def _fit_2d_gaussian(subimage, implementation='agpy'):
"""
Fit 2D Gaussian to a 5x5 pixel area. This is a function internal to
pflib that abstracts away the specific methods used to fit.
See fit_psf for the description of the 2D Gaussian model,
parameters, and return values.
Arguments:
subimage: 5x5 Numpy array of pixels to fit.
algorithm: The name of the fitter to use. Currently, only agpy
is supported.
Returns:
(h_0, w_0, H, A, sigma_h, sigma_w, theta, fit_img)
"""
assert subimage.shape[0] == 5 and subimage.shape[1] == 5
if implementation != 'agpy':
raise NotImplementedError("Currently, only agpy is supported.")
#agpy.gaussfit parameters --
#(H(eight), A(mplitude), h_0, w_0, sigma_h, sigma_w, theta)
#initial fitting parameters based on emperical eyeball shufti
((H, A, h_0, w_0, sigma_h, sigma_w, theta), fit_img) = \
agpy.gaussfit(subimage,
params=(np.mean(subimage), np.amax(subimage),
2.5, 2.5, 1, 1, 0),
return_all=False,
limitedmin=[True] * 7,
limitedmax=[False, False, True, True, True, True,
True],
minpars=np.array([0.00,
(np.amax(subimage) -
np.mean(subimage)) / 3.0,
2.00, 2.00, 0.75, 0.75, 0.00]),
maxpars=np.array([0.00, 0.00, 3.00, 3.00, 2.00,
2.00, 360.00]),
returnfitimage=True)
return (h_0, w_0, H, A, sigma_h, sigma_w, theta, fit_img)
def _fit_2d_gaussian(subimage, implementation='agpy'):
"""
Fit 2D Gaussian to a 5x5 pixel area. This is a function internal to
pflib that abstracts away the specific methods used to fit.
See fit_psf for the description of the 2D Gaussian model,
parameters, and return values.
Arguments:
subimage: 5x5 Numpy array of pixels to fit.
algorithm: The name of the fitter to use. Currently, only agpy
is supported.
Returns:
(h_0, w_0, H, A, sigma_h, sigma_w, theta, fit_img)
"""
assert subimage.shape[0] == 5 and subimage.shape[1] == 5
if implementation != 'agpy':
raise NotImplementedError("Currently, only agpy is supported.")
#agpy.gaussfit parameters --
#(H(eight), A(mplitude), h_0, w_0, sigma_h, sigma_w, theta)
#initial fitting parameters based on emperical eyeball shufti
((H, A, h_0, w_0, sigma_h, sigma_w, theta), fit_img) = \
agpy.gaussfit(subimage,
params=(np.mean(subimage), np.amax(subimage),
2.5, 2.5, 1, 1, 0),
return_all=False,
limitedmin=[True] * 7,
limitedmax=[False, False, True, True, True, True,
True],
minpars=np.array([0.00,
(np.amax(subimage) -
np.mean(subimage)) / 3.0,
2.00, 2.00, 0.75, 0.75, 0.00]),
maxpars=np.array([0.00, 0.00, 3.00, 3.00, 2.00,
2.00, 360.00]),
returnfitimage=True)
return (h_0, w_0, H, A, sigma_h, sigma_w, theta, fit_img)
def fit_peaks(images, image_tuples=None, correlation_matrix=None,
median_diameter=5, r_2_threshold=0.5, c_std=4,
consolidation_radius_sq=4**2, child_number=0,
child_priority_increment=0, silent=False):
"""
The core function of the algorithm. In each image, finds all peaks and
evaluates their resemblance to a Gaussian PSF model.
"""
#renice
os.nice(child_priority_increment)
#generate empty image metadata tuples if empty
if image_tuples is None:
if not silent:
logging.info('pflib.fit_peaks child number ' + str(child_number) +
': no image_tuples metadata passed, using empty ' +
'tuples')
image_tuples = [() for t in images]
elif len(image_tuples) != len(images):
if not silent:
logging.info('pflib.fit_peaks child number ' + str(child_number) +
': number of image metadata tuples does not match number of ' +
'images, using empty tuples')
image_tuples = [() for t in images]
#perform squareroot just this once for the entire function call
consolidation_box = np.ceil(math.sqrt(consolidation_radius_sq))
#copy images to prevent tampering with originals by any function called
images = [np.copy(image).astype(np.int64) for image in images]
images_s3 = images
if not silent:
logging.info("pflib.fit_peaks child number " + str(child_number) +
": copied images")
sys.stdout.flush()
#TODO:the following line does not check that images remain within bounds as
#spalttiefpass and gaussian may increase positive values
#subtract median filter
images_s3 = \
[np.subtract(image, np.minimum(spf.median_filter(image, median_diameter),
image))
for image in images_s3]
if not silent:
logging.info("pflib.fit_peaks child number " + str(child_number) +
": applied median filter")
sys.stdout.flush()
#perform correlation
if correlation_matrix is not None:
images_s3 = [np.maximum(scipy.signal.correlate(image,
correlation_matrix,
mode='same'),
np.zeros_like(image)).astype(np.int64)
for image in images]
#if not silent:
# logging.info("pflib.fit_peaks child number " + str(child_number) +
# ": applied correlation filter")
#sys.stdout.flush()
consolidated_peak_stack = [[] for image in images]
for i, image in enumerate(images_s3):
