expand_dims(images, axis=0)
print("Image is reshaped to : (%d, %d, %d)" % (images.shape[1], images.shape[2], images.shape[3]))
# Save intermediate bilinear scaled image is needed for comparison.
# intermediate_img = None
# if save_intermediate:
# if verbose: print("Saving intermediate image.")
# fn = path[0] + "_intermediate_" + path[1]
# intermediate_img = imresize(true_img, (init_dim_1 * scale_factor, init_dim_2 * scale_factor))
# imsave(fn, intermediate_img)
# Transpose and Process images
if K.image_dim_ordering() == "th":
img_conv = images.transpose((0, 3, 1, 2)).astype(np.float32) / 255.
else:
img_conv = images.astype(np.float32) / 255.
model = model
# model = self.create_model(img_dim_2, img_dim_1, load_weights=True)
# if verbose: print("Model loaded.")
# Create prediction for image patches
result = model.predict(img_conv, batch_size=128, verbose=verbose)
if verbose: print("De-processing images.")
# Deprocess patches
if K.image_dim_ordering() == "th":
result = result.transpose((0, 2, 3, 1)).astype(np.float32) * 255.
else:
result = result.astype(np.float32) * 255.
# Output shape is (original_width * scale, original_height * scale, nb_channels)
if mode == 'patch':
out_shape = (init_dim_1 * scale_factor, init_dim_2 * scale_factor, 3)
result = img_utils.combine_patches(result, out_shape, scale_factor)
else:
result = result[0, :, :, :] # Access the 3 Dimensional image vector
result = np.clip(result, 0, 255).astype('uint8')
if _cv2_available:
# used to remove noisy edges
result = cv2.pyrUp(result)
result = cv2.medianBlur(result, 3)
result = cv2.pyrDown(result)
if verbose: print("\nCompleted De-processing image.")
if return_image:
# Return the image without saving. Useful for testing images.
return result
if verbose: print("Saving image.")
imsave(filename, result)