def main():
# ----------------------------------------
# load kernels
# ----------------------------------------
motion_id = 7
PSF_grid = np.load('./data/Motion_PSF_{}.npz'.format(motion_id))['PSF']
PSF_grid = PSF_grid.astype(np.float32)
gx, gy = PSF_grid.shape[:2]
k_tensor = []
for yy in range(gy):
for xx in range(gx):
PSF_grid[xx, yy] = PSF_grid[xx, yy] / np.sum(PSF_grid[xx, yy],
axis=(0, 1))
k_tensor.append(util.single2tensor4(PSF_grid[xx, yy]))
k_tensor = torch.cat(k_tensor, dim=0)
inv_weight = util_deblur.get_inv_spatial_weight(k_tensor)
# ----------------------------------------
# load model
# ----------------------------------------
stage = 8
model_code = 'iter4700'
global_iter = 4700
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = net(n_iter=stage,
h_nc=64,
in_nc=4,
out_nc=3,
nc=[64, 128, 256, 512],
nb=2,
act_mode="R",
downsample_mode='strideconv',
upsample_mode="convtranspose")
pre_files = '/home/xiu/databag/deblur/models/motion/uabcnet_{}_{}.pth'.format(
motion_id, model_code)
model.load_state_dict(torch.load(pre_files), strict=True)
model.train()
for _, v in model.named_parameters():
v.requires_grad = True
model = model.to(device)
# ----------------------------------------
# load training data
# ----------------------------------------
imgs = glob.glob('/home/xiu/databag/deblur/images/*/**.png',
recursive=True)
imgs.sort()
# ----------------------------------------
# positional lambda\mu for HQS
# ----------------------------------------
ab_buffer = np.ones((gx, gy, 2 * stage, 3), dtype=np.float32) * 0.1
#ab_buffer[:,:,0,:] = 0.01
#ab_buffer = np.loadtxt('/home/xiu/databag/deblur/models/motion/ab_{}.txt'.format(model_code)).astype(np.float32).reshape(gx,gy,stage*2,3)
ab = torch.tensor(ab_buffer, device=device, requires_grad=True)
# ----------------------------------------
# build optimizer
# ----------------------------------------
params = []
params += [{"params": [ab], "lr": 0.0005}]
for key, value in model.named_parameters():
params += [{"params": [value], "lr": 0.0001}]
optimizer = torch.optim.Adam(params, lr=0.0001, betas=(0.9, 0.999))
scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=300,
gamma=0.9)
#64x3 = 192
patch_size = [128, 128]
expand = PSF_grid.shape[2] // 2
patch_num = [2, 2]
batch_num = 1
vis = True
#weight for each patch based on the spectral radius
loss_weight = torch.ones((3, gx, gy), device=device)
loss_weight.requires_grad = False
running = True
while running:
#alpha.beta
img_idx = np.random.randint(len(imgs))
img = imgs[img_idx]
img_H = cv2.imread(img)
w, h = img_H.shape[:2]
loss = 0
save_image = []
for _ in range(batch_num):
#focus on the edges
mode = np.random.randint(8)
px_start = np.random.randint(0, gx - patch_num[0] + 1)
py_start = np.random.randint(0, gy - patch_num[1] + 1)
if mode == 0:
px_start = 0
if mode == 1:
px_start = gx - patch_num[0]
if mode == 2:
py_start = 0
if mode == 3:
py_start = gy - patch_num[1]
x_start = np.random.randint(
0, w - patch_size[0] * patch_num[0] - expand * 2 + 1)
y_start = np.random.randint(
0, h - patch_size[1] * patch_num[1] - expand * 2 + 1)
PSF_patch = PSF_grid[px_start:px_start + patch_num[0],
py_start:py_start + patch_num[1]]
patch_H = img_H[x_start:x_start+patch_size[0]*patch_num[0]+expand*2,\
y_start:y_start+patch_size[1]*patch_num[1]+expand*2]
patch_L = util_deblur.blockConv2d(patch_H, PSF_patch, expand)
block_expand = max(patch_size[0] // 8, expand)
patch_L_wrap = util_deblur.wrap_boundary_liu(
patch_L, (patch_size[0] * patch_num[0] + block_expand * 2,
patch_size[1] * patch_num[1] + block_expand * 2))
patch_L_wrap = np.hstack(
(patch_L_wrap[:, -block_expand:, :],
patch_L_wrap[:, :patch_size[1] * patch_num[1] +
block_expand, :]))
patch_L_wrap = np.vstack(
(patch_L_wrap[-block_expand:, :, :],
patch_L_wrap[:patch_size[0] * patch_num[0] +
block_expand, :, :]))
x = util.uint2single(patch_L_wrap)
x = util.single2tensor4(x)
x_gt = util.uint2single(patch_H[expand:-expand, expand:-expand])
x_gt = util.single2tensor4(x_gt)
inv_weight_patch = torch.ones_like(x_gt)
k_local = []
for h_ in range(patch_num[1]):
for w_ in range(patch_num[0]):
inv_weight_patch[0, 0, w_ * patch_size[0]:(w_ + 1) *
patch_size[0],
h_ * patch_size[1]:(h_ + 1) *
patch_size[1]] = inv_weight[w_ + h_ *
patch_num[0],
0]
inv_weight_patch[0, 1, w_ * patch_size[0]:(w_ + 1) *
patch_size[0],
h_ * patch_size[1]:(h_ + 1) *
patch_size[1]] = inv_weight[w_ + h_ *
patch_num[0],
1]
inv_weight_patch[0, 2, w_ * patch_size[0]:(w_ + 1) *
patch_size[0],
h_ * patch_size[1]:(h_ + 1) *
patch_size[1]] = inv_weight[w_ + h_ *
patch_num[0],
2]
k_local.append(k_tensor[w_ + h_ * patch_num[0]:w_ +
h_ * patch_num[0] + 1])
k = torch.cat(k_local, dim=0)
[x, x_gt, k, inv_weight_patch
] = [el.to(device) for el in [x, x_gt, k, inv_weight_patch]]
ab_patch = F.softplus(ab[px_start:px_start + patch_num[0],
py_start:py_start + patch_num[1]])
cd = []
for h_ in range(patch_num[1]):
for w_ in range(patch_num[0]):
cd.append(ab_patch[w_:w_ + 1, h_])
cd = torch.cat(cd, dim=0)
x_E = model.forward_patchwise(x, k, cd, patch_num, patch_size)
predict = x_E[...,block_expand:block_expand+patch_size[0]*patch_num[0],\
block_expand:block_expand+patch_size[1]*patch_num[1]]
loss += F.l1_loss(predict.div(inv_weight_patch),
x_gt.div(inv_weight_patch))
patch_L = patch_L_wrap.astype(np.uint8)
patch_E = util.tensor2uint(x_E)[block_expand:-block_expand,
block_expand:-block_expand]
show = np.hstack((patch_H[expand:-expand, expand:-expand],
patch_L[block_expand:-block_expand,
block_expand:-block_expand], patch_E))
save_image.append(show)
optimizer.zero_grad()
loss.backward()
optimizer.step()
scheduler.step()
print('iter:{},loss {}'.format(global_iter + 1, loss.item()))
save_image = np.vstack(save_image)
if vis:
cv2.imshow('HL', save_image)
key = cv2.waitKey(1)
global_iter += 1
if global_iter % 100 == 0:
ab_numpy = ab.detach().cpu().numpy().flatten()
torch.save(
model.state_dict(),
'/home/xiu/databag/deblur/models/bench/uabcnet_iter{}.pth'.
format(global_iter))
np.savetxt(
'/home/xiu/databag/deblur/models/bench/ab_iter{}.txt'.format(
global_iter), ab_numpy)
cv2.imwrite(
'/home/xiu/databag/deblur/models/bench/show_iter{}.png'.format(
global_iter), save_image)
if vis and key == ord('q'):
running = False
break
def main():
#0. global config
#scale factor
sf = 4
stage = 8
patch_size = [32,32]
patch_num = [2,2]
#1. local PSF
#shape: gx,gy,kw,kw,3
all_PSFs = load_kernels('./data')
#2. local model
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = net(n_iter=8, h_nc=64, in_nc=4, out_nc=3, nc=[64, 128, 256, 512],
nb=2,sf=sf, act_mode="R", downsample_mode='strideconv', upsample_mode="convtranspose")
#loaded_state_dict= torch.load('./data/uabcnet_final.pth')
loaded_state_dict= torch.load('./data/uabcnet_finetune.pth')
model.load_state_dict(loaded_state_dict,strict=True)
model.eval()
for _, v in model.named_parameters():
v.requires_grad = False
model = model.to(device)
#positional lambda, mu for HQS, set as free trainable parameters here.
