def sparse_signal():
dir_path = os.path.dirname(os.path.realpath(__file__))
img = np.load(os.path.join(dir_path, 'test_img1.npy'))
img = img.astype(float)
img = filters.threshold(utils.normalize(img), 0.4)
return preprocess.Patches(img, 8).patches
def test_threshold_hard():
a = np.ones(10)
b1 = filters.threshold(a.copy(), min_val=1.1)
assert np.allclose(b1, a * 0)
b2 = filters.threshold(a.copy(), min_val=0.5, max_val=0.9)
assert np.allclose(b2, a * 0)
b3 = filters.threshold(a.copy(), min_val=0.9)
assert np.allclose(b3, a)
a[:5] = 0.7
b4 = filters.threshold(a.copy(), min_val=0.9)
a[:5] = 0
assert np.allclose(b4, a)
def reconstruct(rng):
rank, start, stop = rng
new_img = IMAGE_VOLUME[start:stop]
for i in range(new_img.shape[0]):
image = new_img[i]
image = utils.normalize(image)
image = filters.threshold(image, 0.4)
patch_size = (16, 16)
sigma = 10
den = dl.Denoise(image, patch_size[0])
den.dictionary = DICT
denoised = den.denoise(sigma)
new_img[i] = denoised.copy()
slices_dir = slice_dir()
fn = 'slice_%d_%d.npy' % (start, stop)
np.save(os.path.join(slices_dir, fn), new_img)
def reconstruct_adept(data):
rank, start, stop = data
image = IMAGE_VOLUME[start:stop]
fn = 'dct_slice_%d.npy'
slice_path = slice_dir()
for i in range(image.shape[0]):
img = IMAGE_VOLUME[i]
img = filters.threshold(utils.normalize(img), 0.4)
patch_size = (16, 16)
sigma = 10
path = os.path.join(slice_path, fn % i)
D = np.load(path)
den = dl.Denoise(img, patch_size[0])
den.dictionary = D
denoised = den.denoise(sigma)
image[i] = denoised.copy()
slices_dir = slice_dir()
fn = 'slice_%d_%d.npy' % (start, stop)
np.save(os.path.join(slices_dir, fn), image)
help='Visualize the dictionary')
parser.add_argument('--smooth', type=int,
help='Apply gaussion smoothing with kernel size SIZE (odd)')
parser.add_argument('-j', '--n-threads', type=int, default=1,
help="Number of threads to use, default 1")
args = parser.parse_args()
image = misc.imread(args.image)
image = utils.normalize(image)
if args.threshold is not None:
image = filters.threshold(image, args.threshold)
image *= 255
patch_size = args.patch_size
n_atoms = args.num_atoms
iters = args.iters
sigma = args.sigma
t1 = timeit.default_timer()
denoised = ksvd_denoise(image, patch_size, iters, n_atoms,
sigma, verbose=True, retDict=args.vis_dict,
n_threads=args.n_threads)
from __future__ import print_function
import os
import sys
import multiprocessing
import numpy as np
import dictlearn as dl
from dictlearn import utils, filters
DICT = np.load('frame_dict_trained.npy')
IMAGE_VOLUME = np.load('images/image_volume.npy')
IMAGE_VOLUME = utils.normalize(IMAGE_VOLUME)
IMAGE_VOLUME = filters.threshold(IMAGE_VOLUME, 0.4)
def slice_dir():
home = os.path.expanduser('~')
slices_dir = os.path.join(home, 'slices')
try:
os.mkdir(slices_dir)
except OSError:
pass
return slices_dir
def train(data):
rank, start, stop = data
image = IMAGE_VOLUME[start:stop]