def main():
args = get_arguments()
with open("./params.json", 'r') as f:
params = json.load(f)
if check_params(args, params) == False:
return
sess = tf.Session()
net = EspcnNet(filters_size=params['filters_size'],
channels=params['channels'],
ratio=params['ratio'])
lr_image = tf.placeholder(tf.uint8)
sr_image = net.generate(lr_image)
saver = tf.train.Saver()
saver.restore(sess, args.checkpoint)
lr_image_data = misc.imread(args.lr_image)
lr_image_ycbcr_data = rgb2ycbcr(lr_image_data)
lr_image_y_data = lr_image_ycbcr_data[:, :, 0:1]
lr_image_cb_data = lr_image_ycbcr_data[:, :, 1:2]
lr_image_cr_data = lr_image_ycbcr_data[:, :, 2:3]
lr_image_batch = np.zeros((1, ) + lr_image_y_data.shape)
lr_image_batch[0] = lr_image_y_data
sr_image_y_data = sess.run(sr_image, feed_dict={lr_image: lr_image_batch})
sr_image_y_data = my_shuffle(sr_image_y_data[0], args.ratio)
sr_image_ycbcr_data = misc.imresize(
lr_image_ycbcr_data,
params['ratio'] * np.array(lr_image_data.shape[0:2]), 'bicubic')
edge = params['edge'] * params['ratio'] / 2
sr_image_ycbcr_data = np.concatenate(
(sr_image_y_data, sr_image_ycbcr_data[edge:-edge, edge:-edge, 1:3]),
axis=2)
sr_image_data = ycbcr2rgb(sr_image_ycbcr_data)
misc.imsave(args.out_path + '.png', sr_image_data)
if args.hr_image != None:
hr_image_data = misc.imread(args.hr_image)
model_psnr = psnr(hr_image_data, sr_image_data, edge)
print('PSNR of the model: {:.2f}dB'.format(model_psnr))
sr_image_bicubic_data = misc.imresize(
lr_image_data,
params['ratio'] * np.array(lr_image_data.shape[0:2]), 'bicubic')
misc.imsave(args.out_path + '_bicubic.png', sr_image_bicubic_data)
bicubic_psnr = psnr(hr_image_data, sr_image_bicubic_data, 0)
print('PSNR of Bicubic: {:.2f}dB'.format(bicubic_psnr))
def super_resolution(path_string_lr, path_string_truth, generator):
pic = cv2.imread(path_string_lr)
pic_truth = cv2.imread(path_string_truth)
pic_norm = pic / 255.0
orig_shape = pic_norm.shape
decomposed_picture = test_transform.test_image_decomposition(pic_norm,
input_size=16)
decomposed_picture_4dim = f2.ndarray_to_4dim(decomposed_picture[0],
image_size=16)
pic_predicted = generator.predict(decomposed_picture_4dim)
whole_predicted = test_transform.text_image_composition(
pic_predicted, decomposed_picture[2])
predicted_orig_shape = cv2.resize(whole_predicted,
(orig_shape[1] * 2, orig_shape[0] * 2),
interpolation=cv2.INTER_LINEAR)
psnr_item = psnr.psnr(predicted_orig_shape, pic_truth / 255.0)
return psnr_item, pic, pic_truth, predicted_orig_shape
def yuv_diff_temporal(filename_input, start_frame_input, w, h, frame_size, block_size, scale):
# read yuv : input
array_input_y, array_input_cbcr = read_yuv420(filename_input, w, h, frame_size, start_frame=start_frame_input)
input_y = array_input_y.squeeze()
## make a copy of pixel domain
#input_y_pel = input_y.copy()
# normalization
input_y = input_y / 255.
# make a copy of pixel domain
input_y_pel = input_y.copy()
# count for each frame
block_count = 0
# loop based on block
list_df_psnr = []
list_df_ssim = []
for frame_index in range(1, frame_size):
list_psnr_frame = []
list_ssim_frame = []
for y in range(0, h, block_size):
list_psnr_row = []
list_ssim_row = []
for x in range(0, w, block_size):
# pick block from input
sub_input_y_prev = input_y[frame_index-1, y:y+block_size, x:x+block_size]
sub_input_y_curr = input_y[frame_index, y:y+block_size, x:x+block_size]
# PSNR calculation
each_psnr = psnr(sub_input_y_prev, sub_input_y_curr, scale)
each_ssim = ssim(sub_input_y_prev, sub_input_y_curr, data_range=1)
list_psnr_row.append(float("{0:.4f}".format(each_psnr)))
list_ssim_row.append(float("{0:.4f}".format(each_ssim)))
block_count += 1
list_psnr_frame.append(list_psnr_row)
list_ssim_frame.append(list_ssim_row)
# outside of the loop
# stat the block-level PSNR
# list_psnr to df
df_psnr = pd.DataFrame(list_psnr_frame)
df_ssim = pd.DataFrame(list_ssim_frame)
# append to list
list_df_psnr.append(df_psnr)
list_df_ssim.append(df_ssim)
return list_df_psnr, list_df_ssim, input_y_pel
def computeError(img1, img2):
img_psnr = psnr.psnr(img1, img2)
gt_y = cv2.cvtColor(cv2.cvtColor(img1.astype(np.uint8), cv2.COLOR_RGB2BGR),
cv2.COLOR_BGR2YCR_CB)[:, :, 0]
pred_y = cv2.cvtColor(
cv2.cvtColor(img2.astype(np.uint8), cv2.COLOR_RGB2BGR),
cv2.COLOR_BGR2YCR_CB)[:, :, 0]
img_ssim = compare_ssim(gt_y, pred_y, gaussian_weight=True)
return img_psnr, img_ssim
def test_pVDSR_with_sess(ckpt_path, data_path, sess):
folder_list = glob.glob(os.path.join(data_path, 'Set*'))
print 'folder_list', folder_list