start_time = time.clock()
#number of fit attempts
fit_counter = 0
#no dictionary comprehensions in python 2.6 :(
#pixel_bins = {(h,w): [] for h in range(image.shape[0])
# for w in range(image.shape[1])}
#pixel_bins = dict([((h,w), []) for h in range(image.shape[0])
# for w in range(image.shape[1])])
pixel_bins = [[[] for w in range(image.shape[1])]
for h in range(image.shape[0])]
#threshold beneath which a pixel needs not to be checked
candidate_threshold = np.mean(image) + c_std * np.std(image)
for h, w in itertools.product(range(image.shape[0]),
range(image.shape[1])):
#keep track of progress within each image
peak_time = time.clock()
if (not 1 < h < image.shape[0] - 2 or
not 1 < w < image.shape[1] - 2):
continue
if image[h, w] < candidate_threshold:
continue
fit_counter = fit_counter + 1
subimage = images[i][h - 2:h + 3, w - 2:w + 3]
#agpy.gaussfit parameters --
#(height, amplitude, x, y, width_x, width_y, rota)
((fit, fit_err), fit_image) = \
agpy.gaussfit(subimage,
params=(np.mean(subimage),
np.amax(subimage),
2.5, 2.5, 1, 1, 0),
return_all=True,
limitedmin = [True] * 7,
limitedmax = [False, False,
True, True, True, True, True],
minpars=np.array([0.00,
(np.amax(subimage) -
np.mean(subimage)) / 3.0,
2.00, 2.00, 0.75, 0.75,
0.00]),
maxpars=np.array([0.00, 0.00, 3.00, 3.00,
2.00, 2.00, 360.00]),
returnfitimage=True)
#rmse
rmse = math.sqrt(
sum([(subimage[x,y] - fit_image[x,y])**2
for x,y in itertools.product(range(5),range(5))])
/ 9.0)
#coefficient of determination r**2
r_2 = (1.0 -
sum(np.reshape((subimage - fit_image)**2, -1)) /
sum((np.reshape(subimage, -1) - np.mean(subimage))**2))
if r_2 < r_2_threshold:
continue
#Illumina S/N =
#([peak pixel] - [mean of 16 outer edge pixels]) /
# (std[16 outer pixels])
op = [subimage[x, y]
for x in range(subimage.shape[0])
for y in range(subimage.shape[1])
if x == 0 or
x == subimage.shape[0] - 1 or
y == 0 or
y == subimage.shape[1] - 1]
s_n = (np.amax(subimage) - np.mean(op)) / np.std(op)
#peak fitting result tuple
#(0 1 2-- 3------ 4------- 5-------- 6--- 7--, 8--)
#(h, w, fit, fit_err, subimage, fit_image, rmse, r_2, s_n)
#agpy.gaussfit parameters --
# 0----- 1-------- 2 3 4------ 5------ 6---
#(height, amplitude, x, y, width_x, width_y, rota)
peak = (fit[2] + h - 2, fit[3] + w - 2, fit, fit_err, subimage,
fit_image, rmse, r_2, s_n)
#for all fits, consolidate them into the best one within a radius;
#check peaks in the consolidated peak stack
#default is that this is a new peak
if not (0 <= peak[0] <= image.shape[0] and
0 <= peak[1] <= image.shape[1]):
continue
(pixel_bins[int(np.around(peak[0]))]
[int(np.around(peak[1]))].append(peak))
if (not silent
and
(h + w) % int(image.shape[0] * image.shape[1] / 1000) == 0):
logging.info("pflib.fitpeaks child number " +
str(child_number) +
": done fitting at h,w " + str((h,w)) + " in " +
str(time.clock() - peak_time) + " sec; layer " +
str(i) + ' of ' + str(len(images_s3)) +
"; cumulative time " +
str(time.clock() - start_time) +
"; fit_counter " + str(fit_counter))
sys.stdout.flush()
consolidation_time = time.clock()
for h, w in itertools.product(range(image.shape[0]),
range(image.shape[1])):
if pixel_bins[h][w]:
pixel_bins[h][w] = max(pixel_bins[h][w],
key=lambda x: x[7])
else:
pixel_bins[h][w] = None
for h, w in itertools.product(range(image.shape[0]),
range(image.shape[1])):
if not pixel_bins[h][w]:
continue
local_max = True
for j, q in itertools.product(
range(int(max(h - consolidation_box - 2, 0)),
int(min(h + consolidation_box + 3,
image.shape[0]))),
range(int(max(w - consolidation_box - 2, 0)),
int(min(w + consolidation_box + 3,
image.shape[1])))):
if not pixel_bins[j][q] or (j == h and q == w):
continue
if ((pixel_bins[j][q][0] - pixel_bins[h][w][0])**2 +
(pixel_bins[j][q][1] - pixel_bins[h][w][1])**2 <
consolidation_radius_sq):
if pixel_bins[j][q][7] < pixel_bins[h][w][7]:
pixel_bins[j][q] = None
else:
local_max = False
if not local_max:
pixel_bins[h][w] = None
consolidated_peak_stack[i] = \
(image_tuples[i],
[pixel_bins[h][w]
for h, w in itertools.product(range(image.shape[0]),
range(image.shape[1]))
if pixel_bins[h][w]])
if not silent:
logging.info("pflib.fit_peaks child number " + str(child_number) +
": consolidation accomplished in " +
str(time.clock() - consolidation_time) + " sec; " +
str(len(consolidated_peak_stack[i][1])) +
" peaks remaining of " + str(fit_counter) +
" fit_counter")
if not silent:
logging.info("pflib.fit_peaks child number " + str(child_number) +
": image " + str(i + 1) + " of " +
str(len(images_s3)) + " processed in " +
str(time.clock() - start_time) + " sec;" +
" fit_counter " + str(fit_counter))
return consolidated_peak_stack
