def do_super_resolution(self, file_path, output_folder="output"):
filename, extension = os.path.splitext(file_path)
output_folder = output_folder + "/"
org_image = util.load_image(file_path)
util.save_image(output_folder + file_path, org_image)
if len(org_image.shape
) >= 3 and org_image.shape[2] == 3 and self.channels == 1:
scaled_image = util.resize_image_by_pil_bicubic(
org_image, self.scale)
util.save_image(output_folder + filename + "_bicubic" + extension,
scaled_image)
input_ycbcr_image = util.convert_rgb_to_ycbcr(
scaled_image, jpeg_mode=self.jpeg_mode)
output_y_image = self.do(input_ycbcr_image[:, :, 0:1])
util.save_image(output_folder + filename + "_result_y" + extension,
output_y_image)
image = util.convert_y_and_cbcr_to_rgb(output_y_image,
input_ycbcr_image[:, :,
1:3],
jpeg_mode=self.jpeg_mode)
else:
scaled_image = util.resize_image_by_pil_bicubic(
org_image, self.scale)
util.save_image(output_folder + filename + "_bicubic" + extension,
scaled_image)
image = self.do(scaled_image)
util.save_image(output_folder + filename + "_result" + extension,
image)
return 0
def do_super_resolution(self, label_file_path, file_path, output_folder="output"):
filename, extension = os.path.splitext(label_file_path)
output_folder = output_folder + "/"
org_image = util.load_image(label_file_path)
# util.save_image(output_folder + label_file_path, org_image)
input_ycbcr_image = []
if len(org_image.shape) >= 3 and org_image.shape[2] == 3 and self.channels == 1:
for i in range (0,5):
input_image = util.load_image(file_path[i])
scaled_image = util.resize_image_by_pil_bicubic(input_image, 1.0/self.scale)
# util.save_image(output_folder + file_path[i] + "_bicubic" + extension, scaled_image)
scaled_image = util.convert_rgb_to_ycbcr(scaled_image, jpeg_mode=self.jpeg_mode)
input_ycbcr_image.append(scaled_image[:, :, 0:1])
input_ycbcr_image = np.array(input_ycbcr_image)
org_image = util.convert_rgb_to_ycbcr(org_image, jpeg_mode=self.jpeg_mode)
output_y_image = self.do(input_ycbcr_image, org_image[:, :, 0:1])
# util.save_image(output_folder + filename + "_result_y" + extension, output_y_image)
img_up = util.resize_image_by_pil_bilinear(input_ycbcr_image[4], self.scale)
mse_bic = util.compute_mse(org_image[:, :, 0:1], img_up, border_size=self.scale)
mse = util.compute_mse(org_image[:, :, 0:1], output_y_image, border_size=self.scale)
image = util.convert_y_and_cbcr_to_rgb(output_y_image, org_image[:, :, 1:3], jpeg_mode=self.jpeg_mode)
elif len(org_image.shape) >= 3 and org_image.shape[2] == 1:
for i in range (0,5):
input_image = util.load_image(file_path[i])
scaled_image = util.resize_image_by_pil_bicubic(input_image, 1.0/self.scale)
# util.save_image(output_folder + file_path[i] + "_bicubic" + extension, scaled_image)
input_ycbcr_image.append(scaled_image)
input_ycbcr_image = np.array(input_ycbcr_image)
image = self.do(input_ycbcr_image, org_image)
img_up = util.resize_image_by_pil_bilinear(input_ycbcr_image[4], self.scale)
mse_bic = util.compute_mse(org_image, img_up, border_size=self.scale)
mse = util.compute_mse(org_image, image, border_size=self.scale)
else:
print('wrong')
return 0
# util.save_image(output_folder + filename + "_result" + extension, image)
return mse_bic, mse
def do_super_resolution_for_test(self, file_path, output_folder):
filename, extension = os.path.splitext(file_path)
true_image = util.set_image_alignment(util.load_image(file_path),
self.scale)
util.save_image("output/" + file_path, true_image)
input_image = util.load_input_image(file_path,
channels=self.channels,
scale=self.scale,
alignment=self.scale,
convert_ycbcr=True,
jpeg_mode=self.jpeg_mode)
util.save_image("output/" + filename + "_input" + extension,
input_image)
input_color_image = util.load_input_image(file_path,
channels=3,
scale=self.scale,
alignment=self.scale,
convert_ycbcr=False,
jpeg_mode=self.jpeg_mode)
util.save_image("output/" + filename + "_input_c" + extension,
input_color_image)
if len(true_image.shape
) >= 3 and true_image.shape[2] == 3 and self.channels == 1:
true_image = util.convert_rgb_to_y(true_image,
jpeg_mode=self.jpeg_mode)
util.save_image("output/" + filename + "_true" + extension,
true_image)
input_ycbcr_image = util.load_input_image(file_path,
channels=3,
scale=self.scale,
alignment=self.scale,
convert_ycbcr=True,
jpeg_mode=self.jpeg_mode)
