if __name__ == "__main__":
# file is tested with 4 different noise levels
# 25, 33, 40 db
data_dir = "../../sparse_ct/data/"
fname = data_dir + "benchmark_ellipses/6.png"
gt, sinogram, theta, FOCUS = image_to_sparse_sinogram(fname,
channel=1,
n_proj=64,
size=512,
angle1=0.0,
angle2=180.0,
noise_pow=30.0)
recon_sart = SartReconstructor('SART',
sart_n_iter=40,
sart_relaxation=0.15)
recon_dip = DgrReconstructor(
'DGR',
dip_n_iter=4001,
net='skip',
lr=0.01,
reg_std=1. / 100,
w_proj_loss=1.00,
# w_perceptual_loss=0.01,
w_tv_loss=0.00,
w_ssim_loss=0.00)
img_sart = recon_sart.calc(sinogram, theta)
recon_dip.set_for_metric(gt, img_sart, FOCUS=FOCUS, log_dir='log/')
img_dip = recon_dip.calc(sinogram, theta)
IRadonReconstructor,
SartReconstructor,
SartTVReconstructor)
if __name__ == "__main__":
fname = "../data/benchmark_human/20.png"
gt, sinogram, theta, FOCUS = image_to_sparse_sinogram(fname, channel=1, n_proj=32, size=512, noise_pow=40.0 )
n_iter = 40
recon_fbp = IRadonReconstructor('FBP')
recon_sart_95 = SartReconstructor('SART95', sart_n_iter=n_iter, sart_relaxation=0.95)
recon_sart_90 = SartReconstructor('SART90', sart_n_iter=n_iter, sart_relaxation=0.90)
recon_sart_80 = SartReconstructor('SART80', sart_n_iter=n_iter, sart_relaxation=0.80)
recon_sart_50 = SartReconstructor('SART50', sart_n_iter=n_iter, sart_relaxation=0.50)
recon_sart_30 = SartReconstructor('SART30', sart_n_iter=n_iter, sart_relaxation=0.30)
recon_sart_15 = SartReconstructor('SART15', sart_n_iter=n_iter, sart_relaxation=0.15)
recon_sart_07 = SartReconstructor('SART07', sart_n_iter=n_iter, sart_relaxation=0.07)
recon_sart_05 = SartReconstructor('SART05', sart_n_iter=n_iter, sart_relaxation=0.05)
recon_sart_03 = SartReconstructor('SART03', sart_n_iter=n_iter, sart_relaxation=0.03)
recon_sart_02 = SartReconstructor('SART02', sart_n_iter=n_iter, sart_relaxation=0.02)
recons = [
recon_fbp,
recon_sart_95,
recon_sart_90,
recon_sart_80,
def benchmark(
images_path,
recon,
theta,
noise_pow
):
images=get_images(images_path)
log_filename = 'benchmark_{recon}'.format(
recon=recon.name
)
logging.basicConfig(
filename='{}.log'.format(log_filename),
filemode='a',
format='%(asctime)s - %(levelname)s - %(message)s',
level=logging.INFO
)
logging.warning('Starting')
logging.warning('images: %s', images_path)
logging.warning('n_proj: %s', len(theta))
logging.warning('noise_pow: %s', noise_pow)
mse_list = []
psnr_list = []
ssim_list = []
for fname in images:
gt, sinogram, theta, FOCUS = image_to_sparse_sinogram(fname,
channel=1, n_proj=len(theta), size=512,
angle1=0.0, angle2=180.0, noise_pow=noise_pow)
# set metrics
if type(recon) == DgrReconstructor or type(recon) == N2SelfReconstructor:
recon_bm3d = SartReconstructor('SART',
sart_n_iter=40, sart_relaxation=0.15)
img_sart_bm3d = recon_bm3d.calc(sinogram, theta)
recon.set_for_metric(gt, img_sart_bm3d, FOCUS=FOCUS, log_dir='../log/dip')
recon.calc(sinogram, theta)
mse, psnr, ssim = recon.eval(gt)
mse_list.append(mse)
psnr_list.append(psnr)
ssim_list.append(ssim)
logstr = "{}: MSE:{:.5f} PSNR:{:.5f} SSIM:{:.5f}".format(
fname, mse, psnr, ssim
)
logging.info(logstr)
logging.info('Avg: MSE:{:.5f} PSNR:{:.5f} SSIM:{:.5f}'.format(
np.mean(mse_list),
np.mean(psnr_list),
np.mean(ssim_list)
))
logging.info('Std: MSE:{:.5f} PSNR:{:.5f} SSIM:{:.5f}'.format(
np.std(mse_list),
np.std(psnr_list),
np.std(ssim_list)
))
logging.warning('Done.')