ab_buffer = np.loadtxt('./data/ab.txt').reshape((patch_num[0],patch_num[1],2*stage,3)).astype(np.float32)
#ab[2x2,2*stage,3]
#ab_buffer = np.ones((patch_num[0],patch_num[1],2*stage,3),dtype=np.float32)*0.1
ab = torch.tensor(ab_buffer,device=device,requires_grad=False)
#ab = F.softplus(ab)
#3.load training data
imgs_H = glob.glob('/home/xiu/databag/deblur/images/DIV2K_train/*.png',recursive=True)
imgs_H.sort()
global_iter = 0
N_maxiter = 1000
PSF_grid = using_AC254_lens(all_PSFs,patch_num)
all_PSNR = []
for i in range(N_maxiter):
#draw random image.
img_idx = np.random.randint(len(imgs_H))
img_H = cv2.imread(imgs_H[img_idx])
patch_L,patch_H,patch_psf = draw_training_pair(img_H,PSF_grid,sf,patch_num,patch_size)
x = util.uint2single(patch_L)
x = util.single2tensor4(x)
x_gt = util.uint2single(patch_H)
x_gt = util.single2tensor4(x_gt)
k_local = []
for h_ in range(patch_num[1]):
for w_ in range(patch_num[0]):
k_local.append(util.single2tensor4(patch_psf[w_,h_]))
k = torch.cat(k_local,dim=0)
[x,x_gt,k] = [el.to(device) for el in [x,x_gt,k]]
ab_patch = F.softplus(ab)
ab_patch_v = []
for h_ in range(patch_num[1]):
for w_ in range(patch_num[0]):
ab_patch_v.append(ab_patch[w_:w_+1,h_])
ab_patch_v = torch.cat(ab_patch_v,dim=0)
x_E = model.forward_patchwise_SR(x,k,ab_patch_v,patch_num,[patch_size[0],patch_size[1]],sf)
patch_L = cv2.resize(patch_L,dsize=None,fx=sf,fy=sf,interpolation=cv2.INTER_NEAREST)
patch_E = util.tensor2uint((x_E))
psnr = cv2.PSNR(patch_E,patch_H)
all_PSNR.append(psnr)
show = np.hstack((patch_H,patch_L,patch_E))
def main():
args = parse_args()
noise_level_img = args.ID_noise # noise level for noisy image
noise_level_model = noise_level_img # noise level for model
model_name = args.ID_model # 'ffdnet_gray' | 'ffdnet_color' | 'ffdnet_color_clip' | 'ffdnet_gray_clip'
testset_name = 'CBSD68' # test set, 'bsd68' | 'cbsd68' | 'set12'
need_degradation = True # default: True
task_current = 'dn' # 'dn' for denoising | 'sr' for super-resolution
sf = 1 # unused for denoising
if 'color' in model_name:
n_channels = 3 # setting for color image
nc = 96 # setting for color image
nb = 12 # setting for color image
else:
n_channels = 1 # setting for grayscale image
nc = 64 # setting for grayscale image
nb = 15 # setting for grayscale image
if 'clip' in model_name:
use_clip = True # clip the intensities into range of [0, 1]
else:
use_clip = False
model_pool = '/workspace/xuma/PAIR/KAIR/model_zoo' # fixed
model_path = os.path.join(model_pool, model_name+'.pth')
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
from models.network_ffdnet import FFDNet as net
model = net(in_nc=n_channels, out_nc=n_channels, nc=nc, nb=nb, act_mode='R')
model.load_state_dict(torch.load(model_path), strict=True)
model.eval()
for k, v in model.named_parameters():
v.requires_grad = False
model = model.to(device)
img_name, ext = os.path.splitext(args.test_image)
img_L = util.imread_uint(args.test_image, n_channels=n_channels)
img_L = util.uint2single(img_L)
if need_degradation: # degradation process
np.random.seed(seed=0) # for reproducibility
img_L += np.random.normal(0, noise_level_img/255., img_L.shape)
if use_clip:
img_L = util.uint2single(util.single2uint(img_L))
img_L = util.single2tensor4(img_L)
img_L = img_L.to(device)
sigma = torch.full((1,1,1,1), noise_level_model/255.).type_as(img_L)
img_E = model(img_L, sigma)
img_E = util.tensor2uint(img_E)
save_path = os.path.join(args.ID_savepath, os.path.split(img_name)[1]+"_denoising"+str(args.ID_noise)+ext)
util.imsave(img_E, save_path)
print(f"Denoised image is saved to {save_path}")
def main():
# ----------------------------------------
# load kernels
# ----------------------------------------
#PSF_grid = np.load('./data/AC254-075-A-ML-Zemax(ZMX).npz')['PSF']
PSF_grid = np.load('./data/Schuler_PSF_facade.npz')['PSF']
PSF_grid = PSF_grid.astype(np.float32)
gx, gy = PSF_grid.shape[:2]
for xx in range(gx):
for yy in range(gy):
PSF_grid[xx, yy] = PSF_grid[xx, yy] / np.sum(PSF_grid[xx, yy],
axis=(0, 1))
# ----------------------------------------
# load model
# ----------------------------------------
stage = 8
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = net(n_iter=stage,
h_nc=64,
in_nc=4,
out_nc=3,
nc=[64, 128, 256, 512],
nb=2,
act_mode="R",
downsample_mode='strideconv',
upsample_mode="convtranspose")
model_code = 'iter1500'
loaded_state = torch.load(
'/home/xiu/databag/deblur/models/facade/uabcnet_{}.pth'.format(
model_code))
#strip_state = strip_prefix_if_present(loaded_state,prefix="p.")
model.load_state_dict(loaded_state, strict=True)
model.eval()
for _, v in model.named_parameters():
v.requires_grad = False
model = model.to(device)
for img_id in range(1, 237):
#for img_id in range(1,12):
#img_L = cv2.imread('/home/xiu/workspace/UABC/ICCV2021/video1-3/res/2_{:03d}.bmp'.format(img_id))
#img_L = cv2.imread('/home/xiu/workspace/UABC/ICCV2021/video/{:08d}.bmp'.format(img_id))
#img_L = cv2.imread('/home/xiu/databag/deblur/ICCV2021/suo_image/{}/AC254-075-A-ML-Zemax(ZMX).bmp'.format(img_id))
#img_L = cv2.imread('/home/xiu/workspace/UABC/ICCV2021/ResolutionChart/Reso.bmp')
img_L = cv2.imread(
'/home/xiu/databag/deblur/ICCV2021/MPI_data/facade/blurry.jpg')
img_L = img_L.astype(np.float32)
#img_L = np.pad(img_L,((1,1),(61,62),(0,0)),mode='edge')
full_img = img_L.copy()
weight = full_img.copy()
W, H = img_L.shape[:2]
num_patch = [gx, gx]
#positional alpha-beta parameters for HQS
#ab_numpy = np.ones((num_patch*num_patch,17,1,1),dtype=np.float32)*0.1
#ab_numpy[:,0,:,:] = 0.01
ab_numpy = np.loadtxt(
'/home/xiu/databag/deblur/models/facade/ab_{}.txt'.format(
model_code)).astype(np.float32).reshape(gx, gy, stage * 2, 3)
#ab_numpy = ab_numpy[:,1:-1,:,:]
#ab_numpy = ab_numpy[...,None,None]
ab = torch.tensor(ab_numpy, device=device, requires_grad=False)
#save img_L
t0 = time.time()
px_start = 0
for py_start in range(0, gy - gx + 1, 1):
PSF_patch = PSF_grid[px_start:px_start + num_patch[0],
py_start:py_start + num_patch[1]]
#block_expand = 1
patch_L = img_L[px_start * W // gx:(px_start + num_patch[0]) * W //
gx, py_start * H // gy:(py_start + num_patch[1]) *
H // gy, :]
p_W, p_H = patch_L.shape[:2]
expand = max(PSF_grid.shape[2] // 2, p_W // 16)
block_expand = expand
patch_L_wrap = util_deblur.wrap_boundary_liu(
patch_L, (p_W + block_expand * 2, p_H + block_expand * 2))
#centralize
patch_L_wrap = np.hstack((patch_L_wrap[:, -block_expand:, :],
