def make_mask(self, isl, subn, subm, nsrc, src_id, g_list, delc):
import functions as func
# define stuff for calculating gaussian
boxx, boxy = isl.bbox
subn = boxx.stop - boxx.start
subm = boxy.stop - boxy.start
x, y = N.indices((subn, subm))
# construct image of each source in the island
src_image = N.zeros((subn, subm, nsrc), dtype=N.float32)
nn = 1
for isrc in range(nsrc):
if nsrc == 1:
g_sublist = g_list
else:
posn = N.where(src_id == isrc)[0]
g_sublist = []
for i in posn:
g_sublist.append(g_list[i])
for g in g_sublist:
params = func.g2param(g)
params[1] -= delc[0]
params[2] -= delc[1]
gau = func.gaus_2d(params, x, y)
src_image[:, :, isrc] = src_image[:, :, isrc] + gau
# mark each pixel as belonging to one source
# just compare value, should compare with sigma later
mask = N.argmax(src_image, axis=2) + src_id
orig_mask = isl.mask_active
mask[N.where(orig_mask)] = -1
return mask
def make_mask(self, isl, subn, subm, nsrc, src_id, g_list, delc):
import functions as func
# define stuff for calculating gaussian
boxx, boxy = isl.bbox
subn = boxx.stop-boxx.start; subm = boxy.stop-boxy.start
x, y = N.indices((subn, subm))
# construct image of each source in the island
src_image = N.zeros((subn, subm, nsrc), dtype=N.float32)
nn = 1
for isrc in range(nsrc):
if nsrc == 1:
g_sublist = g_list
else:
posn = N.where(src_id == isrc)[0]
g_sublist=[]
for i in posn:
g_sublist.append(g_list[i])
for g in g_sublist:
params = func.g2param(g)
params[1] -= delc[0]; params[2] -= delc[1]
gau = func.gaus_2d(params, x, y)
src_image[:,:,isrc] = src_image[:,:,isrc] + gau
# mark each pixel as belonging to one source
# just compare value, should compare with sigma later
mask = N.argmax(src_image, axis=2) + src_id
orig_mask = isl.mask_active
mask[N.where(orig_mask)] = -1
return mask
def make_subim(self, subn, subm, g_list, delc, mc=False):
import functions as func
subim = N.zeros((subn, subm), dtype=N.float32)
x, y = N.indices((subn, subm))
for g in g_list:
params = func.g2param(g)
params[1] -= delc[0]; params[2] -= delc[1]
if mc:
# draw random variables from distributions given by errors
params_err = func.g2param_err(g)
for i in range(len(params)):
mc_param = N.random.normal(loc=params[i], scale=params_err[i])
params[i] = mc_param
gau = func.gaus_2d(params, x, y)
subim = subim + gau
return subim
def make_subim(self, subn, subm, g_list, delc, mc=False):
import functions as func
subim = N.zeros((subn, subm), dtype=N.float32)
x, y = N.indices((subn, subm))
for g in g_list:
params = func.g2param(g)
params[1] -= delc[0]; params[2] -= delc[1]
if mc:
# draw random variables from distributions given by errors
params_err = func.g2param_err(g)
for i in range(len(params)):
mc_param = N.random.normal(loc=params[i], scale=params_err[i])
params[i] = mc_param
gau = func.gaus_2d(params, x, y)
subim = subim + gau
return subim
def inigaus_nobeam(self, isl, thr, beam, img):
""" To get initial guesses when the source sizes are very different
from the beam, and can also be elongated. Mainly in the context of
a-trous transform images. Need to arrive at a good guess of the sizes
and hence need to partition the image around the maxima first. Tried the
IFT watershed algo but with markers, it segments the island only around
the minima and not the whole island. Cant find a good weighting scheme
for tesselation either. Hence will try this :
Calculate number of maxima. If one, then take moment as initial
guess. If more than one, then moment of whole island is one of the
guesses if mom1 is within n pixels of one of the maxima. Else dont take
whole island moment. Instead, find minima on lines connecting all maxima
and use geometric mean of all minima of a peak as the size of that peak.