if not os.path.exists('./output_img'):
os.mkdir('./output_img')
psnr_dict = {}
for folder_path in folder_list:
psnr_list = []
img_list = get_img_list(folder_path)
for i in range(len(img_list)):
input_list, gt_list, scale_list = get_test_image(img_list, i, 1)
input_y = input_list[0]
gt_y = gt_list[0]
start_t = time.time()
img_vdsr_y = sess.run([output_tensor], feed_dict={input_tensor: np.resize(input_y, (1, input_y.shape[0], input_y.shape[1], 1))})
img_vdsr_y = np.resize(img_vdsr_y, (input_y.shape[0], input_y.shape[1]))
end_t = time.time()
# save test image and results
id = img_list[i][1].split('/')[-1].split('.')[0]
output_id = 'output' + id + '.png'
input_id = 'input' + id + '.png'
misc.imsave(os.path.join('./output_img',output_id),img_vdsr_y)
misc.imsave(os.path.join('./output_img',input_id),input_y)
print "end_t",end_t,"start_t",start_t
print "time consumption",end_t-start_t
print "image_size", input_y.shape
# calculate psnr
psnr_bicub = psnr(input_y, gt_y, scale_list[0])
psnr_vdsr = psnr(img_vdsr_y, gt_y, scale_list[0])
print "PSNR: bicubic %f\tpVDSR %f" % (psnr_bicub, psnr_vdsr)
psnr_list.append([psnr_bicub, psnr_vdsr, scale_list[0]])
psnr_dict[os.path.basename(folder_path)] = psnr_list
if not os.path.exists('./psnr'):
os.mkdir('psnr')
with open('psnr/%s' % "result", 'wb') as f:
pickle.dump(psnr_dict, f)
def fitness(chromosome, stego, secret):
"""Computes fitness for current chromosome"""
if len(chromosome) > 7:
chromosome = helper_individual.packchromosome(chromosome)
# Embed the secret sequence
try:
stego1 = embed(stego, secret, chromosome)
except:
return (0, )
return (psnr(stego, stego1), )
def test():
'''
Test the model, need model.h5 -> run the train module
before testing
'''
srcnn_model = model()
srcnn_model.load_weights("./model/srcnn_model.h5")
img = cv2.imread(IMG_NAME)
shape = img.shape
img = cv2.resize(img, (int(shape[1] / 2), int(shape[0] / 2)),
cv2.INTER_CUBIC)
img = cv2.resize(img, (shape[1], shape[0]), cv2.INTER_CUBIC)
cv2.imwrite(BICUBIC_NAME, img)
img = cv2.cvtColor(img, cv2.COLOR_BGR2YCrCb)
Y = numpy.zeros((1, img.shape[0], img.shape[1], 1))
Y[0, :, :, 0] = img[:, :, 0]
pre = srcnn_model.predict(Y, batch_size=1)
pre[pre[:] > 255] = 255
pre[pre[:] < 0] = 0
pre = pre.astype(numpy.uint8)
img[6:-6, 6:-6, 0] = pre[0, :, :, 0]
img = cv2.cvtColor(img, cv2.COLOR_YCrCb2BGR)
if denoise:
img = cv2.fastNlMeansDenoisingColored(img, None, 5, 5, 7, 21)
cv2.imwrite(OUTPUT_NAME, img)
# PSNR and MSE calculation:
im1 = cv2.imread(IMG_NAME, cv2.IMREAD_COLOR)
im1 = cv2.cvtColor(im1, cv2.COLOR_BGR2YCrCb)[6:-6, 6:-6, 0]
im2 = cv2.imread(BICUBIC_NAME, cv2.IMREAD_COLOR)
im2 = cv2.cvtColor(im2, cv2.COLOR_BGR2YCrCb)[6:-6, 6:-6, 0]
im3 = cv2.imread(OUTPUT_NAME, cv2.IMREAD_COLOR)
im3 = cv2.cvtColor(im3, cv2.COLOR_BGR2YCrCb)[6:-6, 6:-6, 0]
print("test completed... below is the report")
print("PSNR bicubic: ", psnr.psnr(im2, im1))
print("PSNR SRCNN: ", psnr.psnr(im3, im1))
print("MSE bicubic: ", psnr.mse(im2, im1))
print("MSE srcnn: ", psnr.mse(im3, im1))
def yuv_diff_single_frame(label_y, input_y, w, h, block_size, scale):
# normalization
label_y = label_y / 255.
input_y = input_y / 255.
df_psnr = pd.DataFrame()
df_sse = pd.DataFrame()
# count for each frame
block_count = 0
# loop based on block
list_psnr_frame = []
list_sse_frame = []
for y in range(0, h, block_size):
list_psnr_row = []
list_sse_row = []
for x in range(0, w, block_size):
# pick block from input
sub_label_y = label_y[y:y+block_size, x:x+block_size]
sub_input_y = input_y[y:y+block_size, x:x+block_size]
# PSNR calculation
each_psnr = psnr(sub_input_y, sub_label_y, scale)
#each_ssim = ssim(sub_input_y, sub_label_y, data_range=256)
list_psnr_row.append(float("{0:.4f}".format(each_psnr)))
# SSE
each_sse = np.sum((sub_label_y - sub_input_y) ** 2)
list_sse_row.append(float("{0:.4f}".format(each_sse)))
block_count += 1
list_psnr_frame.append(list_psnr_row)
list_sse_frame.append(list_sse_row)
# outside of the loop-block
# stat the block-level PSNR
# list_psnr to df
df_psnr_frm = pd.DataFrame(list_psnr_frame)
df_sse_frm = pd.DataFrame(list_sse_frame)
# append to global table.
df_psnr = df_psnr.append(df_psnr_frm)
df_sse = df_sse.append(df_sse_frm)
return df_psnr, df_sse
def psnr(ref, teste):
# RFZ: precisa converter aqui para grayscale (?)
l = plt.imread(ref)#[:,:,1]
#print(l.shape)
#f = open('arquivo.txt','w')
#f.write(ref + )
#for i in range(1, 9):
k = plt.imread(teste)#[:,:,1]
psrn = psnr.psnr(l, k)
return psnr
def psnr(ref, teste):