patch_L_wrap[:, :p_H + block_expand, :]))
patch_L_wrap = np.vstack((patch_L_wrap[-block_expand:, :, :],
patch_L_wrap[:p_W + block_expand, :, :]))
x = util.uint2single(patch_L_wrap)
x = util.single2tensor4(x)
k_all = []
for h_ in range(num_patch[1]):
for w_ in range(num_patch[0]):
k_all.append(util.single2tensor4(PSF_patch[w_, h_]))
k = torch.cat(k_all, dim=0)
[x, k] = [el.to(device) for el in [x, k]]
ab_patch = F.softplus(ab[px_start:px_start + num_patch[0],
py_start:py_start + num_patch[1]])
cd = []
for h_ in range(num_patch[1]):
for w_ in range(num_patch[0]):
cd.append(ab_patch[w_:w_ + 1, h_])
cd = torch.cat(cd, dim=0)
x_E = model.forward_patchwise(x, k, cd, num_patch,
[W // gx, H // gy])
x_E = x_E[..., block_expand:block_expand + p_W,
block_expand:block_expand + p_H]
patch_L = patch_L_wrap.astype(np.uint8)
patch_E = util.tensor2uint(x_E)
if py_start == 0 or py_start == gy - gx:
full_img[px_start * W // gx:(px_start + num_patch[0]) * W //
gx, py_start * H // gy:(py_start + num_patch[1]) *
H // gy, :] = patch_E
else:
full_img[px_start * W // gx:(px_start + num_patch[0]) * W //
gx, py_start * H // gy:(py_start + num_patch[1]) *
H // gy, :] = patch_E
t1 = time.time()
print(py_start)
print(t1 - t0)
# print(i)
xk = patch_E
# #zk = zk.astype(np.uint8)
xk = full_img.astype(np.uint8)
#cv2.imwrite('/home/xiu/workspace/UABC/ICCV2021/new_image/image/ours-{}.png'.format(img_id),xk)
#cv2.imwrite('/home/xiu/workspace/UABC/ICCV2021/video_deblur/{:08d}.png'.format(img_id),xk)
#cv2.imwrite('/home/xiu/workspace/UABC/ICCV2021/cap_result/1_{:03d}.png'.format(img_id),xk)
cv2.imshow('xx', xk)
cv2.imshow('img_L', img_L.astype(np.uint8))
key = cv2.waitKey(-1)
if key == ord('q'):
break
def main():
# ----------------------------------------
# Preparation
# ----------------------------------------
model_name = '1000_G' # 'usrgan' | 'usrnet' | 'usrgan_tiny' | 'usrnet_tiny'
testset_name = 'set_real' # test set, 'set5' | 'srbsd68'
test_sf = [4] if 'gan' in model_name else [2, 3, 4] # scale factor, from {1,2,3,4}
show_img = False # default: False
save_L = True # save LR image
save_E = True # save estimated image
save_LEH = False # save zoomed LR, E and H images
# ----------------------------------------
# load testing kernels
# ----------------------------------------
# kernels = hdf5storage.loadmat(os.path.join('kernels', 'kernels.mat'))['kernels']
kernels = loadmat(os.path.join('kernels', 'kernels_12.mat'))['kernels']
n_channels = 1 if 'gray' in model_name else 3 # 3 for color image, 1 for grayscale image
model_pool = 'model_zoo' # fixed
testsets = 'testsets' # fixed
results = 'results' # fixed
noise_level_img = 0 # fixed: 0, noise level for LR image
noise_level_model = noise_level_img # fixed, noise level of model, default 0
result_name = testset_name + '_' + model_name
model_path = os.path.join(model_pool, model_name+'.pth')
# ----------------------------------------
# L_path = H_path, E_path, logger
# ----------------------------------------
L_path = os.path.join(testsets, testset_name) # L_path and H_path, fixed, for Low-quality images
E_path = os.path.join(results, result_name) # E_path, fixed, for Estimated images
util.mkdir(E_path)
logger_name = result_name
utils_logger.logger_info(logger_name, log_path=os.path.join(E_path, logger_name+'.log'))
logger = logging.getLogger(logger_name)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# ----------------------------------------
# load model
# ----------------------------------------
if 'tiny' in model_name:
model = net(n_iter=6, h_nc=32, in_nc=4, out_nc=3, nc=[16, 32, 64, 64],
nb=2, act_mode="BR", downsample_mode='strideconv', upsample_mode="convtranspose")
else:
model = net(n_iter=8, h_nc=64, in_nc=4, out_nc=3, nc=[64, 128, 256, 512],
nb=2, act_mode="BR", downsample_mode='strideconv', upsample_mode="convtranspose")
model.load_state_dict(torch.load(model_path), strict=True)
model.eval()
for key, v in model.named_parameters():
v.requires_grad = False
number_parameters = sum(map(lambda x: x.numel(), model.parameters()))
model = model.to(device)
logger.info('Model path: {:s}'.format(model_path))
logger.info('Params number: {}'.format(number_parameters))
logger.info('Model_name:{}, image sigma:{}'.format(model_name, noise_level_img))
logger.info(L_path)
L_paths = util.get_image_paths(L_path)
# --------------------------------
# read images
# --------------------------------
test_results_ave = OrderedDict()
test_results_ave['psnr_sf_k'] = []
for sf in test_sf:
for k_index in range(kernels.shape[1]):
test_results = OrderedDict()
test_results['psnr'] = []
kernel = kernels[0, k_index].astype(np.float64)
## other kernels
# kernel = utils_deblur.blurkernel_synthesis(h=25) # motion kernel
# kernel = utils_deblur.fspecial('gaussian', 25, 1.6) # Gaussian kernel
# kernel = sr.shift_pixel(kernel, sf) # pixel shift; optional
# kernel /= np.sum(kernel)
util.surf(kernel) if show_img else None
idx = 0
for img in L_paths:
# --------------------------------
# (1) classical degradation, img_L
# --------------------------------
idx += 1
img_name, ext = os.path.splitext(os.path.basename(img))
img_H = util.imread_uint(img, n_channels=n_channels) # HR image, int8
img_H = util.modcrop(img_H, np.lcm(sf,8)) # modcrop
# generate degraded LR image
img_L = ndimage.filters.convolve(img_H, kernel[..., np.newaxis], mode='wrap') # blur
img_L = sr.downsample_np(img_L, sf, center=False) # downsample, standard s-fold downsampler
img_L = util.uint2single(img_L) # uint2single
np.random.seed(seed=0) # for reproducibility
img_L += np.random.normal(0, noise_level_img, img_L.shape) # add AWGN
util.imshow(util.single2uint(img_L)) if show_img else None
x = util.single2tensor4(img_L)
k = util.single2tensor4(kernel[..., np.newaxis])
sigma = torch.tensor(noise_level_model).float().view([1, 1, 1, 1])
[x, k, sigma] = [el.to(device) for el in [x, k, sigma]]
# --------------------------------
# (2) inference
# --------------------------------
x = model(x, k, sf, sigma)
# --------------------------------
# (3) img_E
# --------------------------------
img_E = util.tensor2uint(x)
if save_E:
util.imsave(img_E, os.path.join(E_path, img_name+'_x'+str(sf)+'_k'+str(k_index+1)+'_'+model_name+'.png'))
# --------------------------------
# (4) img_LEH
# --------------------------------
img_L = util.single2uint(img_L)
if save_LEH:
k_v = kernel/np.max(kernel)*1.2