output_image = self.do(input_image)
output_color_image = util.convert_y_and_cbcr_to_rgb(
output_image,
input_ycbcr_image[:, :, 1:3],
jpeg_mode=self.jpeg_mode)
util.save_image("output/" + filename + "_result_c" + extension,
output_color_image)
else:
# for monochro or rgb image
output_image = self.do(input_image)
mse = util.compute_mse(true_image,
output_image,
border_size=self.scale)
util.save_image("output/" + filename + "_result" + extension,
output_image)
print("MSE:%f PSNR:%f" % (mse, util.get_psnr(mse)))
return mse
def do_super_resolution_for_test(self, i, label_file_path, file_path, output_folder="output", output=True):
true_image = util.set_image_alignment(util.load_image(label_file_path), self.scale)
output_folder = output_folder + "/"
filename, extension = os.path.splitext(label_file_path)
input_y_image = []
if len(true_image.shape) >= 3 and true_image.shape[2] == 3 and self.channels == 1:
for j in range (i*5, i*5+5):
input_image = util.load_input_image(file_path[i], channels=1, scale=self.scale, alignment=self.scale)
input_y_image.append(input_image) # convert_ycbcr:True->False
input_y_image = np.dstack((input_y_image[0], input_y_image[1],
input_y_image[2], input_y_image[3], input_y_image[4]))
true_ycbcr_image = util.convert_rgb_to_ycbcr(true_image, jpeg_mode=self.jpeg_mode)
output_y_image = self.do(input_y_image, true_ycbcr_image[:, :, 0:1])
mse = util.compute_mse(true_ycbcr_image[:, :, 0:1], output_y_image, border_size=self.scale)
if output:
output_color_image = util.convert_y_and_cbcr_to_rgb(output_y_image, true_ycbcr_image[:, :, 1:3],
jpeg_mode=self.jpeg_mode)
loss_image = util.get_loss_image(true_ycbcr_image[:, :, 0:1], output_y_image, border_size=self.scale)
util.save_image(output_folder + label_file_path, true_image)
util.save_image(output_folder + filename + "_input" + extension, input_y_image)
util.save_image(output_folder + filename + "_true_y" + extension, true_ycbcr_image[:, :, 0:1])
util.save_image(output_folder + filename + "_result" + extension, output_y_image)
util.save_image(output_folder + filename + "_result_c" + extension, output_color_image)
util.save_image(output_folder + filename + "_loss" + extension, loss_image)
else:
for j in range (i*5, i*5+5):
input_image = util.load_input_image(file_path[i], channels=1, scale=self.scale, alignment=self.scale)
input_y_image.append(util.build_input_image(input_image, channels=self.channels, scale=self.scale, alignment=self.scale,
convert_ycbcr=False, jpeg_mode=self.jpeg_mode)) # convert_ycbcr:True->False
input_y_image = np.dstack((input_y_image[0], input_y_image[1],
input_y_image[2], input_y_image[3], input_y_image[4]))
output_image = self.do(input_y_image, true_image)
mse = util.compute_mse(true_image, output_image, border_size=self.scale)
if output:
util.save_image(output_folder + label_file_path, true_image)
util.save_image(output_folder + filename + "_result" + extension, output_image)
print("MSE:%f PSNR:%f" % (mse, util.get_psnr(mse)))
return mse
def do_super_resolution(self, file_path, output_folder):
filename, extension = os.path.splitext(file_path)
org_image = util.load_image(file_path)
input_image = util.resize_image_by_bicubic(org_image, self.scale)
util.save_image("output/" + file_path, org_image)
if len(input_image.shape
) >= 3 or input_image.shape[2] == 3 and self.channels == 1:
# use y_image for rgb_image
input_ycbcr_image = util.convert_rgb_to_ycbcr(
input_image, jpeg_mode=self.jpeg_mode)
input_y_image = input_ycbcr_image[:, :, 0:1].copy()
output_y_image = self.do(input_y_image)
image = util.convert_y_and_cbcr_to_rgb(output_y_image,
input_ycbcr_image,
jpeg_mode=self.jpeg_mode)
else:
image = self.do(org_image)
util.save_image("output/" + filename + "_result" + extension, image)
return 0
def do_super_resolution_for_test(self, file_path, output_folder="output", output=True):
filename, extension = os.path.splitext(file_path)
output_folder = output_folder + "/"
true_image = util.set_image_alignment(util.load_image(file_path), self.scale)
if len(true_image.shape) >= 3 and true_image.shape[2] == 3 and self.channels == 1:
input_y_image = util.build_input_image(true_image, channels=self.channels, scale=self.scale, alignment=self.scale,
convert_ycbcr=True, jpeg_mode=self.jpeg_mode)
true_ycbcr_image = util.convert_rgb_to_ycbcr(true_image, jpeg_mode=self.jpeg_mode)