# RFZ: precisa converter aqui para grayscale (?)
l = plt.imread(ref) #[:,:,1]
#print(l.shape)
#f = open('arquivo.txt','w')
#f.write(ref + )
#for i in range(1, 9):
k = plt.imread(teste) #[:,:,1]
psrn = psnr.psnr(l, k)
return psnr
def index_results(dist_filelist, ref_filelist):
psnr_list = []
ssim_list = []
uiqm_list = []
for dist_file in dist_filelist: #遍历
dist_file_dir = os.path.join(dist_path, dist_file) #参考文件绝对路径
if os.path.isdir(dist_file_dir): #如果是文件夹，跳过
continue
dist_img = scipy.misc.imread(dist_file_dir,
flatten=True).astype(numpy.float32)
#filename = os.path.splitext(dist_file)[0] + os.path.splitext(dist_file)[1] #ref filename
filename = os.path.splitext(dist_file)[0] + '.jpg' #ref filename
ref_img = scipy.misc.imread(ref_path + '\\' + filename,
flatten=True).astype(
numpy.float32) #读取参考图像对应的测试图像
psnr_data = psnr.psnr(ref_img, dist_img) #psnr指标
ssim_data = ssim.ssim_exact(ref_img / 255, dist_img / 255) #ssim指标
dist_img = cv2.imread(dist_path + '\\' +
os.path.splitext(dist_file)[0] + '.jpg') #UIQM指标
uiqm_data = uqim_utils.getUIQM(dist_img) #UIQM
print("img:" + str(filename) + " psnr:" + str(psnr_data) + " ssim:" +
str(ssim_data) + " UIQM:" + str(uiqm_data))
data = str(filename) + " psnr:" + str(psnr_data) + " ssim:" + str(
ssim_data) + " UIQM:" + str(uiqm_data)
psnr_list.append(psnr_data)
ssim_list.append(ssim_data)
uiqm_list.append(uiqm_data)
average = "psnr_average:" + str(
sum(psnr_list) / len(psnr_list)) + " ssim_average:" + str(
sum(ssim_list) / len(ssim_list)) + " UIQM:" + str(
sum(uiqm_list) / len(uiqm_list))
return average
# print(gt_gray.size, enhanced_gray.size)
# ssim needs same sized images
enhanced_gray = ImageOps.grayscale(enhanced_image)
enhanced_image = np.asanyarray(enhanced_image)
enhanced_gray = np.asanyarray(enhanced_gray)
# FR_IQA
if os.path.isfile(gt_path):
ssim, mse = ssim_sk(gt_image, enhanced_image)
metrics["SSIM"][0] += ssim
metrics["SSIM"][1] += 1
psnr_val = psnr(gt_image, enhanced_image)
metrics["PSNR"][0] += psnr_val
metrics["PSNR"][1] += 1
metrics["MSE"][0] += mse
metrics["MSE"][1] += 1
try:
metrics["IE"][0] += entropy_sk(gt_image, enhanced_image)
metrics["IE"][1] += 1
except:
pass
try:
metrics["VIF"][0] += compute_vif(gt_image, enhanced_image)
metrics["VIF"][1] += 1
except:
dist2 = scipy.misc.imread(
dist2_file, flatten=True).astype(numpy.float32)
dist3 = scipy.misc.imread(
dist3_file, flatten=True).astype(numpy.float32)
dist4 = scipy.misc.imread(
dist4_file, flatten=True).astype(numpy.float32)
sharp = scipy.misc.imread(
sharp_file, flatten=True).astype(numpy.float32)
# print(ref_file + ' ' + dist_file + ':')
num += 1
# vifp1 = vifp.vifp_mscale(ref, dist)
ssim1 = ssim.ssim_exact(sharp / 255, dist1 / 255)
ssim2 = ssim.ssim_exact(sharp / 255, dist2 / 255)
ssim3 = ssim.ssim_exact(sharp / 255, dist3 / 255)
ssim4 = ssim.ssim_exact(sharp / 255, dist4 / 255)
psnr1 = psnr.psnr(sharp, dist1)
psnr2 = psnr.psnr(sharp, dist2)
psnr3 = psnr.psnr(sharp, dist3)
psnr4 = psnr.psnr(sharp, dist4)
# reco1 = reco.reco(ref / 255, dist / 255)
# vifp_sum += vifp1
# reco_sum += reco1
# ref_values.append(ref_file)
# dist_values.append(dist_file)
# vifp_values.append(vifp1)
# ssim_values.append(ssim1)
# psnr_values.append(psnr1)
# reco_values.append(reco1)
f.write("%s:\n" % sharp_file)
f.write("%f\t%f\t%f\t%f\n" %
(ssim1, ssim2, ssim3, ssim4))
if 'compress' in mode:
orig_img = img.copy()
res_img = compress(img, x_01, x_01, mu, lamb, m_ratio, display, method,
output)
img = res_img
if 'decompress' in mode:
res_img = decompress(img, x_01, x_01, mu, lamb, m_ratio, display,
method, output)
if orig_img is not None:
psnr_val = psnr(orig_img, res_img)
print(psnr_val)
p_name = os.path.splitext(os.path.basename(output))[0]
p_file = output.replace(p_name, p_name + 'psnr')
p_file = output.replace(
os.path.splitext(os.path.basename(output))[1], '.txt')
with open(p_file, 'w') as f:
f.write(str(psnr_val))
## img_width, img_height = img.shape[:2]
##
## img = cv2.medianBlur(img,3)
## cv2.imshow('image orig', img.astype(np.uint8))
## cv2.waitKey(0)
##
## img = img.astype(np.float32)
def generate():
args = get_arguments()
with open("./params.json", 'r') as f:
params = json.load(f)
if check_params(args, params) == False:
return
sess = tf.Session()
net = ESPCN(filters_size=params['filters_size'],
channels=params['channels'],
ratio=params['ratio'],
batch_size=1,
lr_size=params['lr_size'],
edge=params['edge'])
loss, images, labels = net.build_model()
lr_image = tf.placeholder(tf.uint8)
lr_image_data = misc.imread(args.lr_image)
lr_image_ycbcr_data = rgb2ycbcr(lr_image_data)
lr_image_y_data = lr_image_ycbcr_data[:, :, 0:1]
lr_image_cb_data = lr_image_ycbcr_data[:, :, 1:2]
lr_image_cr_data = lr_image_ycbcr_data[:, :, 2:3]
lr_image_batch = np.zeros((1, ) + lr_image_y_data.shape)
lr_image_batch[0] = lr_image_y_data
sr_image = net.generate(lr_image)
saver = tf.train.Saver()
try:
model_loaded = net.load(sess, saver, args.checkpoint)
except:
raise Exception(
"Failed to load model, does the ratio in params.json match the ratio you trained your checkpoint with?"
)
if model_loaded:
print("[*] Checkpoint load success!")