k_v = util.single2uint(np.tile(k_v[..., np.newaxis], [1, 1, 3]))
k_v = cv2.resize(k_v, (3*k_v.shape[1], 3*k_v.shape[0]), interpolation=cv2.INTER_NEAREST)
img_I = cv2.resize(img_L, (sf*img_L.shape[1], sf*img_L.shape[0]), interpolation=cv2.INTER_NEAREST)
img_I[:k_v.shape[0], -k_v.shape[1]:, :] = k_v
img_I[:img_L.shape[0], :img_L.shape[1], :] = img_L
util.imshow(np.concatenate([img_I, img_E, img_H], axis=1), title='LR / Recovered / Ground-truth') if show_img else None
util.imsave(np.concatenate([img_I, img_E, img_H], axis=1), os.path.join(E_path, img_name+'_x'+str(sf)+'_k'+str(k_index+1)+'_LEH.png'))
if save_L:
util.imsave(img_L, os.path.join(E_path, img_name+'_x'+str(sf)+'_k'+str(k_index+1)+'_LR.png'))
psnr = util.calculate_psnr(img_E, img_H, border=sf**2) # change with your own border
test_results['psnr'].append(psnr)
logger.info('{:->4d}--> {:>10s} -- x{:>2d} --k{:>2d} PSNR: {:.2f}dB'.format(idx, img_name+ext, sf, k_index, psnr))
ave_psnr_k = sum(test_results['psnr']) / len(test_results['psnr'])
logger.info('------> Average PSNR(RGB) of ({}) scale factor: ({}), kernel: ({}) sigma: ({}): {:.2f} dB'.format(testset_name, sf, k_index+1, noise_level_model, ave_psnr_k))
test_results_ave['psnr_sf_k'].append(ave_psnr_k)
logger.info(test_results_ave['psnr_sf_k'])
def main():
# ----------------------------------------
# Preparation
# ----------------------------------------
model_name = 'usrgan' # 'usrgan' | 'usrnet' | 'usrgan_tiny' | 'usrnet_tiny'
testset_name = 'set_real' # test set, 'set_real'
test_image = 'test.png' # 'chip.png', 'comic.png'
#test_image = 'comic.png'
sf = 1 # scale factor, only from {1, 2, 3, 4}
show_img = True # default: False
save_E = True # save estimated image
save_LE = True # save zoomed LR, Estimated images
# ----------------------------------------
# set noise level and kernel
# ----------------------------------------
if 'chip' in test_image:
noise_level_img = 15 # noise level for LR image, 15 for chip
kernel_width_default_x1234 = [
0.6, 0.9, 1.7, 2.2
] # Gaussian kernel widths for x1, x2, x3, x4
else:
noise_level_img = 2 # noise level for LR image, 0.5~3 for clean images
kernel_width_default_x1234 = [
0.4, 0.7, 1.5, 2.0
] # default Gaussian kernel widths of clean/sharp images for x1, x2, x3, x4
noise_level_model = noise_level_img / 255. # noise level of model
kernel_width = kernel_width_default_x1234[sf - 1]
# set your own kernel width
# kernel_width = 2.2
k = utils_deblur.fspecial('gaussian', 25, kernel_width)
k = sr.shift_pixel(k, sf) # shift the kernel
k /= np.sum(k)
util.surf(k) if show_img else None
# scio.savemat('kernel_realapplication.mat', {'kernel':k})
# load approximated bicubic kernels
#kernels = hdf5storage.loadmat(os.path.join('kernels', 'kernel_bicubicx234.mat'))['kernels']
# kernels = loadmat(os.path.join('kernels', 'kernel_bicubicx234.mat'))['kernels']
# kernel = kernels[0, sf-2].astype(np.float64)
kernel = util.single2tensor4(k[..., np.newaxis])
n_channels = 1 if 'gray' in model_name else 3 # 3 for color image, 1 for grayscale image
model_pool = 'model_zoo' # fixed
testsets = 'testsets' # fixed
results = 'results' # fixed
result_name = testset_name + '_' + model_name
model_path = os.path.join(model_pool, model_name + '.pth')
# ----------------------------------------
# L_path, E_path
# ----------------------------------------
L_path = os.path.join(
testsets, testset_name) # L_path, fixed, for Low-quality images
E_path = os.path.join(results,
result_name) # E_path, fixed, for Estimated images
util.mkdir(E_path)
logger_name = result_name
utils_logger.logger_info(logger_name,
log_path=os.path.join(E_path,
logger_name + '.log'))
logger = logging.getLogger(logger_name)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# ----------------------------------------
# load model
# ----------------------------------------
if 'tiny' in model_name:
model = net(n_iter=6,
h_nc=32,
in_nc=4,
out_nc=3,
nc=[16, 32, 64, 64],
nb=2,
act_mode="R",
downsample_mode='strideconv',
upsample_mode="convtranspose")
else:
model = net(n_iter=8,
h_nc=64,
in_nc=4,
out_nc=3,
nc=[64, 128, 256, 512],
nb=2,
act_mode="R",
downsample_mode='strideconv',
upsample_mode="convtranspose")
model.load_state_dict(torch.load(model_path), strict=True)
model.eval()
for key, v in model.named_parameters():
v.requires_grad = False
number_parameters = sum(map(lambda x: x.numel(), model.parameters()))
logger.info('Params number: {}'.format(number_parameters))
model = model.to(device)
logger.info('Model path: {:s}'.format(model_path))
logger.info('model_name:{}, image sigma:{}'.format(model_name,
noise_level_img))
logger.info(L_path)
img = os.path.join(L_path, test_image)
# ------------------------------------
# (1) img_L
# ------------------------------------
img_name, ext = os.path.splitext(os.path.basename(img))
img_L = util.imread_uint(img, n_channels=n_channels)
img_L = util.uint2single(img_L)
util.imshow(img_L) if show_img else None
w, h = img_L.shape[:2]
logger.info('{:>10s}--> ({:>4d}x{:<4d})'.format(img_name + ext, w, h))
# boundary handling
boarder = 8 # default setting for kernel size 25x25
img = cv2.resize(img_L, (sf * h, sf * w), interpolation=cv2.INTER_NEAREST)
img = utils_deblur.wrap_boundary_liu(img, [
int(np.ceil(sf * w / boarder + 2) * boarder),
int(np.ceil(sf * h / boarder + 2) * boarder)
])
img_wrap = sr.downsample_np(img, sf, center=False)
img_wrap[:w, :h, :] = img_L
img_L = img_wrap
util.imshow(util.single2uint(img_L),
title='LR image with noise level {}'.format(
noise_level_img)) if show_img else None
img_L = util.single2tensor4(img_L)
img_L = img_L.to(device)
# ------------------------------------
# (2) img_E
# ------------------------------------
sigma = torch.tensor(noise_level_model).float().view([1, 1, 1, 1])
[img_L, kernel, sigma] = [el.to(device) for el in [img_L, kernel, sigma]]
img_E = model(img_L, kernel, sf, sigma)
img_E = util.tensor2uint(img_E)[:sf * w, :sf * h, ...]
if save_E:
util.imsave(
img_E,
os.path.join(E_path, img_name + '_x' + str(sf) + '_' + model_name +
'.png'))
# --------------------------------
# (3) save img_LE
# --------------------------------
if save_LE:
k_v = k / np.max(k) * 1.2
k_v = util.single2uint(np.tile(k_v[..., np.newaxis], [1, 1, 3]))
k_factor = 3
k_v = cv2.resize(k_v,
(k_factor * k_v.shape[1], k_factor * k_v.shape[0]),
interpolation=cv2.INTER_NEAREST)
img_L = util.tensor2uint(img_L)[:w, :h, ...]