"""
from math import sqrt
from const import fwsig
import scipy.ndimage as nd
import functions as func
im = isl.image-isl.islmean
if img.opts.ini_method == 'curvature':
im_pos = -1.0 * func.make_curvature_map(isl.image-isl.islmean)
thr_pos = 0.0
else:
im_pos = im
thr_pos = -1e9
mask = isl.mask_active
av = img.clipped_mean
inipeak, iniposn, im1 = func.get_maxima(im, mask, thr_pos, isl.shape, beam, im_pos=im_pos)
npeak = len(iniposn)
gaul = []
av, stdnew, maxv, maxp, minv, minp = func.arrstatmask(im, mask)
mom = func.momanalmask_gaus(isl.image-isl.islmean, isl.mask_active, 0, 1.0, True)
if npeak <= 1:
g = (float(maxv), int(round(mom[1])), int(round(mom[2])), mom[3]/fwsig, \
mom[4]/fwsig, mom[5])
gaul.append(g)
if npeak > 1: # markers start from 1=background, watershed starts from 1=background
watershed, markers = func.watershed(im, mask=isl.mask_active)
nshed = N.max(markers)-1 # excluding background
xm, ym = N.transpose([N.where(markers==i) for i in range(1,nshed+2)])[0]
coords = [c for c in N.transpose([xm,ym])[1:]]
alldists = [func.dist_2pt(c1, c2) for c1 in coords for c2 in coords if N.any(c1!=c2)] # has double
meandist = N.mean(alldists) # mean dist between all pairs of markers
compact = []; invmask = []
for ished in range(nshed):
shedmask = N.where(watershed==ished+2, False, True) + isl.mask_active # good unmasked pixels = 0
imm = nd.binary_dilation(~shedmask, N.ones((3,3), int))
xbad, ybad = N.where((imm==1)*(im>im[xm[ished+1], ym[ished+1]]))
imm[xbad, ybad] = 0
invmask.append(imm); x, y = N.where(imm); xcen, ycen = N.mean(x), N.mean(y) # good pixels are now = 1
dist = func.dist_2pt([xcen, ycen], [xm[ished+1], ym[ished+1]])
if dist < max(3.0, meandist/4.0):
compact.append(True) # if not compact, break source + diffuse
else:
compact.append(False)
if not N.all(compact):
avsize = []
ind = N.where(compact)[0]
for i in ind: avsize.append(N.sum(invmask[i]))
avsize = sqrt(N.mean(N.array(avsize)))
for i in range(len(compact)):
if not compact[i]: # make them all compact
newmask = N.zeros(imm.shape, bool)
newmask[max(0,xm[i+1]-avsize/2):min(im.shape[0],xm[i+1]+avsize/2), \
max(0,ym[i+1]-avsize/2):min(im.shape[1],ym[i+1]+avsize/2)] = True
invmask[i] = invmask[i]*newmask
resid = N.zeros(im.shape, dtype=N.float32) # approx fit all compact ones
for i in range(nshed):
mask1 = ~invmask[i]
size = sqrt(N.sum(invmask))/fwsig
xf, yf = coords[i][0], coords[i][1]
p_ini = [im[xf, yf], xf, yf, size, size, 0.0]
x, y = N.indices(im.shape)
p, success = func.fit_gaus2d(im*invmask[i], p_ini, x, y)
resid = resid + func.gaus_2d(p, x, y)
gaul.append(p)
resid = im - resid
if not N.all(compact): # just add one gaussian to fit whole unmasked island
maxv = N.max(resid) # assuming resid has only diffuse emission. can be false
x, y = N.where(~isl.mask_active); xcen = N.mean(x); ycen = N.mean(y)
invm = ~isl.mask_active
#bound = invm - nd.grey_erosion(invm, footprint = N.ones((3,3), int)) # better to use bound for ellipse fitting
mom = func.momanalmask_gaus(invm, N.zeros(invm.shape, dtype=N.int16), 0, 1.0, True)
g = (maxv, xcen, ycen, mom[3]/fwsig, mom[4]/fwsig, mom[5]-90.)
gaul.append(g)
coords.append([xcen, ycen])
return gaul
