def __init__(self, img):
self.img = img#threadsHolding(img)
#self.pixels = pixels
print("Extracting LBP...")
self.lbpR1P8 = normalize(np.float32(lbp.lbp(self.img, 1, 8, ignore_zeros=False)))
self.lbpR2P16 = normalize(np.float32(lbp.lbp(self.img, 2, 16, ignore_zeros=False)))
self.lbpR3P16 = normalize(np.float32(lbp.lbp(self.img, 3, 16, ignore_zeros=False)))
#self.lbpR3P24 = normalize(np.float32(lbp.lbp(self.img, 3, 24, ignore_zeros=False)))
self.lbpR1P8_R2P16 = np.float32(np.concatenate([self.lbpR1P8, self.lbpR2P16]))
self.lbpR1P8_R3P16 = np.float32(np.concatenate([self.lbpR1P8, self.lbpR3P16]))
self.lbpR2P16_R3P16 = np.float32(np.concatenate([self.lbpR2P16, self.lbpR3P16]))
#self.lbpR1P8pic = np.float32(lbp.lbp_transform(self.img, 1, 8, ignore_zeros=False))
#self.lbpR2P16pic = np.float32(lbp.lbp_transform(self.img, 2, 16, ignore_zeros=False))
self.lbpR3P16pic = np.float32(lbp.lbp_transform(self.img, 3, 16, ignore_zeros=False))
#self.lbpR3P24pic = np.float32(lbp.lbp_transform(self.img, 3, 24, ignore_zeros=False))
'''
for imagename in image_targets: image_counter+=1 filename = args.trainingSetPath + imagename print 'Processing image: ', image_counter, filename # Abro la imagen en escala de grises image = mahotas.imread(filename, as_grey=True) if (args.featuresGray != ""): # Extraer histograma de grises gray_hist, bins = np.histogram(image.flatten(), 256, [0, 256]) gray_features[imagename] = gray_hist # if (args.featuresLbp != ""): # Extraer lbp (sobre la misma imagen del anterior, en escala de grises) radius = 3 points = 4 * radius # Number of points to be considered as neighbourers lbp_hist = lbp.lbp(image, radius, points, ignore_zeros=False) lbp_features[imagename] = lbp_hist # if (args.featuresSurf != ""): # Extraer surf (sobre la misma imagen del anterior, en escala de grises) surf_features = surf.surf(image)[:, 5:] surf_all_hist[imagename] = surf_features # # Abro la imagen en colores image = mahotas.imread(filename, as_grey=False) # if (args.featuresColor != ""): # Extraer histograma de colores color_hist, bins = np.histogram(image.flatten(), 256, [0, 256]) color_features[imagename] = color_hist # if (args.featuresHara != ""): # Extraer histograma haraclick (sobre la misma imagen del anterior, en colores) hara_hist = mahotas.features.haralick(image).mean(0)
def test_shape():
A = np.arange(32*32).reshape((32,32))
B = np.arange(64*64).reshape((64,64))
features0 = lbp(A, 3, 12)
features1 = lbp(B, 3, 12)
assert features0.shape == features1.shape
def test_histogram_large():
A = np.arange(32*32).reshape((32,32))
for r in (2,3,4,5):
assert lbp(A,r,12).sum() == A.size
def test_nonzero():
A = np.arange(32*32).reshape((32,32))
features = lbp(A, 3, 12)
features_ignore_zeros = lbp(A * (A> 256), 3, 12, ignore_zeros=True)
assert features.sum() > 0
assert not np.all(features == features_ignore_zeros)
def computefeatures(img, featsets, progress=None, preprocessing=None, **kwargs):
"""
features = computefeatures(img,featsets,progress=None,**kwargs)
Compute features defined by featsets on image img.
featsets can be either a list of feature groups.
Feature groups:
--------------
+ 'har': 13 Haralick features [in matslic]
+ 'har3d': 26 Haralick features (actually this works in 2D as well)
+ 'skl': Skeleton features [in matslic] (syn: 'skel')
+ 'img': Image features [in matslic]
+ 'hul': Hull features [in matslic] (syn: 'hull')
+ 'edg': Edge features [in matslic] (syn: 'edge')
+ 'zer': Zernike moments [in matslic]
+ 'tas' : Threshold Adjacency Statistics
+ 'pftas' : Parameter-free Threshold Adjacency Statistics
Feature set names:
+ 'SLF7dna'
+ 'mcell': field level features
+ 'field+': all field level features. The exact exact number of
features computed is liable to change (increase) in
newer versions
img can be a list of images. In this case, a two-dimensional feature vector will be returned, where
f[i,j] is the j-th feature of the i-th image. Also, in this case, imgs will be unload after feature calculation.
Parameters
----------
* *progress*: if progress is not None, then it should be an integer.
Every *progress* images, an output message will be printed.
* *preprocessing*: Whether to do preprocessing (default: True).