img_I = cv2.resize(img_L, (sf * img_L.shape[1], sf * img_L.shape[0]),
interpolation=cv2.INTER_NEAREST)
img_I[:k_v.shape[0], :k_v.shape[1], :] = k_v
util.imshow(np.concatenate([img_I, img_E], axis=1),
title='LR / Recovered') if show_img else None
util.imsave(
np.concatenate([img_I, img_E], axis=1),
os.path.join(
E_path,
img_name + '_x' + str(sf) + '_' + model_name + '_LE.png'))
def main():
# ----------------------------------------
# Preparation
# ----------------------------------------
model_name = 'usrgan' # 'usrgan' | 'usrnet' | 'usrgan_tiny' | 'usrnet_tiny'
testset_name = 'set5' # test set, 'set5' | 'srbsd68'
need_degradation = False # default: True
sf = 4 # scale factor, only from {2, 3, 4}
show_img = False # default: False
save_L = True # save LR image
save_E = True # save estimated image
# load approximated bicubic kernels
#kernels = hdf5storage.loadmat(os.path.join('kernels', 'kernels_bicubicx234.mat'))['kernels']
kernels = loadmat(os.path.join('kernels',
'kernels_bicubicx234.mat'))['kernels']
kernel = kernels[0, sf - 2].astype(np.float64)
kernel = util.single2tensor4(kernel[..., np.newaxis])
task_current = 'sr' # fixed, 'sr' for super-resolution
n_channels = 3 # fixed, 3 for color image
model_pool = '/home/dengzeshuai/pretrained_models/USRnet/' # fixed
testsets = 'testsets' # fixed
results = 'results' # fixed
noise_level_img = 0 # fixed: 0, noise level for LR image
noise_level_model = noise_level_img # fixed, noise level of model, default 0
result_name = testset_name + '_' + model_name + '_bicubic'
border = sf if task_current == 'sr' else 0 # shave boader to calculate PSNR and SSIM
model_path = os.path.join(model_pool, model_name + '.pth')
# ----------------------------------------
# L_path, E_path, H_path
# ----------------------------------------
L_path = os.path.join(
testsets, testset_name) # L_path, fixed, for Low-quality images
H_path = L_path # H_path, 'None' | L_path, for High-quality images
E_path = os.path.join(results,
result_name) # E_path, fixed, for Estimated images
util.mkdir(E_path)
if H_path == L_path:
need_degradation = True
logger_name = result_name
utils_logger.logger_info(logger_name,
log_path=os.path.join(E_path,
logger_name + '.log'))
logger = logging.getLogger(logger_name)
need_H = True if H_path is not None else False
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# ----------------------------------------
# load model
# ----------------------------------------
from models.network_usrnet import USRNet as net
if 'tiny' in model_name:
model = net(n_iter=6,
h_nc=32,
in_nc=4,
out_nc=3,
nc=[16, 32, 64, 64],
nb=2,
act_mode="R",
downsample_mode='strideconv',
upsample_mode="convtranspose")
else:
model = net(n_iter=8,
h_nc=64,
in_nc=4,
out_nc=3,
nc=[64, 128, 256, 512],
nb=2,
act_mode="R",
downsample_mode='strideconv',
upsample_mode="convtranspose")
model.load_state_dict(torch.load(model_path), strict=True)
model.eval()
for key, v in model.named_parameters():
v.requires_grad = False
number_parameters = sum(map(lambda x: x.numel(), model.parameters()))
logger.info('Params number: {}'.format(number_parameters))
model = model.to(device)
logger.info('Model path: {:s}'.format(model_path))
test_results = OrderedDict()
test_results['psnr'] = []
test_results['ssim'] = []
test_results['psnr_y'] = []
test_results['ssim_y'] = []
logger.info('model_name:{}, image sigma:{}'.format(model_name,
noise_level_img))
logger.info(L_path)
L_paths = util.get_image_paths(L_path)
H_paths = util.get_image_paths(H_path) if need_H else None
for idx, img in enumerate(L_paths):
# ------------------------------------
# (1) img_L
# ------------------------------------
img_name, ext = os.path.splitext(os.path.basename(img))
logger.info('{:->4d}--> {:>10s}'.format(idx + 1, img_name + ext))
img_L = util.imread_uint(img, n_channels=n_channels)
img_L = util.uint2single(img_L)
# degradation process, bicubic downsampling
if need_degradation:
img_L = util.modcrop(img_L, sf)
img_L = util.imresize_np(img_L, 1 / sf)
# img_L = util.uint2single(util.single2uint(img_L))
# np.random.seed(seed=0) # for reproducibility
# img_L += np.random.normal(0, noise_level_img/255., img_L.shape)
w, h = img_L.shape[:2]
if save_L:
util.imsave(
util.single2uint(img_L),
os.path.join(E_path, img_name + '_LR_x' + str(sf) + '.png'))
img = cv2.resize(img_L, (sf * h, sf * w),
interpolation=cv2.INTER_NEAREST)
img = utils_deblur.wrap_boundary_liu(img, [
int(np.ceil(sf * w / 8 + 2) * 8),
int(np.ceil(sf * h / 8 + 2) * 8)
])
img_wrap = sr.downsample_np(img, sf, center=False)
img_wrap[:w, :h, :] = img_L
img_L = img_wrap
util.imshow(util.single2uint(img_L),
title='LR image with noise level {}'.format(
noise_level_img)) if show_img else None
img_L = util.single2tensor4(img_L)
img_L = img_L.to(device)
# ------------------------------------
# (2) img_E
# ------------------------------------
sigma = torch.tensor(noise_level_model).float().view([1, 1, 1, 1])
[img_L, kernel,
sigma] = [el.to(device) for el in [img_L, kernel, sigma]]
img_E = model(img_L, kernel, sf, sigma)
img_E = util.tensor2uint(img_E)
img_E = img_E[:sf * w, :sf * h, :]
if need_H:
# --------------------------------
# (3) img_H
# --------------------------------
img_H = util.imread_uint(H_paths[idx], n_channels=n_channels)
img_H = img_H.squeeze()
img_H = util.modcrop(img_H, sf)
# --------------------------------
# PSNR and SSIM
# --------------------------------
psnr = util.calculate_psnr(img_E, img_H, border=border)
ssim = util.calculate_ssim(img_E, img_H, border=border)
test_results['psnr'].append(psnr)
test_results['ssim'].append(ssim)
logger.info('{:s} - PSNR: {:.2f} dB; SSIM: {:.4f}.'.format(
img_name + ext, psnr, ssim))
util.imshow(np.concatenate([img_E, img_H], axis=1),
title='Recovered / Ground-truth') if show_img else None
if np.ndim(img_H) == 3: # RGB image
img_E_y = util.rgb2ycbcr(img_E, only_y=True)
img_H_y = util.rgb2ycbcr(img_H, only_y=True)
psnr_y = util.calculate_psnr(img_E_y, img_H_y, border=border)
ssim_y = util.calculate_ssim(img_E_y, img_H_y, border=border)
test_results['psnr_y'].append(psnr_y)
test_results['ssim_y'].append(ssim_y)
# ------------------------------------
# save results
# ------------------------------------
if save_E:
util.imsave(
img_E,
os.path.join(
E_path,
img_name + '_x' + str(sf) + '_' + model_name + '.png'))
if need_H:
ave_psnr = sum(test_results['psnr']) / len(test_results['psnr'])
ave_ssim = sum(test_results['ssim']) / len(test_results['ssim'])
logger.info(
'Average PSNR/SSIM(RGB) - {} - x{} --PSNR: {:.2f} dB; SSIM: {:.4f}'
.format(result_name, sf, ave_psnr, ave_ssim))
if np.ndim(img_H) == 3:
ave_psnr_y = sum(test_results['psnr_y']) / len(
test_results['psnr_y'])
ave_ssim_y = sum(test_results['ssim_y']) / len(
test_results['ssim_y'])
logger.info(
'Average PSNR/SSIM( Y ) - {} - x{} - PSNR: {:.2f} dB; SSIM: {:.4f}'
.format(result_name, sf, ave_psnr_y, ave_ssim_y))
def main():
# ----------------------------------------
# Preparation
# ----------------------------------------
model_name = '300000_G' # 'msrresnet_x4_gan' | 'msrresnet_x4_psnr'
testset_name = 'testing_lr_images' # test set, 'set5' | 'srbsd68'
need_degradation = False # default: True
x8 = False # default: False, x8 to boost performance, default: False
sf = 3 #[int(s) for s in re.findall(r'\d+', model_name)][0] # scale factor
show_img = False # default: False
task_current = 'sr' # 'dn' for denoising | 'sr' for super-resolution
n_channels = 3 # fixed
model_pool = './superresolution/msrresnet_psnr/models/' # fixed
testsets = 'testsets' # fixed