output_y_image = self.do(input_y_image)
mse = util.compute_mse(true_ycbcr_image[:, :, 0:1], output_y_image, border_size=self.scale)
if output:
output_color_image = util.convert_y_and_cbcr_to_rgb(output_y_image, true_ycbcr_image[:, :, 1:3],
jpeg_mode=self.jpeg_mode)
loss_image = util.get_loss_image(true_ycbcr_image[:, :, 0:1], output_y_image, border_size=self.scale)
util.save_image(output_folder + file_path, true_image)
util.save_image(output_folder + filename + "_input" + extension, input_y_image)
util.save_image(output_folder + filename + "_true_y" + extension, true_ycbcr_image[:, :, 0:1])
util.save_image(output_folder + filename + "_result" + extension, output_y_image)
util.save_image(output_folder + filename + "_result_c" + extension, output_color_image)
util.save_image(output_folder + filename + "_loss" + extension, loss_image)
else:
input_image = util.load_input_image(file_path, channels=1, scale=self.scale, alignment=self.scale)
output_image = self.do(input_image)
mse = util.compute_mse(true_image, output_image, border_size=self.scale)
if output:
util.save_image(output_folder + file_path, true_image)
util.save_image(output_folder + filename + "_result" + extension, output_image)
print("MSE:%f PSNR:%f" % (mse, util.get_psnr(mse)))
return mse
def do_super_resolution_for_test(self, file_path, output_folder="output", output=True):
filename, extension = os.path.splitext(file_path)
output_folder = output_folder + "/"
true_image = util.set_image_alignment(util.load_image(file_path), self.scale) #读取文件夹中图片并根据scale图片对准
if len(true_image.shape) >= 3 and true_image.shape[2] == 3 and self.channels == 1: #图片预处理
input_y_image = util.build_input_image(true_image, channels=self.channels, scale=self.scale, alignment=self.scale,
convert_ycbcr=True, jpeg_mode=self.jpeg_mode) #从true_image中创建网络输入图片测试LR图片
true_ycbcr_image = util.convert_rgb_to_ycbcr(true_image, jpeg_mode=self.jpeg_mode) # 将true_image从RGB格式转为YCBCR格式HR图片
output_y_image = self.do(input_y_image) #输入测试LR图片到网络中得到输出
mse = util.compute_mse(true_ycbcr_image[:, :, 0:1], output_y_image, border_size=self.scale)#计算HR图与输出图的mse
if output:
output_color_image = util.convert_y_and_cbcr_to_rgb(output_y_image, true_ycbcr_image[:, :, 1:3],
jpeg_mode=self.jpeg_mode) #把输出图片YCBCR再转为RGB
loss_image = util.get_loss_image(true_ycbcr_image[:, :, 0:1], output_y_image, border_size=self.scale)#将HR图和LR输出图相减
util.save_image(output_folder + file_path, true_image)
util.save_image(output_folder + filename + "_input" + extension, input_y_image)#保存测试LR图片
util.save_image(output_folder + filename + "_true_y" + extension, true_ycbcr_image[:, :, 0:1])#保存测试HR图片
util.save_image(output_folder + filename + "_result" + extension, output_y_image)#保存测试输出图片
util.save_image(output_folder + filename + "_result_c" + extension, output_color_image)#保存输出图片转为RGB图
util.save_image(output_folder + filename + "_loss" + extension, loss_image)#保存HR图与LR输出图的差值
else:
input_image = util.load_input_image(file_path, channels=1, scale=self.scale, alignment=self.scale)
output_image = self.do(input_image)
mse = util.compute_mse(true_image, output_image, border_size=self.scale)
if output:
util.save_image(output_folder + file_path, true_image)
util.save_image(output_folder + filename + "_result" + extension, output_image)
print("MSE:%f PSNR:%f" % (mse, util.get_psnr(mse)))
return mse
# http://www.cv-foundation.org/openaccess/content_cvpr_2016/html/Kim_Deeply-Recursive_Convolutional_Network_CVPR_2016_paper.html
#
import os
import numpy as np
import super_resolution_utilty as util
print("Data Augmentation For Training Data")
training_filenames = util.get_files_in_directory("data/99/")
augmented_directory = "data/992/"
util.make_dir(augmented_directory)
for file_path in training_filenames:
org_image = util.load_image(file_path)
_, filename = os.path.split(file_path)
filename, extension = os.path.splitext(filename)
util.save_image(augmented_directory + filename + extension, org_image)
ud_image = np.flipud(org_image)
util.save_image(augmented_directory + filename + "_v" + extension,
ud_image)
lr_image = np.fliplr(org_image)
util.save_image(augmented_directory + filename + "_h" + extension,
lr_image)
lrud_image = np.flipud(lr_image)
util.save_image(augmented_directory + filename + "_hv" + extension,
lrud_image)