If False and img.channeldata[procprotein|procdna] are empty, they are filled in
using img.channeldata[protein|dna] respectively.
* *options*: currently passed through to pyslic.preprocessimage
"""
if type(featsets) == str:
featsets = _featsfor(featsets)
if type(img) == list:
features = []
for i, im in enumerate(img):
f = computefeatures(im, featsets, progress=None, **kwargs)
features.append(f)
im.unload()
if progress is not None and (i % progress) == 0:
print "Processed %s images..." % i
return numpy.array(features)
regions = img.regions
if regions is not None and regions.max() > 1 and "region" not in kwargs:
precomputestats(img)
return numpy.array(
[
computefeatures(img, featsets, progress=progress, region=r, **kwargs)
for r in xrange(1, regions.max() + 1)
]
)
scale = img.scale
if scale is None:
scale = _Default_Scale
is_surf = len(featsets) == 1 and featsets[0] in ("surf", "surf-ref", "surfp")
if preprocessing is None:
preprocessing = not is_surf
if preprocessing:
preprocessimage(img, kwargs.get("region"), options=kwargs.get("options", {}))
else:
if "procprotein" not in img.channeldata:
img.channeldata["procprotein"] = img.get("protein")
img.channeldata["resprotein"] = 0 * img.get("protein")
if ("dna" in img.channeldata) and ("procdna" not in img.channeldata):
img.channeldata["procdna"] = img.get("dna")
features = numpy.array([])
protein = img.get("protein")
procprotein = img.get("procprotein")
resprotein = img.get("resprotein")
dna = img.channeldata.get("dna")
procdna = img.channeldata.get("procdna")
if procprotein.size < _Min_image_size:
if is_surf:
return np.array([])
if featsets == _featsfor("SLF7dna"):
return np.array([np.nan for i in xrange(90)])
if featsets == _featsfor("field-dna+"):
return np.array([np.nan for i in xrange(173)])
raise ValueError
lbppat = re.compile(r"lbp\(([0-9]+), ?([0-9]+)\)")
for F in featsets:
if F in ["edg", "edge"]:
feats = edgefeatures(procprotein)
elif F == "raw-har":
feats = haralickfeatures(procprotein).mean(0)
elif F[:3] == "har":
img = procprotein
har_scale = kwargs.get("haralick.scale", _Default_Haralick_Scale)
if F == "har" and scale != har_scale:
img = img.copy()
img = ndimage.zoom(img, scale / _Default_Haralick_Scale)
if len(F) > 3:
rate = int(F[3])
if rate != 1:
C = np.ones((rate, rate))
img = np.array(img, np.uint16)
img = ndimage.convolve(img, C)
img = img[::rate, ::rate]
if not img.size:
feats = np.zeros(13)
else:
bins = kwargs.get("haralick.bins", _Default_Haralick_Bins)
if bins != 256:
min = img.min()
max = img.max()
ptp = max - min
if ptp:
img = np.array((img - min).astype(float) * bins / ptp, np.uint8)
feats = haralickfeatures(img)
feats = feats.mean(0)
elif F in ["hul", "hull"]:
feats = hullfeatures(procprotein)
elif F == "hullsize":
feats = hullsizefeatures(procprotein)
elif F == "hullsizedna":
feats = hullsizefeatures(procdna)
elif F == "img":
feats = imgfeaturesdna(procprotein, procdna)
elif F in ("obj-field", "obj-field-dna"):
feats = imgfeaturesdna(procprotein, procdna, isfield=True)
elif F == "mor":
feats = morphologicalfeatures(procprotein)
elif F == "nof":
feats = noffeatures(procprotein, resprotein)
elif F in ["skl", "skel"]:
feats = imgskelfeatures(procprotein)
elif F == "zer":
feats = zernike(procprotein, 12, 34.5, scale)
elif F == "tas":
feats = tas(protein)
elif F == "pftas":
feats = pftas(procprotein)
elif F == "overlap":
feats = overlapfeatures(protein, dna, procprotein, procdna)
elif lbppat.match(F):
radius, points = lbppat.match(F).groups()
feats = lbp(protein, int(radius), int(points))
elif F in ("surf", "surf-ref", "surfp"):
if F == "surfp":
from warnings import warn
warn("surfp is deprecated. Use surf-ref", DeprecationWarning)
if len(featsets) > 1:
raise ValueError("pyslic.features.computefeatures: surf-ref must be computed on its own")
if F == "surf":
dna = None
return surf_ref(protein, dna)
else:
raise Exception("Unknown feature set: %s" % F)
features = numpy.r_[features, feats]
return features
def computefeatures(img, featsets, progress=None, preprocessing=True, **kwargs):
'''
features = computefeatures(img,featsets,progress=None,**kwargs)
Compute features defined by featsets on image img.
featsets can be either a list of feature groups.