results = 'results' # fixed
noise_level_img = 0 # fixed: 0, noise level for LR image
result_name = testset_name + '_' + model_name + '_msrresnet'
border = sf if task_current == 'sr' else 0 # shave boader to calculate PSNR and SSIM
model_path = os.path.join(model_pool, model_name + '.pth')
# ----------------------------------------
# L_path, E_path, H_path
# ----------------------------------------
L_path = os.path.join(testsets,
testset_name) # L_path, for Low-quality images
H_path = None # H_path, for High-quality images
E_path = os.path.join(results, result_name) # E_path, for Estimated images
util.mkdir(E_path)
if H_path == L_path:
need_degradation = True
logger_name = result_name
utils_logger.logger_info(logger_name,
log_path=os.path.join(E_path,
logger_name + '.log'))
logger = logging.getLogger(logger_name)
need_H = True if H_path is not None else False
os.environ["CUDA_VISIBLE_DEVICES"] = '8'
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# ----------------------------------------
# load model
# ----------------------------------------
from models.network_msrresnet import MSRResNet1 as net
model = net(in_nc=n_channels, out_nc=n_channels, nc=64, nb=16, upscale=3)
model.load_state_dict(torch.load(model_path), strict=False)
model.eval()
for k, v in model.named_parameters():
v.requires_grad = False
model = model.to(device)
logger.info('Model path: {:s}'.format(model_path))
number_parameters = sum(map(lambda x: x.numel(), model.parameters()))
logger.info('Params number: {}'.format(number_parameters))
test_results = OrderedDict()
test_results['psnr'] = []
test_results['ssim'] = []
test_results['psnr_y'] = []
test_results['ssim_y'] = []
logger.info('model_name:{}, image sigma:{}'.format(model_name,
noise_level_img))
logger.info(L_path)
L_paths = util.get_image_paths(L_path)
H_paths = util.get_image_paths(H_path) if need_H else None
for idx, img in enumerate(L_paths):
# ------------------------------------
# (1) img_L
# ------------------------------------
img_name, ext = os.path.splitext(os.path.basename(img))
# logger.info('{:->4d}--> {:>10s}'.format(idx+1, img_name+ext))
img_L = util.imread_uint(img, n_channels=n_channels)
img_L = util.uint2single(img_L)
# degradation process, bicubic downsampling
if need_degradation:
img_L = util.modcrop(img_L, sf)
img_L = util.imresize_np(img_L, 1 / sf)
# img_L = util.uint2single(util.single2uint(img_L))
# np.random.seed(seed=0) # for reproducibility
# img_L += np.random.normal(0, noise_level_img/255., img_L.shape)
util.imshow(util.single2uint(img_L),
title='LR image with noise level {}'.format(
noise_level_img)) if show_img else None
img_L = util.single2tensor4(img_L)
img_L = img_L.to(device)
# ------------------------------------
# (2) img_E
# ------------------------------------
if not x8:
img_E = model(img_L)
else:
img_E = utils_model.test_mode(model, img_L, mode=3, sf=sf)
img_E = util.tensor2uint(img_E)
if need_H:
# --------------------------------
# (3) img_H
# --------------------------------
img_H = util.imread_uint(H_paths[idx], n_channels=n_channels)
img_H = img_H.squeeze()
img_H = util.modcrop(img_H, sf)
# --------------------------------
# PSNR and SSIM
# --------------------------------
psnr = util.calculate_psnr(img_E, img_H, border=border)
ssim = util.calculate_ssim(img_E, img_H, border=border)
test_results['psnr'].append(psnr)
test_results['ssim'].append(ssim)
logger.info('{:s} - PSNR: {:.2f} dB; SSIM: {:.4f}.'.format(
img_name + ext, psnr, ssim))
util.imshow(np.concatenate([img_E, img_H], axis=1),
title='Recovered / Ground-truth') if show_img else None
if np.ndim(img_H) == 3: # RGB image
img_E_y = util.rgb2ycbcr(img_E, only_y=True)
img_H_y = util.rgb2ycbcr(img_H, only_y=True)
psnr_y = util.calculate_psnr(img_E_y, img_H_y, border=border)
ssim_y = util.calculate_ssim(img_E_y, img_H_y, border=border)
test_results['psnr_y'].append(psnr_y)
test_results['ssim_y'].append(ssim_y)
# ------------------------------------
# save results
# ------------------------------------
util.imsave(img_E, os.path.join(E_path, img_name + '.png'))
if need_H:
ave_psnr = sum(test_results['psnr']) / len(test_results['psnr'])
ave_ssim = sum(test_results['ssim']) / len(test_results['ssim'])
logger.info(
'Average PSNR/SSIM(RGB) - {} - x{} --PSNR: {:.2f} dB; SSIM: {:.4f}'
.format(result_name, sf, ave_psnr, ave_ssim))
if np.ndim(img_H) == 3:
ave_psnr_y = sum(test_results['psnr_y']) / len(
test_results['psnr_y'])
ave_ssim_y = sum(test_results['ssim_y']) / len(
test_results['ssim_y'])
logger.info(
'Average PSNR/SSIM( Y ) - {} - x{} - PSNR: {:.2f} dB; SSIM: {:.4f}'
.format(result_name, sf, ave_psnr_y, ave_ssim_y))
def main():
# ----------------------------------------
# Preparation
# ----------------------------------------
model_name = 'usrnet_tiny' # 'usrgan' | 'usrnet' | 'usrgan_tiny' | 'usrnet_tiny'
testset_name = 'srcvte' # test set, 'set5' | 'srbsd68' | 'srcvte'
test_sf = [
4
] # if 'gan' in model_name else [2, 3, 4] # scale factor, from {1,2,3,4}
load_kernels = False
show_img = False # default: False
save_L = False # save LR image
save_E = True # save estimated image
save_LEH = False # save zoomed LR, E and H images
# ----------------------------------------
# load testing kernels
# ----------------------------------------
# kernels = hdf5storage.loadmat(os.path.join('kernels', 'kernels.mat'))['kernels']
kernels = loadmat(os.path.join(
'kernels', 'kernels_12.mat'))['kernels'] if load_kernels else None
n_channels = 1 if 'gray' in model_name else 3 # 3 for color image, 1 for grayscale image
model_pool = '/home/dengzeshuai/pretrained_models/USRnet/' # fixed
testsets = '/home/datasets/sr/' # fixed
results = 'results' # fixed
noise_level_img = 0 # fixed: 0, noise level for LR image
noise_level_model = noise_level_img # fixed, noise level of model, default 0
result_name = testset_name + '_' + model_name + '_blur'
model_path = os.path.join(model_pool, model_name + '.pth')
# ----------------------------------------
# L_path = H_path, E_path, logger
# ----------------------------------------
L_path = os.path.join(
testsets,
testset_name) # L_path and H_path, fixed, for Low-quality images
if testset_name == 'srcvte':
L_path = os.path.join(testsets, testset_name, 'LR_val')
H_path = os.path.join(testsets, testset_name, 'HR_val')
video_names = os.listdir(H_path)
E_path = os.path.join(results,
result_name) # E_path, fixed, for Estimated images
util.mkdir(E_path)
logger_name = result_name
utils_logger.logger_info(logger_name,
log_path=os.path.join(E_path,
logger_name + '.log'))
logger = logging.getLogger(logger_name)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# ----------------------------------------
# load model
# ----------------------------------------
if 'tiny' in model_name:
model = net(n_iter=6,
h_nc=32,
in_nc=4,
out_nc=3,
nc=[16, 32, 64, 64],
nb=2,
act_mode="R",
downsample_mode='strideconv',
upsample_mode="convtranspose")
else:
model = net(n_iter=8,
h_nc=64,
in_nc=4,
out_nc=3,
nc=[64, 128, 256, 512],
nb=2,
act_mode="R",
downsample_mode='strideconv',
upsample_mode="convtranspose")
model.load_state_dict(torch.load(model_path), strict=True)
model.eval()
for key, v in model.named_parameters():
v.requires_grad = False
number_parameters = sum(map(lambda x: x.numel(), model.parameters()))
model = model.to(device)
logger.info('Model path: {:s}'.format(model_path))
logger.info('Params number: {}'.format(number_parameters))
logger.info('Model_name:{}, image sigma:{}'.format(model_name,
noise_level_img))
logger.info(L_path)
L_paths = util.get_image_paths(L_path)