if __name__ == "__main__":
fname = "../data/ct1.jpg"
gt, sinogram, FOCUS = image_to_sparse_sinogram(fname,
channel=1,
n_proj=32,
size=512,
noise_pow=25.0)
n_iter = 40
recons = [
IRadonReconstructor('FBP'),
SartReconstructor('SART', sart_n_iter=n_iter, sart_relaxation=0.15),
SartTVReconstructor('SART+TVw0.2',
sart_n_iter=n_iter,
sart_relaxation=0.15,
tv_weight=0.2,
tv_n_iter=100),
SartTVReconstructor('SART+TVw0.5',
sart_n_iter=n_iter,
sart_relaxation=0.15,
tv_weight=0.5,
tv_n_iter=100),
SartTVReconstructor('SART+TVw0.8',
sart_n_iter=n_iter,
sart_relaxation=0.15,
tv_weight=0.8,
tv_n_iter=100),
def test(fname, label, n_proj=32, noise_pow=25.0):
dgr_iter = 4000
lr = 0.01
net = 'skip'
noise_std = 1./100
gt, sinogram, theta, FOCUS = image_to_sparse_sinogram(fname,
channel=1, n_proj=n_proj, size=512,
angle1=0.0, angle2=180.0, noise_pow=noise_pow)
logging.warning('Starting')
logging.warning('fname: %s %s',label, fname)
logging.warning('n_proj: %s', n_proj)
logging.warning('noise_pow: %s', noise_pow)
logging.warning('dgr_n_iter: %s', dgr_iter)
logging.warning('dgr_lr: %s', lr)
logging.warning('dgr_net: %s', net)
logging.warning('dgr_noise_std: %s', noise_std)
recons = [
IRadonReconstructor('FBP'),
SartReconstructor('SART', sart_n_iter=40, sart_relaxation=0.15),
SartTVReconstructor('SART+TV',
sart_n_iter=40, sart_relaxation=0.15,
tv_weight=0.5, tv_n_iter=100),
SartBM3DReconstructor('SART+BM3D',
sart_n_iter=40, sart_relaxation=0.15,
bm3d_sigma=0.5),
DgrReconstructor('DIP_1.00_0.00_0.00_0.00',
dip_n_iter=dgr_iter,
net=net,
lr=lr,
reg_std=noise_std,
w_proj_loss=1.0,
w_perceptual_loss=0.0,
w_tv_loss=0.0
),
DgrReconstructor('DIP_0.99_0.01_0.00_0.00',
dip_n_iter=dgr_iter,
net=net,
lr=lr,
reg_std=noise_std,
w_proj_loss=0.99,
w_perceptual_loss=0.01,
w_tv_loss=0.0
),
DgrReconstructor('DIP_0.90_0.10_0.00_0.00',
dip_n_iter=dgr_iter,
net=net,
lr=lr,
reg_std=noise_std,
w_proj_loss=0.90,
w_perceptual_loss=0.10,
w_tv_loss=0.0
),
DgrReconstructor('DIP_0.50_0.50_0.00_0.00',
dip_n_iter=dgr_iter,
net=net,
lr=lr,
reg_std=noise_std,
w_proj_loss=0.5,
w_perceptual_loss=0.5,
w_tv_loss=0.0
),
DgrReconstructor('DIP_0.10_0.90_0.00_0.00',
dip_n_iter=dgr_iter,
net=net,
lr=lr,
reg_std=noise_std,
w_proj_loss=0.10,
w_perceptual_loss=0.90,
w_tv_loss=0.0
),
DgrReconstructor('DIP_0.01_0.99_0.00_0.00',
dip_n_iter=dgr_iter,
net=net,
lr=lr,
reg_std=noise_std,
w_proj_loss=0.01,
w_perceptual_loss=0.99,
w_tv_loss=0.0
),
DgrReconstructor('DIP_0.00_1.00_0.00_0.00',
dip_n_iter=dgr_iter,
net=net,
lr=lr,
reg_std=noise_std,
w_proj_loss=0.0,
w_perceptual_loss=1.0,
w_tv_loss=0.0
),
DgrReconstructor('DIP_0.99_0.00_0.01_0.00',
dip_n_iter=dgr_iter,
net=net,
lr=lr,
reg_std=noise_std,
w_proj_loss=0.99,
w_perceptual_loss=0.0,
w_tv_loss=0.01
