def train_test_vdsr(is_train, scale=4):
from MODELS.VDSR import VDSR
model = VDSR(scale=scale,
sess=tf.Session(),
save_path='./TRAINED_MODEL/VDSRx{}/'.format(scale))
if is_train:
train_img, label_img, test_img, test_label_img = prepareImageData(
'./DATABASE/DIV2K/', 0.2, scale=4, subimg_h=32, subimg_w=32)
print('\n\nSRModel (' + 'VDSR' + ' x ' + str(scale) +
') Trainning ... ...')
model.train(train_img,
label_img,
test_img,
test_label_img,
Epoch=int(5e4),
iter_view=500)
else:
or_image = Image.open('./DATABASE/Set5/butterfly_GT.bmp').convert(
"YCbCr")
lr_image = DownScale(or_image, or_image.size, scale=4)
lr_image = np.array(lr_image)
sr_image = model.test(image=lr_image)
sr_image.show()
print(
'----<The PSNR, SSIM between HR image and SR image is: PSNR:{:.5f}, SSIM:{:.5f}>----'
.format(psnr(np.array(or_image), np.array(sr_image)),
ssim(np.array(or_image), np.array(sr_image))))
def test_SpCodeVDSR(scale, fn, downscale_fn, config):
from MODELS.VDSR import VDSR
model = VDSR(config=config,
sess=tf.Session(),
save_path='./TRAINED_MODEL/SpCode-VDSRx{}/'.format(scale))
model.scale = scale
hr_image = np.asarray(Image.open(fn).convert("YCbCr"))
sr_image = model.SpCodeVDSRSR(fn, downscale_fn)
print(
'----<The PSNR, SSIM between HR image and SR image is: PSNR:{:.5f}, SSIM:{:.5f}>----'
.format(psnr(hr_image, sr_image), ssim(hr_image, sr_image)))
def ksvd(Y, A0, K, n, Acc):
num_signal = Y.shape[1] #待分解的信号个数
#A = overDct(m,n)
#A = overDct(529,529)
#A = A[0:512,0:512]
A = A0
IterK = 0
#由于的DCT字典所以不必归一化
Err = 10**7
PSNR = 0
X = np.matrix(np.zeros((n, num_signal)))
while (IterK < Acc):
IterK += 1
#lsomp做稀疏表示
for i in range(0, num_signal):
b = Y[:, i]
rec = lsomp(A, b, K)
X[:, i] = rec
#用KSVD做字典学习
for j0 in range(0, n):
Xj0 = X[j0, :]
Xj0 = np.array(Xj0)
Xj0.shape = [num_signal]
InstacnceSet = np.nonzero(Xj0)
InstacnceSet = np.array(InstacnceSet)
InstacnceSet.shape = [InstacnceSet.shape[1]]
InstacnceSet = list(InstacnceSet)
if (max(np.abs(Xj0)) == 0):
Err = Y - A * X
Err = np.diagonal(np.transpose(Err) * Err)
Err = list(Err)
i = Err.index(max(Err))
A[:, j0] = Y[:, i]
A[:, j0] = A[:, j0] / np.linalg.norm(
np.transpose(A[:, j0]) * A[:, j0])
A[:, j0] = A[:, j0] / (np.sign(A[1, j0]))
else:
tempX = X[:, InstacnceSet]
tempX[j0, :] = 0
Ej0 = Y[:, InstacnceSet] - A * tempX
U, Delta, V = np.linalg.svd(Ej0)
A[:, j0] = U[:, 0]
X[j0, InstacnceSet] = Delta[0] * V[0, :]
PSNR = psnr(np.array(Y), np.array(A * X))
curt = datetime.datetime.now()
print('The No.' + str(IterK) + ' iteration.' + 'PSNR = ' + str(PSNR) +
' time:' + str(curt))
return A, X, PSNR
def test_vdsr(scale, fn, downscale_fn, config):
from MODELS.VDSR import VDSR
model = VDSR(config=config,
sess=tf.Session(),
save_path='./TRAINED_MODEL/VDSRx{}/'.format(scale))
model.scale = scale
or_image = Image.open(fn).convert("YCbCr")
lr_image = downscale_fn(or_image, or_image.size, scale=scale)
lr_image = np.array(lr_image)
sr_image = model.test(image=lr_image)
sr_image.show()
print(
'----<The PSNR, SSIM between HR image and SR image is: PSNR:{:.5f}, SSIM:{:.5f}>----'
.format(psnr(np.array(or_image), np.array(sr_image)),
ssim(np.array(or_image), np.array(sr_image))))
def mod(Y, A0, K, n):
#A = overDct(Y.shape[0],n)
A = A0
num_signal = Y.shape[1]
IterK = 0
X = np.matrix(np.zeros((n, num_signal)))
PSNR = 0
while (IterK < 40):
IterK += 1
for i in range(num_signal):
rec = lsomp(A, Y[:, i], K)
X[:, i] = rec
#用MOD
A = Y * X.T * (X * X.T).I
PSNR = psnr(np.array(Y), np.array(A * X))
curt = datetime.datetime.now()
print('The No.' + str(IterK) + ' iteration.' + 'PSNR = ' + str(PSNR) +
' time:' + str(curt))
return A, X, PSNR
def train_test_SpCodeVDSR(is_train, scale):
from MODELS.VDSR import VDSR
model = VDSR(scale=scale,
sess=tf.Session(),
save_path='./TRAINED_MODEL/SpCode-VDSRx{}/'.format(scale))
if is_train:
x_test, y_test, x_train, y_train, train_size, test_size = prepareSparseData(
'./DATABASE/TrainDatax4.npz', 0.2)
print('\n\SpCode- ( ' + 'VDSR' + ' x ' + str(scale) +
' ) Trainning ... ...')
model.train(x_train,
y_train,
x_test,
y_test,
Epoch=int(5e4),
iter_view=500)
else:
filename = './DATABASE/Set5/woman_GT.bmp'
hr_image = np.asarray(Image.open(filename).convert("YCbCr"))
sr_image = model.SpCodeVDSRSR(filename)
print(
'----<The PSNR, SSIM between HR image and SR image is: PSNR:{:.5f}, SSIM:{:.5f}>----'
.format(psnr(hr_image, sr_image), ssim(hr_image, sr_image)))