need_H = True if H_path is not None else False
H_paths = util.get_image_paths(H_path) if need_H else None
# --------------------------------
# read images
# --------------------------------
test_results_ave = OrderedDict()
test_results_ave['psnr_sf_k'] = []
test_results_ave['ssim_sf_k'] = []
test_results_ave['psnr_y_sf_k'] = []
test_results_ave['ssim_y_sf_k'] = []
for sf in test_sf:
loop = kernels.shape[1] if load_kernels else 1
for k_index in range(loop):
test_results = OrderedDict()
test_results['psnr'] = []
test_results['ssim'] = []
test_results['psnr_y'] = []
test_results['ssim_y'] = []
if load_kernels:
kernel = kernels[0, k_index].astype(np.float64)
else:
## other kernels
# kernel = utils_deblur.blurkernel_synthesis(h=25) # motion kernel
kernel = utils_deblur.fspecial('gaussian', 25,
1.6) # Gaussian kernel
kernel = sr.shift_pixel(kernel, sf) # pixel shift; optional
kernel /= np.sum(kernel)
util.surf(kernel) if show_img else None
# idx = 0
for idx, img in enumerate(L_paths):
# --------------------------------
# (1) classical degradation, img_L
# --------------------------------
img_name, ext = os.path.splitext(os.path.basename(img))
if testset_name == 'srcvte':
video_name = os.path.basename(os.path.dirname(img))
img_L = util.imread_uint(img, n_channels=n_channels)
img_L = util.uint2single(img_L)
# generate degraded LR image
# img_L = ndimage.filters.convolve(img_H, kernel[..., np.newaxis], mode='wrap') # blur
# img_L = sr.downsample_np(img_L, sf, center=False) # downsample, standard s-fold downsampler
# img_L = util.uint2single(img_L) # uint2single
# np.random.seed(seed=0) # for reproducibility
# img_L += np.random.normal(0, noise_level_img, img_L.shape) # add AWGN
util.imshow(util.single2uint(img_L)) if show_img else None
x = util.single2tensor4(img_L)
k = util.single2tensor4(kernel[..., np.newaxis])
sigma = torch.tensor(noise_level_model).float().view(
[1, 1, 1, 1])
[x, k, sigma] = [el.to(device) for el in [x, k, sigma]]
# --------------------------------
# (2) inference
# --------------------------------
x = model(x, k, sf, sigma)
# --------------------------------
# (3) img_E
# --------------------------------
img_E = util.tensor2uint(x)
if save_E:
if testset_name == 'srcvte':
save_path = os.path.join(E_path, video_name)
util.mkdir(save_path)
# util.imsave(img_E, os.path.join(save_path, img_name+'_k'+str(k_index+1)+'.png'))
util.imsave(img_E,
os.path.join(save_path, img_name + '.png'))
else:
util.imsave(
img_E,
os.path.join(
E_path, img_name + '_x' + str(sf) + '_k' +
str(k_index + 1) + '_' + model_name + '.png'))
# --------------------------------
# (4) img_H
# --------------------------------
if need_H:
img_H = util.imread_uint(H_paths[idx],
n_channels=n_channels)
img_H = img_H.squeeze()
img_H = util.modcrop(img_H, sf)
psnr = util.calculate_psnr(
img_E, img_H, border=sf) # change with your own border
ssim = util.calculate_ssim(img_E, img_H, border=sf)
test_results['psnr'].append(psnr)
test_results['ssim'].append(ssim)
if np.ndim(img_H) == 3: # RGB image
img_E_y = util.rgb2ycbcr(img_E, only_y=True)
img_H_y = util.rgb2ycbcr(img_H, only_y=True)
psnr_y = util.calculate_psnr(img_E_y,
img_H_y,
border=sf)
ssim_y = util.calculate_ssim(img_E_y,
img_H_y,
border=sf)
test_results['psnr_y'].append(psnr_y)
test_results['ssim_y'].append(ssim_y)
logger.info(
'{:->4d} --> {:>4s}--> {:>10s} -- x{:>2d} --k{:>2d} PSNR: {:.2f}dB SSIM: {:.4f}'
.format(idx, video_name, img_name + ext, sf,
k_index, psnr_y, ssim_y))
else:
logger.info(
'{:->4d} --> {:>4s}--> {:>10s} -- x{:>2d} --k{:>2d} PSNR: {:.2f}dB SSIM: {:.4f}'
.format(idx, video_name, img_name + ext, sf,
k_index, psnr, ssim))
if need_H:
ave_psnr = sum(test_results['psnr']) / len(
test_results['psnr'])
ave_ssim = sum(test_results['ssim']) / len(
test_results['ssim'])
logger.info(
'Average PSNR/SSIM(RGB) - {} - x{} --PSNR: {:.2f} dB; SSIM: {:.4f}'
.format(result_name, sf, ave_psnr, ave_ssim))
logger.info(
'------> Average PSNR(RGB) - {} - x{}, kernel:{} sigma:{} --PSNR: {:.2f} dB; SSIM: {:.4f}'
.format(testset_name, sf, k_index + 1, noise_level_model,
ave_psnr, ave_ssim))
if np.ndim(img_H) == 3:
ave_psnr_y = sum(test_results['psnr_y']) / len(
test_results['psnr_y'])
ave_ssim_y = sum(test_results['ssim_y']) / len(
test_results['ssim_y'])
logger.info(
'------> Average PSNR(Y) - {} - x{}, kernel:{} sigma:{} --PSNR: {:.2f} dB; SSIM: {:.4f}'
.format(testset_name, sf, k_index + 1,
noise_level_model, ave_psnr_y, ave_ssim_y))
test_results_ave['psnr_sf_k'].append(ave_psnr)
test_results_ave['ssim_sf_k'].append(ave_ssim)
if np.ndim(img_H) == 3:
test_results_ave['psnr_y_sf_k'].append(ave_psnr_y)
test_results_ave['ssim_y_sf_k'].append(ave_ssim_y)
logger.info(test_results_ave['psnr_sf_k'])
logger.info(test_results_ave['ssim_sf_k'])
if np.ndim(img_H) == 3:
logger.info(test_results_ave['psnr_y_sf_k'])
logger.info(test_results_ave['ssim_y_sf_k'])
def main():
# ----------------------------------------
# Preparation
# ----------------------------------------
noise_level_img = 7.65 / 255.0 # default: 0, noise level for LR image
noise_level_model = noise_level_img # noise level of model, default 0
model_name = 'drunet_gray' # 'drunet_gray' | 'drunet_color' | 'ircnn_gray' | 'ircnn_color'
testset_name = 'Set3C' # test set, 'set5' | 'srbsd68'
x8 = True # default: False, x8 to boost performance
iter_num = 8 # number of iterations
modelSigma1 = 49
modelSigma2 = noise_level_model * 255.
show_img = False # default: False
save_L = True # save LR image
save_E = True # save estimated image
save_LEH = False # save zoomed LR, E and H images
border = 0
# --------------------------------
# load kernel
# --------------------------------
kernels = hdf5storage.loadmat(os.path.join('kernels',
'Levin09.mat'))['kernels']
sf = 1
task_current = 'deblur' # 'deblur' for deblurring
n_channels = 3 if 'color' in model_name else 1 # fixed
model_zoo = 'model_zoo' # fixed
testsets = 'testsets' # fixed
results = 'results' # fixed
result_name = testset_name + '_' + task_current + '_' + model_name
model_path = os.path.join(model_zoo, model_name + '.pth')
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
torch.cuda.empty_cache()
# ----------------------------------------
# L_path, E_path, H_path
# ----------------------------------------
L_path = os.path.join(testsets,
testset_name) # L_path, for Low-quality images
E_path = os.path.join(results, result_name) # E_path, for Estimated images
util.mkdir(E_path)
logger_name = result_name
utils_logger.logger_info(logger_name,
log_path=os.path.join(E_path,
logger_name + '.log'))
logger = logging.getLogger(logger_name)
# ----------------------------------------
# load model
# ----------------------------------------
if 'drunet' in model_name:
from models.network_unet import UNetRes as net
model = net(in_nc=n_channels + 1,
out_nc=n_channels,
nc=[64, 128, 256, 512],
nb=4,
act_mode='R',
downsample_mode="strideconv",
upsample_mode="convtranspose")
model.load_state_dict(torch.load(model_path), strict=True)
model.eval()
for _, v in model.named_parameters():
v.requires_grad = False
model = model.to(device)
elif 'ircnn' in model_name:
from models.network_dncnn import IRCNN as net
model = net(in_nc=n_channels, out_nc=n_channels, nc=64)
model25 = torch.load(model_path)
former_idx = 0
logger.info('model_name:{}, image sigma:{:.3f}, model sigma:{:.3f}'.format(
model_name, noise_level_img, noise_level_model))
logger.info('Model path: {:s}'.format(model_path))
logger.info(L_path)
L_paths = util.get_image_paths(L_path)
test_results_ave = OrderedDict()
