def my_filter(imageIn):
MAX_CCS = 4000
count = 0
image = imageIn
#imageIn.remove_border()
ccs = image.cc_analysis()
print "filter started on",len(ccs) ,"elements..."
if len(ccs) < 1:
raise ImageSegmentationError("there are no ccs")
if len(ccs) > MAX_CCS:
raise ImageSegmentationError("there are more than " + str(MAX_CCS) + " ccs.")
median_black_area = median([cc.black_area()[0] for cc in ccs])
#filter long vertical runs left over from margins
median_height = median([cc.nrows for cc in ccs])
for cc in ccs:
if((cc.nrows / cc.ncols > 6) and (cc.nrows > 1.5 * median_height) ):
cc.fill_white()
del cc
count = count + 1
for cc in ccs:
if(cc.black_area()[0] > (median_black_area * 10)):
cc.fill_white()
del cc
count = count + 1
for cc in ccs:
if(cc.black_area()[0] < (median_black_area / 10)):
cc.fill_white()
del cc
count = count + 1
print "filter done.",len(ccs)-count,"elements left."
def chars_make_words(lines_glyphs,threshold=None):
"""Groups the given glyphs to words based upon the horizontal distance
between adjacent glyphs.
Signature:
``chars_make_words (glyphs, threshold=None)``
with
*glyphs*:
A list of ``Cc`` data types, each of which representing a character.
All glyphs must stem from the same single line of text.
*threshold*:
Horizontal white space greater than *threshold* will be considered
a word separating gap. When ``None``, the threshold value is
calculated automatically as 2.5 times teh median white space
between adjacent glyphs.
The result is a nested list of glyphs with each sublist representing
a word. This is the same data structure as used in `Textline.words`_
.. _`Textline.words`: gamera.toolkits.ocr.classes.Textline.html
"""
glyphs = lines_glyphs[:]
wordlist = []
if(threshold == None):
spacelist = []
total_space = 0
for i in range(len(glyphs) - 1):
spacelist.append(glyphs[i + 1].ul_x - glyphs[i].lr_x)
if(len(spacelist) > 0):
threshold = median(spacelist)
threshold = threshold * 2.5
else:
threshold = 0
word = []
for i in range(len(glyphs)):
if i > 0:
if((glyphs[i].ul_x - glyphs[i - 1].lr_x) > threshold):
wordlist.append(word)
word = []
word.append(glyphs[i])
if(len(word) > 0):
wordlist.append(word)
return wordlist
def my_filter(imageIn):
count = 0
image = imageIn
ccs = image.cc_analysis()
print "filter started on", len(ccs), "elements..."
median_black_area = median([cc.black_area()[0] for cc in ccs])
#also check for height?
for cc in ccs:
if (cc.black_area()[0] > (median_black_area * 10)):
cc.fill_white()
del cc
count = count + 1
for cc in ccs:
if (cc.black_area()[0] < (median_black_area / 10)):
cc.fill_white()
del cc
count = count + 1
print "filter done.", len(ccs) - count, "elements left."
def my_filter(imageIn): count = 0 image = imageIn ccs = image.cc_analysis() print "filter started on",len(ccs) ,"elements..." median_black_area = median([cc.black_area()[0] for cc in ccs]) #also check for height? for cc in ccs: if(cc.black_area()[0] > (median_black_area * 10)): cc.fill_white() del cc count = count + 1 for cc in ccs: if(cc.black_area()[0] < (median_black_area / 10)): cc.fill_white() del cc count = count + 1 print "filter done.",len(ccs)-count,"elements left."
