def extract_candidate_groups(segmentation,
min_aspect=0.1,
max_aspect=1.8,
maxgap=2,
maxrange=3):
"""Select a basic collection of candidate components. Aspect ratios <1 are tall and skinny,
>1 are wide. `maxgap` is the maximum gap between bounding boxes. `maxrange` gives the
maximum number of compponents to be combined."""
assert morph.ordered_by_xcenter(
segmentation), "call morph.sort_by_xcenter first"
boxes = [None] + morph.find_objects(segmentation)
n = len(boxes)
result = []
for i in range(1, n):
for r in range(1, maxrange + 1):
if i + r > n: continue
if r > 1 and max_boxgap(boxes[i:i + r]) > maxgap:
continue
box = box_union(boxes[i:i + r])
a = sl.aspect(box)
if r > 1 and 1.0 / a > max_aspect: continue
if 1.0 / a < min_aspect: continue
assert sum(segmentation[box]) > 0
j = i + r - 1
seg = segmentation[box] * (segmentation[box] >=
i) * (segmentation[box] <= j)
result.append(Segment(first=i, last=j, bbox=box))
return result
def extract_candidate_groups(segmentation, min_aspect=0.1, max_aspect=1.8, maxgap=2, maxrange=3):
"""Select a basic collection of candidate components. Aspect ratios <1 are tall and skinny,
>1 are wide. `maxgap` is the maximum gap between bounding boxes. `maxrange` gives the
maximum number of compponents to be combined."""
assert morph.ordered_by_xcenter(segmentation), "call morph.sort_by_xcenter first"
boxes = [None] + morph.find_objects(segmentation)
n = len(boxes)
result = []
for i in range(1, n):
for r in range(1, maxrange + 1):
if i + r > n:
continue
if r > 1 and max_boxgap(boxes[i : i + r]) > maxgap:
continue
box = box_union(boxes[i : i + r])
a = sl.aspect(box)
if r > 1 and 1.0 / a > max_aspect:
continue
if 1.0 / a < min_aspect:
continue
assert sum(segmentation[box]) > 0
j = i + r - 1
seg = segmentation[box] * (segmentation[box] >= i) * (segmentation[box] <= j)
result.append(Segment(first=i, last=j, bbox=box))
return result
def setImageMasked(self, image, mask=None, lo=None, hi=None):
"""Set the image to be iterated over. This should be an RGB image,
ndim==3, dtype=='B'. This picks a subset of the segmentation to iterate
over, using a mask and lo and hi values.
"""
# print("$$ setImageMasked: image=%s mask=%s lo=%s hi=%s" % (
# desc(image), hexs(mask), hexs(lo), hexs(hi)))
assert image.dtype == dtype('B') or image.dtype == dtype(
'i'), "image must be type B or i"
if image.ndim == 3:
image = rgb2int(image)
imageDescribe("image", image)
assert image.ndim == 2, "wrong number of dimensions"
self.image = image
labels = image
if lo is not None:
labels[labels < lo] = 0
if hi is not None:
labels[labels > hi] = 0
if mask is not None:
labels = bitwise_and(labels, mask)
imageDescribe("labels", labels)
labels, correspondence = morph.renumber_labels_ordered(
labels, correspondence=1)
self.labels = labels
self.correspondence = correspondence
self.objects = [None] + morph.find_objects(labels)
# print("$$ setImageMasked: objects=%d" % len(self.objects))
sizes = [o for o in self.objects]
sizes.sort(key=lambda u: -sliceSize(u))
for i, o in enumerate(sizes[:20]):
print("%5d: %s %s %d" % (i, o, sliceDims(o), sliceSize(o)))
def bbox(image):
"""Compute the bounding box for the pixels in the image."""
assert len(image.shape)==2,"wrong shape: "+str(image.shape)
image = array(image!=0,'uint32')
cs = morph.find_objects(image)
if len(cs)<1: return None
c = cs[0]
return (c[0].start,c[1].start,c[0].stop,c[1].stop)
def bbox(image):
"""Compute the bounding box for the pixels in the image."""
assert len(image.shape) == 2, "wrong shape: " + str(image.shape)
image = array(image != 0, 'uint32')
cs = morph.find_objects(image)
if len(cs) < 1: return None
c = cs[0]
return (c[0].start, c[1].start, c[0].stop, c[1].stop)
def extract_csegs(segmentation,aligned=None):
"""Given a segmentation, extracts a list of segment objects."""
if aligned is None: aligned = []
def get(c): return aligned[c] if c<len(aligned) else ""
boxes = morph.find_objects(segmentation)
# n = len(boxes)
result = [Segment(first=i+1,last=i+1,bbox=box,out=[(get(i),0.0)],
img=1*(segmentation[:,box[1]]==i+1)) \
for i,box in enumerate(boxes) if box is not None]
return result
def setImageMasked(self,image,mask=None,lo=None,hi=None):
"""Set the image to be iterated over. This should be an RGB image,
ndim==3, dtype=='B'. This picks a subset of the segmentation to iterate
over, using a mask and lo and hi values.."""
assert image.dtype==dtype('B') or image.dtype==dtype('i'),"image must be type B or i"
if image.ndim==3: image = rgb2int(image)
assert image.ndim==2,"wrong number of dimensions"
self.image = image
labels = image
if lo is not None: labels[labels<lo] = 0
if hi is not None: labels[labels>hi] = 0
if mask is not None: labels = bitwise_and(labels,mask)
labels,correspondence = morph.renumber_labels_ordered(labels,correspondence=1)
self.labels = labels
self.correspondence = correspondence
self.objects = [None]+morph.find_objects(labels)
def compute_lines(segmentation, scale):
"""Given a line segmentation map, computes a list
of tuples consisting of 2D slices and masked images."""