test_results_ave['psnr'] = [] # record average PSNR for each kernel
for k_index in range(kernels.shape[1]):
logger.info('-------k:{:>2d} ---------'.format(k_index))
test_results = OrderedDict()
test_results['psnr'] = []
k = kernels[0, k_index].astype(np.float64)
util.imshow(k) if show_img else None
for idx, img in enumerate(L_paths):
# --------------------------------
# (1) get img_L
# --------------------------------
img_name, ext = os.path.splitext(os.path.basename(img))
img_H = util.imread_uint(img, n_channels=n_channels)
img_H = util.modcrop(img_H, 8) # modcrop
img_L = ndimage.filters.convolve(img_H,
np.expand_dims(k, axis=2),
mode='wrap')
util.imshow(img_L) if show_img else None
img_L = util.uint2single(img_L)
np.random.seed(seed=0) # for reproducibility
img_L += np.random.normal(0, noise_level_img,
img_L.shape) # add AWGN
# --------------------------------
# (2) get rhos and sigmas
# --------------------------------
rhos, sigmas = pnp.get_rho_sigma(sigma=max(0.255 / 255.,
noise_level_model),
iter_num=iter_num,
modelSigma1=modelSigma1,
modelSigma2=modelSigma2,
w=1.0)
rhos, sigmas = torch.tensor(rhos).to(device), torch.tensor(
sigmas).to(device)
# --------------------------------
# (3) initialize x, and pre-calculation
# --------------------------------
x = util.single2tensor4(img_L).to(device)
img_L_tensor, k_tensor = util.single2tensor4(
img_L), util.single2tensor4(np.expand_dims(k, 2))
[k_tensor, img_L_tensor] = util.todevice([k_tensor, img_L_tensor],
device)
FB, FBC, F2B, FBFy = sr.pre_calculate(img_L_tensor, k_tensor, sf)
# --------------------------------
# (4) main iterations
# --------------------------------
for i in range(iter_num):
# --------------------------------
# step 1, FFT
# --------------------------------
tau = rhos[i].repeat(1, 1, 1, 1)
x = sr.data_solution(x, FB, FBC, F2B, FBFy, tau, sf)
if 'ircnn' in model_name:
current_idx = np.int(
np.ceil(sigmas[i].cpu().numpy() * 255. / 2.) - 1)
if current_idx != former_idx:
model.load_state_dict(model25[str(current_idx)],
strict=True)
model.eval()
for _, v in model.named_parameters():
v.requires_grad = False
model = model.to(device)
former_idx = current_idx
# --------------------------------
# step 2, denoiser
# --------------------------------
if x8:
x = util.augment_img_tensor4(x, i % 8)
if 'drunet' in model_name:
x = torch.cat(
(x, sigmas[i].repeat(1, 1, x.shape[2], x.shape[3])),
dim=1)
x = utils_model.test_mode(model,
x,
mode=2,
refield=32,
min_size=256,
modulo=16)
elif 'ircnn' in model_name:
x = model(x)
if x8:
if i % 8 == 3 or i % 8 == 5:
x = util.augment_img_tensor4(x, 8 - i % 8)
else:
x = util.augment_img_tensor4(x, i % 8)
# --------------------------------
# (3) img_E
# --------------------------------
img_E = util.tensor2uint(x)
if n_channels == 1:
img_H = img_H.squeeze()
if save_E:
util.imsave(
img_E,
os.path.join(
E_path, img_name + '_k' + str(k_index) + '_' +
model_name + '.png'))
# --------------------------------
# (4) img_LEH
# --------------------------------
if save_LEH:
img_L = util.single2uint(img_L)
k_v = k / np.max(k) * 1.0
k_v = util.single2uint(np.tile(k_v[..., np.newaxis],
[1, 1, 3]))
k_v = cv2.resize(k_v, (3 * k_v.shape[1], 3 * k_v.shape[0]),
interpolation=cv2.INTER_NEAREST)
img_I = cv2.resize(img_L,
(sf * img_L.shape[1], sf * img_L.shape[0]),
interpolation=cv2.INTER_NEAREST)
img_I[:k_v.shape[0], -k_v.shape[1]:, :] = k_v
img_I[:img_L.shape[0], :img_L.shape[1], :] = img_L
util.imshow(np.concatenate([img_I, img_E, img_H], axis=1),
title='LR / Recovered / Ground-truth'
) if show_img else None
util.imsave(
np.concatenate([img_I, img_E, img_H], axis=1),
os.path.join(E_path,
img_name + '_k' + str(k_index) + '_LEH.png'))
if save_L:
util.imsave(
util.single2uint(img_L),
os.path.join(E_path,
img_name + '_k' + str(k_index) + '_LR.png'))
psnr = util.calculate_psnr(
img_E, img_H, border=border) # change with your own border
test_results['psnr'].append(psnr)
logger.info('{:->4d}--> {:>10s} --k:{:>2d} PSNR: {:.2f}dB'.format(
idx + 1, img_name + ext, k_index, psnr))
# --------------------------------
# Average PSNR
# --------------------------------
ave_psnr = sum(test_results['psnr']) / len(test_results['psnr'])
logger.info(
'------> Average PSNR of ({}), kernel: ({}) sigma: ({:.2f}): {:.2f} dB'
.format(testset_name, k_index, noise_level_model, ave_psnr))
test_results_ave['psnr'].append(ave_psnr)
def main():
# ----------------------------------------
# load kernels
# ----------------------------------------
PSF_grid = np.load('./data/AC254-075-A-ML-Zemax(ZMX).npz')['PSF']
PSF_grid = PSF_grid.astype(np.float32)
gx, gy = PSF_grid.shape[:2]
for xx in range(gx):
for yy in range(gy):
PSF_grid[xx, yy] = PSF_grid[xx, yy] / np.sum(PSF_grid[xx, yy],
axis=(0, 1))
# ----------------------------------------
# load model
# ----------------------------------------
stage = 8
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = net(n_iter=stage,
h_nc=64,
in_nc=4,
out_nc=3,
nc=[64, 128, 256, 512],
nb=2,
act_mode="R",
downsample_mode='strideconv',
upsample_mode="convtranspose")
model_code = 'iter17000'
loaded_state = torch.load(
'/home/xiu/databag/deblur/models/ZEMAX/uabcnet_{}.pth'.format(
model_code))
model.load_state_dict(loaded_state, strict=True)
model.eval()
for _, v in model.named_parameters():
v.requires_grad = False
model = model.to(device)
img_names = glob.glob(
'/home/xiu/databag/deblur/ICCV2021/suo_image/*/AC254-075-A-ML-Zemax(ZMX).bmp'
)
img_names.sort()
for img_id, img_name in enumerate(img_names):
img_L = cv2.imread(img_name)
img_L = img_L.astype(np.float32)
W, H = img_L.shape[:2]
num_patch = [6, 8]
#positional alpha-beta parameters for HQS
ab_numpy = np.loadtxt(
'/home/xiu/databag/deblur/models/ZEMAX/ab_{}.txt'.format(
model_code)).astype(np.float32).reshape(gx, gy, stage * 2, 3)
ab = torch.tensor(ab_numpy, device=device, requires_grad=False)
#save img_L
t0 = time.time()
px_start = 0
py_start = 0
PSF_patch = PSF_grid[px_start:px_start + num_patch[0],
py_start:py_start + num_patch[1]]
#block_expand = 1
patch_L = img_L[px_start * W // gx:(px_start + num_patch[0]) * W // gx,
py_start * H // gy:(py_start + num_patch[1]) * H //
gy, :]
p_W, p_H = patch_L.shape[:2]
expand = max(PSF_grid.shape[2] // 2, p_W // 16)
block_expand = expand
patch_L_wrap = util_deblur.wrap_boundary_liu(
patch_L, (p_W + block_expand * 2, p_H + block_expand * 2))
#centralize
patch_L_wrap = np.hstack((patch_L_wrap[:, -block_expand:, :],
patch_L_wrap[:, :p_H + block_expand, :]))
patch_L_wrap = np.vstack((patch_L_wrap[-block_expand:, :, :],
patch_L_wrap[:p_W + block_expand, :, :]))
x = util.uint2single(patch_L_wrap)
x = util.single2tensor4(x)
k_all = []
for h_ in range(num_patch[1]):
for w_ in range(num_patch[0]):
k_all.append(util.single2tensor4(PSF_patch[w_, h_]))
k = torch.cat(k_all, dim=0)
[x, k] = [el.to(device) for el in [x, k]]
ab_patch = F.softplus(ab[px_start:px_start + num_patch[0],
py_start:py_start + num_patch[1]])
cd = []
for h_ in range(num_patch[1]):
for w_ in range(num_patch[0]):
cd.append(ab_patch[w_:w_ + 1, h_])
cd = torch.cat(cd, dim=0)
x_E = model.forward_patchwise(x, k, cd, num_patch, [W // gx, H // gy])
x_E = x_E[..., block_expand:block_expand + p_W,
block_expand:block_expand + p_H]
patch_L = patch_L_wrap.astype(np.uint8)
patch_E = util.tensor2uint(x_E)
t1 = time.time()
print('[{}/{}]: {} s per frame'.format(img_id, len(img_names),
t1 - t0))
xk = patch_E
xk = xk.astype(np.uint8)
cv2.imshow('res', xk)
cv2.imshow('input', patch_L.astype(np.uint8))
key = cv2.waitKey(-1)
if key == ord('q'):
break