else:
print("[*] Checkpoint load failed/no checkpoint found")
return
sr_image_y_data = sess.run(sr_image, feed_dict={lr_image: lr_image_batch})
sr_image_y_data = shuffle(sr_image_y_data[0], params['ratio'])
sr_image_ycbcr_data = misc.imresize(
lr_image_ycbcr_data,
params['ratio'] * np.array(lr_image_data.shape[0:2]), 'bicubic')
edge = params['edge'] * params['ratio'] / 2
sr_image_ycbcr_data = np.concatenate(
(sr_image_y_data, sr_image_ycbcr_data[edge:-edge, edge:-edge, 1:3]),
axis=2)
sr_image_data = ycbcr2rgb(sr_image_ycbcr_data)
misc.imsave(args.out_path + '.png', sr_image_data)
if args.hr_image != None:
hr_image_data = misc.imread(args.hr_image)
model_psnr = psnr(hr_image_data, sr_image_data, edge)
print('PSNR of the model: {:.2f}dB'.format(model_psnr))
sr_image_bicubic_data = misc.imresize(
lr_image_data,
params['ratio'] * np.array(lr_image_data.shape[0:2]), 'bicubic')
misc.imsave(args.out_path + '_bicubic.png', sr_image_bicubic_data)
bicubic_psnr = psnr(hr_image_data, sr_image_bicubic_data, 0)
print('PSNR of Bicubic: {:.2f}dB'.format(bicubic_psnr))
print "Frame=%d VIFP=%f SSIM=%f" % (frame_num, vifp_value, ssim_value)
frame_num += 1
else:
# Inputs are image files
ref = scipy.misc.imread(ref_file, flatten=True).astype(numpy.float32)
dist = scipy.misc.imread(dist_file, flatten=True).astype(numpy.float32)
width, height = ref.shape[1], ref.shape[0]
print "Comparing %s to %s, resolution %d x %d" % (ref_file, dist_file,
width, height)
vifp_value = vifp.vifp_mscale(ref, dist)
print "VIFP=%f" % (vifp_value)
ssim_value = ssim.ssim_exact(ref / 255, dist / 255)
print "SSIM=%f" % (ssim_value)
# FIXME this is buggy, disable for now
# ssim_value2 = ssim.ssim(ref/255, dist/255)
# print "SSIM approx=%f" % (ssim_value2)
psnr_value = psnr.psnr(ref, dist)
print "PSNR=%f" % (psnr_value)
# niqe_value = niqe.niqe(dist/255)
# print "NIQE=%f" % (niqe_value)
reco_value = reco.reco(ref / 255, dist / 255)
print "RECO=%f" % (reco_value)
plt.figure(figsize=(8, 8))
for quality in range(0, 101, 1):
subprocess.check_call('gm convert %s -quality %d %s' %
(ref_file, quality, dist_file),
shell=True)
file_size = os.path.getsize(dist_file)
dist = scipy.misc.imread(dist_file, flatten=True).astype(numpy.float32)
quality_values.append(quality)
size_values.append(int(file_size / 1024))
vifp_values.append(vifp.vifp_mscale(ref, dist))
ssim_values.append(ssim.ssim_exact(ref / 255, dist / 255))
psnr_values.append(psnr.psnr(ref, dist))
# niqe_values.append( niqe.niqe(dist/255) )
reco_values.append(reco.reco(ref / 255, dist / 255))
plt.plot(quality_values, vifp_values, label='VIFP')
plt.plot(quality_values, ssim_values, label='SSIM')
# plt.plot(niqe_values, label='NIQE')
plt.plot(quality_values, reco_values, label='RECO')
plt.plot(quality_values, numpy.asarray(psnr_values) / 100.0, label='PSNR/100')
plt.legend(loc='lower right')
plt.xlabel('JPEG Quality')
plt.ylabel('Metric')
plt.savefig('jpg_demo_quality.png')
plt.figure(figsize=(8, 8))
import numpy as np
import cv2
from psnr import psnr
# img = cv2.imread('dataset/test/crop/070_im.png')
mask = cv2.imread('mask.png',0)
p = 0
l2 = 0
l1 = 0
for i in range(1,101):
img = cv2.imread('dataset/val/paris_eval_gt/%03d_im.png'%i)
img=cv2.resize(img,(128,128),interpolation=cv2.INTER_CUBIC)
dst = cv2.inpaint(img,mask,3,cv2.INPAINT_NS)
p = p+psnr(img[32:32+64,32:32+64,:].astype(np.float32),dst[32:32+64,32:32+64,:].astype(np.float32))
cv2.imwrite('result/test/old_all/%03d_im.png'%i,dst)
t = img[32:32+64,32:32+64,:].astype(np.float32)/127.5-dst[32:32+64,32:32+64,:].astype(np.float32)/127.5
l2 = l2 + np.mean(np.square(t))
t = np.abs(t)
l1 = l1 + np.mean(t)
print(l1/100.0)
print(l2/100.0)
print(p/100.0)
# reconst_img = scipy.misc.imread(reconst_path, flatten=False, model='RGB').astype(numpy.float32)
# cps_path =
name = os.path.splitext(os.path.basename(reconst_path))[0]
cps_path = glob.glob(os.path.join(stream_dir, model, name+'*'))[0]
stream_file_size = os.path.getsize(cps_path)
reconst_img = cv2.imread(reconst_path).astype(numpy.float32)
# print cps_img.shape #exit()
path_list.append(reconst_path)
size_values.append( int(stream_file_size/1024) ) ## kb
# vifp_values.append( vifp.vifp_mscale(ref_img, cps_img) )
ssim_values.append( ssim.ssim_exact(ref_img/255, reconst_img/255) )
psnr_values.append( psnr.psnr(ref_img, reconst_img) )
# reco_values.append( reco.reco(ref_img/255, cps_img/255) )
# cps_path_list = glob.glob(os.path.join(stream_dir, model, ref_name+'*'))
# # print ref_path
# for cps_path in cps_path_list:
# file_size = os.path.getsize(cps_path)
# # reconst_path = os.path.join(reconst_dir, model, ref_name) ## have to be same format with reference image
# print size_values;exit()
for i in range(len(size_values)):
# csv_writer.writerow
output_row=[os.path.basename(ref_path), os.path.basename(path_list[i]), str(ori_shape), str(ori_size), str(size_values[i]), str(float(ori_size)/size_values[i]), str(psnr_values[i]), str(ssim_values[i])]
def generate():
args = get_arguments()
with open("./params2.json", 'r') as f:
params = json.load(f)
if check_params(args, params) == False:
return
sess = tf.Session()
net = ESPCN(filters_size=params['filters_size'],
channels=params['channels'],
ratio=params['ratio'],
batch_size=1,
lr_size=params['lr_size'],
edge=params['edge'])
loss, images, labels = net.build_model()
files = [f for f in os.listdir(args.lr_image_dir) if os.path.isfile(os.path.join(args.lr_image_dir, f))]
saver = tf.train.Saver()
if net.load(sess, saver, args.checkpoint):
print("[*] Checkpoint load success!")