),
DgrReconstructor('DIP_0.90_0.00_0.10_0.00',
dip_n_iter=dgr_iter,
net=net,
lr=lr,
reg_std=noise_std,
w_proj_loss=0.9,
w_perceptual_loss=0.0,
w_tv_loss=0.1
),
DgrReconstructor('DIP_0.50_0.00_0.50_0.00',
dip_n_iter=dgr_iter,
net=net,
lr=lr,
reg_std=noise_std,
w_proj_loss=0.5,
w_perceptual_loss=0.0,
w_tv_loss=0.5
),
DgrReconstructor('DIP_0.10_0.00_0.90_0.00',
dip_n_iter=dgr_iter,
net=net,
lr=lr,
reg_std=noise_std,
w_proj_loss=0.1,
w_perceptual_loss=0.0,
w_tv_loss=0.9
),
DgrReconstructor('DIP_0.01_0.00_0.99_0.00',
dip_n_iter=dgr_iter,
net=net,
lr=lr,
reg_std=noise_std,
w_proj_loss=0.01,
w_perceptual_loss=0.0,
w_tv_loss=0.99
),
DgrReconstructor('DIP_0.00_0.00_1.0_0.00',
dip_n_iter=dgr_iter,
net=net,
lr=lr,
reg_std=noise_std,
w_proj_loss=0.00,
w_perceptual_loss=0.0,
w_tv_loss=1.0
),
DgrReconstructor('DIP_0.33_0.33_0.33_0.00',
dip_n_iter=dgr_iter,
net=net,
lr=lr,
reg_std=noise_std,
w_proj_loss=0.33,
w_perceptual_loss=0.33,
w_tv_loss=0.33
),
DgrReconstructor('DIP_0.8_0.10_0.10_0.00',
dip_n_iter=dgr_iter,
net=net,
lr=lr,
reg_std=noise_std,
w_proj_loss=0.8,
w_perceptual_loss=0.1,
w_tv_loss=0.1
),
DgrReconstructor('DIP_0.98_0.01_0.01_0.00',
dip_n_iter=dgr_iter,
net=net,
lr=lr,
reg_std=noise_std,
w_proj_loss=0.98,
w_perceptual_loss=0.01,
w_tv_loss=0.01
),
DgrReconstructor('DIP_0.10_0.80_0.10_0.00',
dip_n_iter=dgr_iter,
net=net,
lr=lr,
reg_std=noise_std,
w_proj_loss=0.10,
w_perceptual_loss=0.80,
w_tv_loss=0.10
),
DgrReconstructor('DIP_0.01_0.98_0.01_0.00',
dip_n_iter=dgr_iter,
net=net,
lr=lr,
reg_std=noise_std,
w_proj_loss=0.01,
w_perceptual_loss=0.98,
w_tv_loss=0.01
),
]
img_sart_bm3d = recons[3].calc(sinogram, theta)
imgs = []
for recon in recons:
if type(recon) == DgrReconstructor:
recon.set_for_metric(gt, img_sart_bm3d, FOCUS=FOCUS, log_dir='../log/dip')
imgs.append(recon.calc(sinogram))
mse, psnr, ssim = recon.eval(gt)
recon.save_result()
logstr = "{}: MSE:{:.5f} PSNR:{:.5f} SSIM:{:.5f}".format(
recon.name, mse, psnr, ssim
)
logging.info(logstr)
plot_grid([gt] + imgs,
FOCUS=FOCUS, save_name=label+'.png', dpi=500)
logging.warning('Done. Results saved as %s', label+'.png')
def benchmark_all(recon, data_list, theta_list):
for data in data_list:
for theta in theta_list:
benchmark(
data,
recon,
theta,
40.0
)
recon_fbp = IRadonReconstructor('FBP')
recon_sart = SartReconstructor('SART',
sart_n_iter=40,
sart_relaxation=0.15)
recon_sart_tv = SartTVReconstructor('SART+TVw0.9',
sart_n_iter=40, sart_relaxation=0.15,
tv_weight=0.9, tv_n_iter=100)
recon_sart_bm3d = SartBM3DReconstructor('SART+BM3Ds0.35',
sart_n_iter=40, sart_relaxation=0.15,
bm3d_sigma=0.35)
recon_dgr = DgrReconstructor('DGR_0.80_0.00_0.10_0.10',
dip_n_iter=4001,
net='skip',
lr=0.01,
reg_std=1./100,
w_proj_loss=0.80,