Feature groups:
--------------
+ 'har': 13 Haralick features [in matslic]
+ 'har3d': 26 Haralick features (actually this works in 2D as well)
+ 'skl': Skeleton features [in matslic] (syn: 'skel')
+ 'img': Image features [in matslic]
+ 'hul': Hull features [in matslic] (syn: 'hull')
+ 'edg': Edge features [in matslic] (syn: 'edge')
+ 'zer': Zernike moments [in matslic]
+ 'tas' : Threshold Adjacency Statistics
+ 'pftas' : Parameter-free Threshold Adjacency Statistics
Feature set names:
+ 'SLF7dna'
+ 'mcell': field level features
+ 'field+': all field level features. The exact exact number of
features computed is liable to change (increase) in
newer versions
img can be a list of images. In this case, a two-dimensional feature vector will be returned, where
f[i,j] is the j-th feature of the i-th image. Also, in this case, imgs will be unload after feature calculation.
Parameters
----------
* *progress*: if progress is not None, then it should be an integer.
Every *progress* images, an output message will be printed.
* *preprocessing*: Whether to do preprocessing (default: True).
If False and img.channeldata[procprotein|procdna] are empty, they are filled in
using img.channeldata[protein|dna] respectively.
* *options*: currently passed through to pyslic.preprocessimage
'''
if type(featsets) == str:
featsets = _featsfor(featsets)
if type(img) == list:
features=[]
for i,im in enumerate(img):
f = computefeatures(im,featsets,progress=None,**kwargs)
features.append(f)
im.unload()
if progress is not None and (i % progress) == 0:
print 'Processed %s images...' % i
return numpy.array(features)
regions = img.regions
if regions is not None and regions.max() > 1 and 'region' not in kwargs:
precomputestats(img)
return numpy.array([
computefeatures(img, featsets, progress=progress, region=r, **kwargs)
for r in xrange(1,regions.max()+1)])
scale = img.scale
if scale is None:
scale = _Default_Scale
if preprocessing:
preprocessimage(img, kwargs.get('region',1), options=kwargs.get('options',{}))
else:
if not img.channeldata['procprotein']:
img.channeldata['procprotein'] = img.get('protein')
img.channeldata['resprotein'] = 0*img.get('protein')
if not img.channeldata['procdna']:
img.channeldata['procdna'] = img.get('dna')
features = numpy.array([])
protein = img.get('protein')
procprotein = img.get('procprotein')
resprotein = img.get('resprotein')
dna = img.channeldata.get('dna')
procdna = img.channeldata.get('procdna')
if procprotein.size < _Min_image_size:
return np.array([np.nan for i in xrange(90)])
lbppat = re.compile(r'lbp\(([0-9]+), ?([0-9]+)\)')
for F in featsets:
if F in ['edg','edge']:
feats = edgefeatures(procprotein)
elif F == 'raw-har':
feats = haralickfeatures(procprotein).mean(0)
elif F[:3] == 'har':
img = procprotein
har_scale = kwargs.get('haralick.scale',_Default_Haralick_Scale)
if F == 'har' and scale != har_scale:
img = img.copy()
img = ndimage.zoom(img, scale/_Default_Haralick_Scale)
if len(F) > 3:
rate = int(F[3])
if rate != 1:
C = np.ones((rate,rate))
img = np.array(img,np.uint16)
img = ndimage.convolve(img,C)
img = img[::rate,::rate]
if not img.size:
feats = np.zeros(13)
else:
bins = kwargs.get('haralick.bins',_Default_Haralick_Bins)
if bins != 256:
min = img.min()
max = img.max()
ptp = max - min
if ptp:
img = np.array((img-min).astype(float) * bins/ptp, np.uint8)
feats = haralickfeatures(img)
feats = feats.mean(0)
elif F == 'har3d':
feats = haralickfeatures(procprotein)
feats = numpy.r_[feats.mean(0),feats.ptp(0)]
elif F in ['hul', 'hull']:
feats = hullfeatures(procprotein)
elif F == 'hullsize':
feats = hullsizefeatures(procprotein)
elif F == 'hullsizedna':
feats = hullsizefeatures(procdna)
elif F == 'img':
feats = imgfeaturesdna(procprotein, procdna)
elif F in ('obj-field', 'obj-field-dna'):
feats = imgfeaturesdna(procprotein, procdna, isfield=True)
elif F == 'mor':
feats = morphologicalfeatures(procprotein)
elif F == 'nof':
feats = noffeatures(procprotein,resprotein)
elif F in ['skl', 'skel']:
feats = imgskelfeatures(procprotein)
elif F == 'zer':
feats = zernike(procprotein,12,34.5,scale)
elif F == 'tas':
feats = tas(protein)
elif F == 'pftas':
feats = pftas(procprotein)
elif F == 'overlap':
feats = overlapfeatures(protein, dna, procprotein, procdna)
elif lbppat.match(F):
radius,points = lbppat.match(F).groups()
feats = lbp(protein, int(radius), int(points))
else:
raise Exception('Unknown feature set: %s' % F)
features = numpy.r_[features,feats]
return features