else:
print("[*] Checkpoint load failed/no checkpoint found")
return
frame_range = (87, 10000)
for fileName in files:
try:
ts = time()
frame_cnt = int(fileName[5:10])
if frame_cnt < frame_range[0] or frame_cnt > frame_range[1]:
print('Ignoring frame ' + str(frame_cnt))
continue
else:
print('start sr for frame ' + str(frame_cnt))
input_file = os.path.join(args.lr_image_dir, fileName)
output_file = os.path.join(args.out_path_dir, fileName)
lr_image = tf.placeholder(tf.uint8)
lr_image_data = misc.imread(input_file) # pip install pillow
lr_image_ycbcr_data = rgb2ycbcr(lr_image_data)
lr_image_y_data = lr_image_ycbcr_data[:, :, 0:1]
lr_image_cb_data = lr_image_ycbcr_data[:, :, 1:2]
lr_image_cr_data = lr_image_ycbcr_data[:, :, 2:3]
lr_image_batch = np.zeros((1,) + lr_image_y_data.shape)
lr_image_batch[0] = lr_image_y_data
print('preprocessed %d ms' % ((time()-ts)*1000))
ts = time()
sr_image = net.generate(lr_image)
print('network generated %d ms' % ((time()-ts)*1000))
ts = time()
sr_image_y_data = sess.run(sr_image, feed_dict={lr_image: lr_image_batch})
print('run %d ms' % ((time()-ts)*1000))
ts = time()
sr_image_y_data = shuffle(sr_image_y_data[0], args.ratio)
sr_image_ycbcr_data = misc.imresize(lr_image_ycbcr_data,
params['ratio'] * np.array(lr_image_data.shape[0:2]),
'bicubic')
edge = params['edge'] * params['ratio'] / 2
sr_image_ycbcr_data = np.concatenate((sr_image_y_data, sr_image_ycbcr_data[edge:-edge,edge:-edge,1:3]), axis=2)
print('mixed %d ms' % ((time()-ts)*1000))
ts = time()
sr_image_data = ycbcr2rgb(sr_image_ycbcr_data)
#sr_image_data = sr_image_ycbcr_data
print('converted %d ms' % ((time()-ts)*1000))
ts = time()
misc.imsave(output_file, sr_image_data)
print(output_file + ' generated %d ms' % ((time()-ts)*1000))
ts = time()
if args.hr_image_dir != None:
hr_image_path = os.path.join(args.hr_image_dir, fileName)
hr_image_data = misc.imread(hr_image_path)
model_psnr = psnr(hr_image_data, sr_image_data, edge)
print('PSNR of the model: {:.2f}dB'.format(model_psnr))
sr_image_bicubic_data = misc.imresize(lr_image_data,
params['ratio'] * np.array(lr_image_data.shape[0:2]),
'bicubic')
bicubic_path = os.path.join(args.out_path_dir, fileName + '_bicubic.png')
misc.imsave(bicubic_path, sr_image_bicubic_data)
bicubic_psnr = psnr(hr_image_data, sr_image_bicubic_data, 0)
print('PSNR of Bicubic: {:.2f}dB'.format(bicubic_psnr))
except IndexError:
print('Index error caught')
except IOError:
print('Cannot identify image file: ' + fileName)
except ValueError:
print('Cannot parse file name: ' + fileName)
ssim_value = ssim.ssim(ref, dist)
print "Frame=%d VIFP=%f SSIM=%f" % (frame_num, vifp_value, ssim_value)
frame_num += 1
else:
# Inputs are image files
ref = scipy.misc.imread(ref_file, flatten=True).astype(numpy.float32)
dist = scipy.misc.imread(dist_file, flatten=True).astype(numpy.float32)
width, height = ref.shape[1], ref.shape[0]
print "Comparing %s to %s, resolution %d x %d" % (ref_file, dist_file, width, height)
vifp_value = vifp.vifp_mscale(ref, dist)
print "VIFP=%f" % (vifp_value)
ssim_value = ssim.ssim_exact(ref / 255, dist / 255)
print "SSIM=%f" % (ssim_value)
# FIXME this is buggy, disable for now
# ssim_value2 = ssim.ssim(ref/255, dist/255)
# print "SSIM approx=%f" % (ssim_value2)
psnr_value = psnr.psnr(ref, dist)
print "PSNR=%f" % (psnr_value)
# niqe_value = niqe.niqe(dist/255)
# print "NIQE=%f" % (niqe_value)
reco_value = reco.reco(ref / 255, dist / 255)
print "RECO=%f" % (reco_value)
_weight = weight(i, j, i + x, j + y) * _flag
tem_weight += _weight
_weight = _weight.repeat(3).reshape(N1, N1, 3) # 扩充成三维
tem_img += _n_img[i + x:i + N1 + x,
j + y:j + N1 + y, :] * _weight
tem_weight += (tem_weight == 0) # 避免除以零
tem_weight = tem_weight.repeat(3).reshape(N1, N1, 3) # 扩充成三维
final_img[i:i + N1, j:j + N1:, :] += tem_img / tem_weight
num_img[i:i + N1, j:j + N1:, :] += tem_weight > 0
num_img += (num_img == 0) # 避免除以零
final_img = final_img / num_img # 每个像素点都进行了多次估计，直接平均
final_img = final_img[blank:H + blank, blank:W + blank, :] + n_img
outFileName = 'result/NLM/' + noiseFile[:-4] + '_' + str(N1) + '_' + str(
N2) + '_' + str(sigma) + '_' + str(h) + noiseFile[-4:]
cv2.imwrite(outFileName, final_img)
final_img = cv2.imread(outFileName)
print("\nP=%d B=%d sigma=%d h = %f" % (N1, N2, sigma, h),
psnr(img, final_img))
end = time.clock()
print('time:%d' % (end - start))
# The results
# noiseFile = 'Lena_RS1_0.5.png'
# P=9 B=15 sigma=0 h = 1.000000 [33.802286078114356, 34.37252084511732, 37.22837467495257, 35.13439386606141] time:569
#
# noiseFile = 'Lena_RS1_0.3.png'
# P=15 B=21 sigma=0 h = 1.000000 [31.073116649651563, 31.0775574382444, 33.684435010928624, 31.945036366274863] time:785
def testFromH5Folder(input_dataset_dir, output_dataset_dir, upscale_factor, input_field_name, model_name, cuda_flag, if_save):
h5_file_list = [x for x in listdir(input_dataset_dir) if is_h5_file(x)]
h5_file_list.sort()
h5_len = len(h5_file_list)
model_PSNR = np.zeros(h5_len)
model_SSIM = np.zeros(h5_len)
bicubic_PSNR = np.zeros(h5_len)
bicubic_SSIM = np.zeros(h5_len)
model_time = np.zeros(h5_len)
# Load SR Model
model = Net_SRCNN(upscale_factor=UPSCALE_FACTOR)
model = torch.load(MODEL_NAME)
if torch.cuda.is_available() & cuda_flag:
model = model.cuda()
else:
model = model.cpu()
