def n4_illum_correction_downsample(img, DOWN=2, RATIO=1.05, FILTERINGSIZE=50, OFFSET=10):
"""Faster but more insensitive to local illum bias.
"""
fil = sitk.ShrinkImageFilter()
cc = sitk.GetArrayFromImage(fil.Execute(sitk.GetImageFromArray(img), [DOWN, DOWN]))
bw = adaptive_thresh(cc, R=RATIO, FILTERINGSIZE=FILTERINGSIZE/DOWN)
himg = homogenize_intensity_n4(cc, -bw)
himg = cc - himg
# himg[himg < 0] = 0
bias = resize_img(himg, img.shape)
img = img - bias
return convert_positive(img, OFFSET)
def background_subtraction_wavelet_hazen(img,
THRES=100,
ITER=5,
WLEVEL=6,
OFFSET=50):
"""Wavelet background subtraction.
"""
back = wavelet_subtraction_hazen(img.astype(np.float),
ITER=ITER,
THRES=THRES,
WLEVEL=WLEVEL)
img = img - back
return convert_positive(img, OFFSET)
def background_subtraction_wavelet(img, level=7, OFFSET=10):
'''
It might be radical but works in many cases in terms of segmentation.
Use "background_subtraction_wavelet_hazen" for a proper implementation.
'''
from pywt import WaveletPacket2D
from skimage.transform import resize
def wavelet_subtraction(img, level):
"""6- 7 level is recommended"""
if level == 0:
return img
wp = WaveletPacket2D(data=img.astype(np.uint16), wavelet='haar', mode='sym')
back = resize(np.array(wp['a'*level].data), img.shape, order=3, mode='reflect')/(2**level)
img = img - back
return img
img = wavelet_subtraction(img, level)
return convert_positive(img, OFFSET)
def rolling_ball(img, RADIUS=100, SIGMA=3, OFFSET=50):
"""Rolling ball background subtraction.
"""
back = rolling_ball_subtraction_hazen(img.astype(np.float), RADIUS)
img = img - back
return convert_positive(img, OFFSET)