def performGreekOCR(options):
import mahotas as mh
import numpy as np
from pylab import imshow, gray, show
#from os import path
from gamera.plugins import numpy_io
# features = ["aspect_ratio", "volume64regions", "moments", "nholes_extended"]
#I think these are size-invariant
# features = ["aspect_ratio","moments","nholes","nholes_extended","skeleton_features","top_bottom","volume","volume16regions","volume64regions","zernike_moments"]
MAX_CCS = 6500
features = ["aspect_ratio","moments","ncols_feature","nholes","nholes_extended","nrows_feature","skeleton_features","top_bottom","volume","volume16regions","volume64regions","zernike_moments"]
image_files = []
g = GreekOCR(splits=options["split"],feats=features)
g.mode = options["mode"] + "body"
g.autogroup = options["autogroup"]
g.debug = options["debug"]
g.load_trainingdata(options['trainingdata'])
g_appcrit = GreekOCR(splits=options["split"], feats=features)
g_appcrit.mode = options["mode"] + "appcrit"
g_appcrit.autogroup = options["autogroup"]
g_appcrit.debug = options["debug"]
g_appcrit.load_trainingdata(options['trainingdata'])
if options["hocrfile"]:
g.hocr = (options["hocrfile"])
if options["settingsfile"]:
g.load_settings(options["settingsfile"])
g_appcrit.load_settings(options["settingsfile"])
if options["otsu"]:
otsu_factors_string = options["otsu"].split(',')
otsu_factors = [float(x) for x in otsu_factors_string]
else:
otsu_factors = [0]
if options["directory"]:
image_files = os.listdir(options["directory"])
image_files = [os.path.join(options["directory"],x) for x in image_files]
test = re.compile(".png$",re.IGNORECASE)
image_files = filter(test.search, image_files)
image_files.sort()
elif options["imagefile"]:
image_files = options["imagefile"]
image_file_count = 1;
image_path = os.path.abspath(image_files[0])
image_split_path = os.path.split(image_path)
book_code = os.path.split(image_split_path[0])[1]
book_id = 0
if options.has_key("sql") and options["sql"]:
book_id = sql_make_book_and_return_id(book_code)
# if options.has_key("hocrout") and options["hocrout"]:
# hocr_tree = hocr_make_tree_and_return(book_code)
for image_file in image_files:
image_path = os.path.abspath(image_file)
image_split_path = os.path.split(image_path)
book_code = os.path.split(image_split_path[0])[1]#directory name
image_file_name = image_split_path[1]
imageBase, imageEx = os.path.splitext(image_file_name)
threshold_info = ""
print "Now working with image: " + image_file_name
internal_image_file_path = os.path.join(book_code, image_file_name)
if imageEx == ".jp2":
try:
jp2Image = mh.imread(image_file, as_grey=True)
jp2Image = jp2Image.astype(np.uint8)
imageIn = numpy_io.from_numpy(jp2Image)
except:
print "Unexpected error:", sys.exc_info()[0]
raise
else:
try:
imageIn = load_image(image_file)
except:
continue
imageType = imageIn.data.pixel_type
if imageType != ONEBIT:
if imageType != GREYSCALE:
imageIn = imageIn.to_greyscale()
otsu_thresh = imageIn.otsu_find_threshold()
for otsu_factor in otsu_factors:
if options.has_key("hocrout") and options["hocrout"]:
hocr_tree = hocr_make_tree_and_return(book_code)
if imageIn.data.pixel_type == ONEBIT:
threshold_info = "thresh_128"
otsu_thresh = 1.0
image = imageIn
if options["debug"]:
print "image is ONEBIT; doing no threshold optimization."
else:
current_thresh = otsu_thresh * otsu_factor
if current_thresh > 253.0:
current_thresh = 253.0
current_thresh = int(current_thresh)
threshold_info = "thresh_" + str(int(current_thresh))# + "=" + str(otsu_factor)
image = imageIn.threshold(current_thresh)
print "Otsu factor: ", otsu_factor, " threshold: ", current_thresh
if options["hocrfile"]:
hocr_to_use = string.replace(options["hocrfile"],"%s",imageBase)
g.hocr = hocr_to_use
if options["debug"]:
print "using '" + hocr_to_use + "' as hocr file"
if options.has_key("filter") and options["filter"] == True:
count = 0
ccs = image.cc_analysis()
if options.has_key("debug"):
if options["debug"] == True:
print "filter started on",len(ccs) ,"elements..."
#filter long vertical runs left over from margins
## #Agressive run filtering
## median_height = median([cc.nrows for cc in ccs])
## for cc in ccs:
## #TODO: add another condition that keeps these at edges of page
## if((cc.nrows / cc.ncols > 6) and (cc.nrows > 1.5 * median_height) ):
## cc.fill_white()
## del cc
## count = count + 1
median_black_area = median([cc.black_area()[0] for cc in ccs])
for cc in ccs:
if(cc.black_area()[0] > (median_black_area * 10)):
cc.fill_white()
del cc
count = count + 1
for cc in ccs:
if(cc.black_area()[0] < (median_black_area / 10)):
cc.fill_white()
del cc
count = count + 1
if options.has_key("debug") and options["debug"] == True:
print "filter done.",len(ccs)-count,"elements left."
if (len(ccs) < 5) or (len(ccs) > MAX_CCS):
print "Error: there are " + str(len(ccs)) + " ccs. Max is " + str( MAX_CCS) + " Omitting this image."
#raise ImageSegmentationError("Error: there are " + str(len(ccs)) + " ccs. Max is " + str( MAX_CCS) + " Omitting this image.")
else:
if options.has_key("deskew") and options["deskew"] == True:
#from gamera.toolkits.otr.otr_staff import *
if options.has_key("debug") and options["debug"] == True:
print "\ntry to skew correct..."