if not os.path.exists(output_dataset_dir):
os.makedirs(output_dataset_dir)
for idx in tqdm(range(h5_len), desc='SR of upscale factor = ' + str(upscale_factor)):
h5_file_name = h5_file_list[idx]
h5_file = h5py.File(input_dataset_dir + '/' + h5_file_name, 'r')
img_HR_y = h5_file['HR'].value
img_LR_y = h5_file['LR'].value
img_LR_bic_y = h5_file['LR_bic_y'].value
img_LR_bic_cb = h5_file['LR_bic_cb'].value
img_LR_bic_cr = h5_file['LR_bic_cr'].value
img_HR_RGB = h5_file['HR_RGB'].value
img_HR_y = img_HR_y.astype(np.float32)
img_LR_y = img_LR_y.astype(np.float32)
img_LR_bic_y = img_LR_bic_y.astype(np.float32)
img_LR_bic_cb = img_LR_bic_cb.astype(np.float32)
img_LR_bic_cr = img_LR_bic_cr.astype(np.float32)
img_HR_RGB = img_HR_RGB.astype(np.float32)
img_HR_y = img_HR_y / 255.0
img_LR_y = img_LR_y / 255.0
img_HR_RGB = img_HR_RGB / 255.0
if input_field_name == 'LR':
img_LR_4d = img_LR_y.reshape(1, 1, img_LR_y.shape[0], img_LR_y.shape[1])
elif input_field_name == 'LR_bic_y':
img_LR_4d = img_LR_bic_y.reshape(1, 1, img_LR_bic_y.shape[0], img_LR_bic_y.shape[1])
image = Variable(torch.from_numpy(img_LR_4d))
if torch.cuda.is_available() & cuda_flag:
image = image.cuda()
start = time.time()
target = model(image)
end = time.time()
target = target.cpu()
img_HR_y_net = target.data[0][0].numpy()
if if_save:
img_HR_ycbcr_net = np.zeros(img_HR_RGB.shape)
img_HR_ycbcr_net[:,:,0] = img_HR_y_net
img_HR_ycbcr_net[:,:,1] = img_LR_bic_cb
img_HR_ycbcr_net[:,:,2] = img_LR_bic_cr
img_HR_ycbcr_net = img_HR_ycbcr_net.clip(0.0, 1.0)
img_HR_RGB_net = ycbcr2rgb(img_HR_ycbcr_net)
img_HR_RGB_net = img_HR_RGB_net.clip(0.0, 1.0)
img_HR_RGB_net *= 255.0
img_HR_BGR_net = rgb2bgr(img_HR_RGB_net)
image_name = 'img_' + str(idx) + '_net.png'
cv2.imwrite(output_dataset_dir + '/' + image_name, img_HR_BGR_net.astype(np.uint8))
# Compute Stat
model_PSNR[idx] = psnr((img_HR_y*255.0).astype(int), (img_HR_y_net*255.0).astype(int))
model_SSIM[idx] = ssim((img_HR_y*255.0).astype(int), (img_HR_y_net*255.0).astype(int))
bicubic_PSNR[idx] = psnr((img_HR_y*255.0).astype(int), (img_LR_bic_y*255.0).astype(int))
bicubic_SSIM[idx] = ssim((img_HR_y*255.0).astype(int), (img_LR_bic_y*255.0).astype(int))
model_time[idx] = (end-start)
print("===> Test on" + input_dataset_dir +" Complete: Model PSNR: {:.4f} dB, Model SSIM: {:.4f} , Bicubic PSNR: {:.4f} dB, Bicubic SSIM: {:.4f} , Average time: {:.4f}"
.format(np.average(model_PSNR), np.average(model_SSIM), np.average(bicubic_PSNR), np.average(bicubic_SSIM), np.average(model_time)*1000))
int(opt.imageSize / 4):int(opt.imageSize / 4 +
opt.imageSize / 2)] = fake.data
vutils.save_image(real_cpu, 'val_real_samples.png', normalize=True)
vutils.save_image(input_cropped.data,
'val_cropped_samples.png',
normalize=True)
vutils.save_image(recon_image.data, 'val_recon_samples.png', normalize=True)
p = 0
l1 = 0
l2 = 0
fake = fake.data.numpy()
real_center = real_center.data.numpy()
from psnr import psnr
import numpy as np
t = real_center - fake
l2 = np.mean(np.square(t))
l1 = np.mean(np.abs(t))
real_center = (real_center + 1) * 127.5
fake = (fake + 1) * 127.5
for i in range(opt.batchSize):
p = p + psnr(real_center[i].transpose(1, 2, 0), fake[i].transpose(1, 2, 0))
print(l2)
print(l1)
print(p / opt.batchSize)
#loading the image
mat_f = loadmat(file_name)
true = mat_f['img'] #original image
img = mat_f['imm256'] #Cai's processed Image
M, N = img.shape
img = img.reshape(1, M, N, 1)
model_file = 'sigma' + str(
sigma) + 'p' + pp + '.mat' #name of the model file to be used
load_from = os.path.join('models', model_file)
session_conf = tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False)
sess = tf.Session(config=session_conf)
cnn = denoiser_net(is_load=True, filename=load_from)
global_step = tf.Variable(0, name="global_step", trainable=False)
sess.run(tf.global_variables_initializer())
out, _ = sess.run([cnn.out, global_step], feed_dict={cnn.input_x: img})
out = out.reshape(M / 2, N / 2)
out[out > 1] = 1
out[out < 0] = 0
mat_f['cnn'] = out
savemat(file_name, mat_f) #mat_f now contains the cnn denoised image as well.
print("PSNR: ")
print(psnr(out, true, 1.0))
args = parse_args()
pred_imgs_path = os.path.join(args.output_path,
args.model_flag + '_' + args.exp_num)
imgs_list_file = '/home/chenli/code/SRN-Deblur/test_datalist.txt'
imgs_datalist = open(imgs_list_file, 'rt').read().splitlines()
imgs_datalist = list(map(lambda x: x.split(' '), imgs_datalist))
imgsName = [x[1] for x in imgs_datalist]
psnr_values = []
ssim_values = []
for imgName in imgsName:
real = scipy.misc.imread(imgName, flatten=True).astype(numpy.float32)
split_name = imgName.split('/')
pred_imgName = os.path.join(pred_imgs_path, split_name[-3], split_name[-2],
split_name[-1])
pred = scipy.misc.imread(pred_imgName, flatten=True).astype(numpy.float32)
ssim_value = ssim.ssim_exact(real / 255, pred / 255)
psnr_value = psnr.psnr(real, pred)
print('psnr:%.5f ssim:%.5f' % (psnr_value, ssim_value))
ssim_values.append(ssim_value)
psnr_values.append(psnr_value)
final_psnr_value = numpy.mean(psnr_values)
final_ssim_value = numpy.mean(ssim_values)
print('final psnr:%.5f final ssim:%.5f' % (final_psnr_value, final_ssim_value))
def main():
'''
训练时时并行的，测试时也应当并行，不然会报告如下的错误：
Missing key(s) in state_dict: ...(如：conv1.weight)
'''
print('testing processing....')