lobjects = morph.find_objects(segmentation)
lines = []
for i, o in enumerate(lobjects):
if o is None: continue
if sl.dim1(o) < 2 * scale or sl.dim0(o) < scale: continue
mask = (segmentation[o] == i + 1)
if amax(mask) == 0: continue
result = record()
result.label = i + 1
result.bounds = o
result.mask = mask
lines.append(result)
return lines
def compute_lines(segmentation,scale):
"""Given a line segmentation map, computes a list
of tuples consisting of 2D slices and masked images."""
lobjects = morph.find_objects(segmentation)
lines = []
for i,o in enumerate(lobjects):
if o is None: continue
if sl.dim1(o)<2*scale or sl.dim0(o)<scale: continue
mask = (segmentation[o]==i+1)
if amax(mask)==0: continue
result = record()
result.label = i+1
result.bounds = o
result.mask = mask
lines.append(result)
return lines
def seg_boxes(seg,math=0):
"""Given a color segmentation, return a list of bounding boxes.
Bounding boxes are returned as tuples (y0,y1,x0,x1). With
math=0, raster coordinates are used, with math=1, Postscript
coordinates are used (however, the order of the values in the
tuple doesn't change)."""
seg = array(seg,'uint32')
slices = morph.find_objects(seg)
h = seg.shape[0]
result = []
for i in range(len(slices)):
if slices[i] is None: continue
(ys,xs) = slices[i]
if math:
result += [(h-ys.stop-1,h-ys.start-1,xs.start,xs.stop)]
else:
result += [(ys.start,ys.stop,xs.start,xs.stop)]
return result
def seg_boxes(seg, math=0):
"""Given a color segmentation, return a list of bounding boxes.
Bounding boxes are returned as tuples (y0,y1,x0,x1). With
math=0, raster coordinates are used, with math=1, Postscript
coordinates are used (however, the order of the values in the
tuple doesn't change)."""
seg = array(seg, 'uint32')
slices = morph.find_objects(seg)
h = seg.shape[0]
result = []
for i in range(len(slices)):
if slices[i] is None: continue
(ys, xs) = slices[i]
if math:
result += [(h - ys.stop - 1, h - ys.start - 1, xs.start, xs.stop)]
else:
result += [(ys.start, ys.stop, xs.start, xs.stop)]
return result
def extract_chars(segmentation,h=32,w=32,f=0.5,minscale=0.5):
"""Extract all the characters from the segmentation and yields them
as an interator. Also yields a forward and a backwards transformation."""
bin = (segmentation>0)
if amax(bin)==0: raise ocrolib.RecognitionError("empty segmentation")
ls,ly,lx = vertical_stddev(bin)
boxes = morph.find_objects(segmentation)
for i,b in enumerate(boxes):
sub = (segmentation==i+1)
if amax(sub)==amin(sub): continue
cs,cy,cx = vertical_stddev(sub)
# limit the character sigma to be at least minscale times the
# line sigma (this causes dots etc. not to blow up ridiculously large)
scale = f*h/(4*max(cs,minscale*ls))
m = diag([1.0/scale,1.0/scale])
offset = array([cy,cx])-dot(m,array([h/2,w/2]))
def transform(image,m=m,offset=offset):
return interpolation.affine_transform(1.0*image,m,offset=offset,order=1,output_shape=(h,w))
def itransform_add(result,image,m=m,cx=cx,cy=cy):
im = inv(m)
ioffset = array([h/2,w/2])-dot(im,array([cy,cx]))
result += interpolation.affine_transform(1.0*image,im,offset=ioffset,order=1,output_shape=segmentation.shape)
cimage = transform(sub)
yield cimage,transform,itransform_add
def extract_chars(segmentation, h=32, w=32, f=0.5, minscale=0.5):
"""Extract all the characters from the segmentation and yields them
as an interator. Also yields a forward and a backwards transformation."""
bin = (segmentation > 0)
if amax(bin) == 0: raise ocrolib.RecognitionError("empty segmentation")
ls, ly, lx = vertical_stddev(bin)
boxes = morph.find_objects(segmentation)
for i, b in enumerate(boxes):
sub = (segmentation == i + 1)
if amax(sub) == amin(sub): continue
cs, cy, cx = vertical_stddev(sub)
# limit the character sigma to be at least minscale times the
# line sigma (this causes dots etc. not to blow up ridiculously large)
scale = f * h / (4 * max(cs, minscale * ls))
m = diag([1.0 / scale, 1.0 / scale])
offset = array([cy, cx]) - dot(m, array([h / 2, w / 2]))
def transform(image, m=m, offset=offset):
return interpolation.affine_transform(1.0 * image,
m,
offset=offset,
order=1,
output_shape=(h, w))
def itransform_add(result, image, m=m, cx=cx, cy=cy):
im = inv(m)
ioffset = array([h / 2, w / 2]) - dot(im, array([cy, cx]))
result += interpolation.affine_transform(
1.0 * image,
im,
offset=ioffset,
order=1,
output_shape=segmentation.shape)
cimage = transform(sub)
yield cimage, transform, itransform_add
def binary_objects(binary):
labels, n = morph.label(binary)
objects = morph.find_objects(labels)
return objects
def binary_objects(binary):
labels,n = morph.label(binary)
objects = morph.find_objects(labels)
return objects