#加载模型
test_model = VRCNN(opt.upscale_factor)
test_model = torch.nn.DataParallel(test_model,device_ids=gpus_list,output_device=gpus_list[1])
test_model = test_model.cuda(gpus_list[0])
print('---------- Networks architecture -------------')
print_network(test_model)
print('----------------------------------------------')
#加载预训练模型
model_name = os.path.join(opt.model_save_folder,opt.exp_name,opt.test_model)
print('model_name=',model_name)
if os.path.exists(model_name):
pretrained_dict=torch.load(model_name,map_location=lambda storage, loc: storage)
model_dict=test_model.state_dict()
# 1. filter out unnecessary keys
pretrained_dict = {k: v for k, v in pretrained_dict.items() if k in model_dict}
# 2. overwrite entries in the existing state dict
model_dict.update(pretrained_dict)
test_model.load_state_dict(model_dict)
print('Pre-trained SR model is loaded.')
if not os.path.exists(opt.pre_result):
os.mkdir(opt.pre_result)
with open(opt.train_log + '/psnr_ssim-xr-200.txt', 'a') as psnr_ssim:
with torch.no_grad():
ave_psnr = 0
ave_ssim = 0
single_ave_psnr = 0
single_ave_ssim = 0
numb = 2
valSet = ValidationsetLoader(opt.val_dataset_hr,opt.val_dataset_lr)
valLoader = DataLoader(dataset=valSet,batch_size=opt.test_val_batchSize,shuffle=False)
val_bar = tqdm(valLoader)
for data in val_bar:
test_model.eval()
# dual_net.eval()
batch_lr_y, label, SR_cb,SR_cr,idx,bicubic_restore = data
batch_lr_y,label = Variable(batch_lr_y).cuda(gpus_list[0]), Variable(label).cuda(gpus_list[0])
output = test_model(batch_lr_y)
SR_ycbcr = np.concatenate((np.array(output.squeeze(0).data.cpu()), SR_cb, SR_cr), axis=0).transpose(1,2,0)
SR_rgb = ycbcr2rgb(SR_ycbcr) * 255.0
SR_rgb = np.clip(SR_rgb, 0, 255)
SR_rgb = ToPILImage()(SR_rgb.astype(np.uint8))
#ToTensor() ---image(0-255)==>image(0-1), (H,W,C)==>(C,H,W)
SR_rgb = ToTensor()(SR_rgb)
#将给定的Tensor保存成image文件。如果给定的是mini-batch tensor，那就用make-grid做成雪碧图，再保存。与utils.make_grid()配套使用
if not os.path.exists(opt.pre_result+'/'+opt.exp_name):
os.mkdir(opt.pre_result+'/'+opt.exp_name)
utils.save_image(SR_rgb, opt.pre_result+'/' +opt.exp_name +'/' + 'my'+str(numb).rjust(3,'0')+'.png')
numb = numb + 1
psnr_value = psnr(np.array(torch.squeeze(label).data.cpu())*255,np.array(torch.squeeze(output).data.cpu())*255)
ave_psnr = ave_psnr + psnr_value
single_ave_psnr = single_ave_psnr + psnr_value
ssim_value = calculate_ssim(np.array(torch.squeeze(label).data.cpu())*255,np.array(torch.squeeze(output).data.cpu())*255)
ave_ssim = ave_ssim + ssim_value
single_ave_ssim = single_ave_ssim + ssim_value
val_bar.set_description('===>{}th video {}th frame, wsPSNR:{:.4f} dB,wsSSIM:{:.6f}'.format(idx // 98 + 1,idx % 98 + 1,psnr_value,ssim_value))
if idx == 293 or idx == 97 or idx == 195 or idx == 391:
print("===> {}th video Avg. wsPSNR: {:.4f} dB".format(idx // 98+1,single_ave_psnr / 98))
print("===> {}th video Avg. wsSSIM: {:.6f}".format(idx // 98+1,single_ave_ssim / 98))
psnr_ssim.write('===>{}th video avg wsPSNR:{:.4f} dB,wsSSIM:{:.6f}\n'.format(idx // 98+1,single_ave_psnr / 98,single_ave_ssim / 98))
single_ave_psnr = 0
single_ave_ssim = 0
print("===> All Avg. wsPSNR: {:.4f} dB".format(ave_psnr / len(valLoader)))
print("===> ALL Avg. wsSSIM: {:.6f}".format(ave_ssim / len(valLoader)))
psnr_ssim.write('===>all videos avg wsPSNR:{:.4f} dB,wsSSIM:{:.6f}\n'.format(ave_psnr / len(valLoader),ave_ssim / len(valLoader)))
print('testing finished!')
dist_file_dir = os.path.join(dist_path, dist_file) #参考文件绝对路径
if os.path.isdir(dist_file_dir): #如果是文件夹，跳过
continue
dist_img = scipy.misc.imread(dist_file_dir,
flatten=True).astype(numpy.float32)
#filename = os.path.splitext(dist_file)[0] + os.path.splitext(dist_file)[1] #ref filename
filename = os.path.splitext(dist_file)[0] + '.png' #ref filename
ref_img = scipy.misc.imread(ref_path + '\\' + filename,
flatten=True).astype(
numpy.float32) #读取参考图像对应的测试图像
psnr_data = psnr.psnr(ref_img, dist_img) #psnr指标
ssim_data = ssim.ssim_exact(ref_img / 255, dist_img / 255) #ssim指标
dist_img = cv2.imread(dist_path + '\\' + os.path.splitext(dist_file)[0] +
'.png') #UIQM指标
uiqm_data = uqim_utils.getUIQM(dist_img) #UIQM
print("img:" + str(filename) + " psnr:" + str(psnr_data) + " ssim:" +
str(ssim_data) + " UIQM:" + str(uiqm_data))
data = str(filename) + " psnr:" + str(psnr_data) + " ssim:" + str(
ssim_data) + " UIQM:" + str(uiqm_data)
psnr_list.append(psnr_data)
ssim_list.append(ssim_data)
uiqm_list.append(uiqm_data)
ori_shape = ref_img.shape
ori_size = 1
for dim in ori_shape:
ori_size = ori_size * dim
ori_size = int(ori_size / 1024)
print cps_path_list
for cps_path in cps_path_list:
file_size = os.path.getsize(cps_path)
cps_img = scipy.misc.imread(cps_path, flatten=False,
mode='RGB').astype(numpy.float32)
print cps_img.shape #exit()
size_values.append(int(file_size / 1024)) ## kb
# vifp_values.append( vifp.vifp_mscale(ref_img, cps_img) )
ssim_values.append(ssim.ssim_exact(ref_img / 255, cps_img / 255))
psnr_values.append(psnr.psnr(ref_img, cps_img))
# reco_values.append( reco.reco(ref_img/255, cps_img/255) )
# print size_values;exit()
for i in range(len(size_values)):
csv_writer.writerow
output_row = [
os.path.basename(ref_path),
os.path.basename(cps_path_list[i]),
str(ori_shape),
str(ori_size),
str(size_values[i]),
str(float(ori_size) / size_values[i]),
str(psnr_values[i]),
str(ssim_values[i])
]
csv_writer.writerow(output_row)
for j in range(col):
if mask.mask[i][j] == 1:
img_r[i][j] = img[i][j][0]
img_g[i][j] = img[i][j][1]
img_b[i][j] = img[i][j][2]
cv2.imwrite(data_path.image_r_path, img_r)
cv2.imwrite(data_path.image_g_path, img_g)
cv2.imwrite(data_path.image_b_path, img_b)
if __name__ == '__main__':
color_to_grey.color_to_grey_1()
bm3d_r()
bm3d_g()
bm3d_b()
refresh()
grey_to_color.grey_to_color()
print("for iteration 1:")
psnr.psnr()
for i in range(2, iter_time + 1):
Threshold_Hard3D = Threshold_Hard3D * 0.6
color_to_grey.color_to_grey_2()
bm3d_r()
bm3d_g()
bm3d_b()
refresh()
grey_to_color.grey_to_color()
print("for iteration %d:" % i)
psnr.psnr()
weight_img[i: i + N1, j: j + N1:, :] += tem_weight > 0 # 记录每个像素点被估计的次数，最后平均
# 用凯撒窗的代码，对应上面两行
# final_img[i: i + N1, j: j + N1:, :] += tem_img / tem_weight * Kaiser
# weight_img[i: i + N1, j: j + N1:, :] += Kaiser
weight_img += (weight_img == 0) # 避免除以零
final_img = final_img / weight_img # 每个像素点都进行了多次估计，直接平均
final_img = final_img[blank: H + blank, blank: W + blank, :]
_bias = np.repeat(f, 3).reshape(H, W, 3) * bias # 像素点rgb都等于零对应YCbCr等于(16, 128, 128),修正final_img
final_img = yuv_to_bgr(final_img + _bias) + n_img
outFileName = 'result/NLM/' + noiseFile[:-4] + '_' + str(N1) + '_' + str(N2) + '_' + str(sigma) + '_' + str(
h) + noiseFile[-4:]
cv2.imwrite(outFileName, final_img)
final_img = cv2.imread(outFileName)
print("\nP=%d B=%d sigma=%d h = %f" % (N1, N2, sigma, h), psnr(img, final_img))
end = time.clock()
print('time:%d' % (end - start))
# The results
# noiseFile = 'Lena_RS1_0.5.png' 用YUV三个通道计算权重,无凯撒窗
# P=9 B=15 sigma=0 h = 1.000000 [33.93627406850448, 34.510636345073365, 37.35516440584259, 35.26735827314015] time:675
#
# noiseFile = 'Lena_RS1_0.3.png'
# P=15 B=21 sigma=0 h = 1.000000 [31.195591935621792, 31.183651976211202, 33.775171616441405, 32.05147184275813]time:685
#
# noiseFile = 'test.png'(150*150取自Lena.png)
# 用RGB计算权重
# P=9 B=15 sigma=0 h = 1.000000 [31.411476278247996, 32.27519874157173, 35.13957722424206, 32.942084081353926]
# 仅Y通道计算权重
# P=9 B=15 sigma=0 h = 1.000000 [31.325063582425457, 32.3117315982166, 35.16094564206194, 32.93258027423467]
targetRGB[1, :, :] = (targetRGB[1, :, :] * 0.5) + 0.5
targetRGB[2, :, :] = (targetRGB[2, :, :] * 0.5) + 0.5
imageRGB_rand_recon[imageRGB_rand_recon > 1] = 1
imageRGB_adap_recon[imageRGB_adap_recon > 1] = 1
corrupt_image_rand[corrupt_image_rand > 1] = 1
corrupt_image_adap[corrupt_image_adap > 1] = 1
imageRGB_rand_recon[imageRGB_rand_recon < 0] = 0
imageRGB_adap_recon[imageRGB_adap_recon < 0] = 0
corrupt_image_rand[corrupt_image_rand < 0] = 0
corrupt_image_adap[corrupt_image_adap < 0] = 0
# Compute Stat here
netE_rand_PSNR[img_idx] = psnr(
(targetRGB * 255.0).astype(int),
(imageRGB_rand_recon * 255.0).astype(int))
netE_adap_PSNR[img_idx] = psnr(
(targetRGB * 255.0).astype(int),
(imageRGB_adap_recon * 255.0).astype(int))
netE_rand_MSE[img_idx] = mse((targetRGB * 255.0).astype(int),
(imageRGB_rand_recon * 255.0).astype(int))
netE_adap_MSE[img_idx] = mse((targetRGB * 255.0).astype(int),
(imageRGB_adap_recon * 255.0).astype(int))
rand_corrupt_PSNR[img_idx] = psnr(
(targetRGB * 255.0).astype(int),
(corrupt_image_rand * 255.0).astype(int))
adap_corrupt_PSNR[img_idx] = psnr(
(targetRGB * 255.0).astype(int),
niqe_values = []
reco_values = []
plt.figure(figsize=(8, 8))
for quality in range(0, 101, 1):
subprocess.check_call('gm convert %s -quality %d %s'%(ref_file, quality, dist_file), shell=True)
file_size = os.path.getsize(dist_file)
dist = scipy.misc.imread(dist_file, flatten=True).astype(numpy.float32)
quality_values.append( quality )
size_values.append( int(file_size/1024) )
vifp_values.append( vifp.vifp_mscale(ref, dist) )
ssim_values.append( ssim.ssim_exact(ref/255, dist/255) )
psnr_values.append( psnr.psnr(ref, dist) )
# niqe_values.append( niqe.niqe(dist/255) )
reco_values.append( reco.reco(ref/255, dist/255) )
plt.plot(quality_values, vifp_values, label='VIFP')
plt.plot(quality_values, ssim_values, label='SSIM')
# plt.plot(niqe_values, label='NIQE')
plt.plot(quality_values, reco_values, label='RECO')
plt.plot(quality_values, numpy.asarray(psnr_values)/100.0, label='PSNR/100')
plt.legend(loc='lower right')
plt.xlabel('JPEG Quality')
plt.ylabel('Metric')
plt.savefig('jpg_demo_quality.png')
plt.figure(figsize=(8, 8))