rotation = image.rotation_angle_projections(-10,10)[0]
img = image.rotate(rotation,0)
if options.has_key("debug") and options["debug"] == True:
print "rotated with",rotation,"angle"
if options.has_key("mode") and options["mode"] == "teubner":
(body_image, app_crit_image) = splitAppCritTeubner(image)
output = g.process_image(body_image)
if app_crit_image:
print "there is an app. crit image"
appcrit_output = g_appcrit.process_image(app_crit_image)
else:
print "there is no app. crit image"
appcrit_output = ""
output = output + appcrit_output
else:
output = g.process_image(image)
output_file_name_base = options["unicodeoutfile"] + imageBase + "_" +imageEx[1:] + "_" + threshold_info
if options.has_key("debug") and options["debug"] == True:
g.save_debug_images(output_file_name_base)
if options.has_key("mode") and options["mode"] == "teubner" and app_crit_image:
#TODO: make more general
g_appcrit.save_debug_images(output_file_name_base + "_appcrit")
if options.has_key("hocrout") and options["hocrout"]:
#if we turned this on, we would make a separate div for each page of input
#hocr_tree = hocr_make_page_and_return_div(internal_image_file_path,image_file_count,book_id,hocr_tree)
g.store_hocr(internal_image_file_path,hocr_tree)
if options.has_key("mode") and options["mode"] == "teubner" and app_crit_image:
g_appcrit.store_hocr(internal_image_file_path,hocr_tree)
if options.has_key("sql") and options["sql"]:
page_id = sql_make_page_and_return_id(internal_image_file_path,image_file_count,book_id)
g.store_sql(image_path,page_id)
if options.has_key("unicodeoutfile"):
if options.has_key("hocrout") and options["hocrout"]:
g.save_text_hocr(hocr_tree, output_file_name_base + ".html")
else:
g.save_text_unicode( output_file_name_base + ".txt")
if options.has_key("mode") and options["mode"] == "teubner":
#TODO: make the above more general
g_appcrit.save_text_unicode( output_file_name_base + "_appcrit.txt")
elif options.has_key("teubneroutfile"):
g.save_text_teubner(options["teubneroutfile"])
else:
print output
image_file_count += 1
def __call__(self, Ex=-1, Ey=-1, iterations=2):
# bbox with contained cc indices
class Bbox:
def __init__(self, allccs, indices):
self.ccs = allccs
self.indices = indices
if len(indices) == 1:
self.rect = Rect(allccs[indices[0]])
else:
self.rect = allccs[indices[0]].union_images(
[allccs[i] for i in indices])
def extend(self, Ex, Ey, img):
ul_y = max(0, self.rect.ul_y - Ey)
ul_x = max(0, self.rect.ul_x - Ex)
lr_y = min(img.lr_y, self.rect.lr_y + Ey)
lr_x = min(img.lr_x, self.rect.lr_x + Ex)
nrows = lr_y - ul_y + 1
ncols = lr_x - ul_x + 1
self.rect = Rect(Point(ul_x, ul_y), Dim(ncols, nrows))
def merge(self, other):
self.indices += other.indices
self.rect.union(other.rect)
# does one merging step
def merge_boxes(bboxes):
from gamera import graph
bboxes.sort(lambda b1, b2: b1.rect.ul_y - b2.rect.ul_y)
g = graph.Graph(graph.UNDIRECTED)
# build graph where edge means overlap of two boxes
for i in range(len(bboxes)):
g.add_node(i)
for i in range(len(bboxes)):
for j in range(i + 1, len(bboxes)):
if bboxes[j].rect.ul_y > bboxes[i].rect.lr_y:
break
if bboxes[i].rect.intersects(bboxes[j].rect):
if not g.has_edge(i, j):
g.add_edge(i, j)
new_bboxes = []
for sg in g.get_subgraph_roots():
seg = [n() for n in g.BFS(sg)]
bbox = bboxes[seg[0]]
for i in range(1, len(seg)):
bbox.merge(bboxes[seg[i]])
new_bboxes.append(bbox)
return new_bboxes
# the actual plugin
from gamera.core import Dim, Rect, Point, Cc
from gamera.plugins.listutilities import median
page = self.image_copy()
ccs = page.cc_analysis()
# compute average CC size
if Ex == -1:
Ex = 2 * median([c.ncols for c in ccs])
if Ey == -1:
Ey = median([c.nrows for c in ccs])
# create merged segments
bboxes = [Bbox(ccs, [i]) for i in range(len(ccs))]
for bb in bboxes:
bb.extend(Ex, Ey, page)
for i in range(iterations):
oldlen = len(bboxes)
bboxes = merge_boxes(bboxes)
if oldlen == len(bboxes):
break
seg_ccs = []
for i, bbox in enumerate(bboxes):
label = i + 1
ccs_of_segment = [ccs[j] for j in bbox.indices]
for cc in ccs_of_segment:
self.highlight(cc, label)
seg_ccs.append(
Cc(self, label, ccs_of_segment[0].union_rects(ccs_of_segment)))
return seg_ccs
def __call__(self, Ex=-1, Ey=-1, iterations=2):
# bbox with contained cc indices
class Bbox:
def __init__(self, allccs, indices):
self.ccs = allccs
self.indices = indices
if len(indices) == 1:
self.rect = Rect(allccs[indices[0]])
else:
self.rect = allccs[indices[0]].union_images([allccs[i] for i in indices])
def extend(self, Ex, Ey, img):
ul_y = max(0, self.rect.ul_y - Ey)
ul_x = max(0, self.rect.ul_x - Ex)
lr_y = min(img.lr_y, self.rect.lr_y + Ey)
lr_x = min(img.lr_x, self.rect.lr_x + Ex)
nrows = lr_y - ul_y + 1
ncols = lr_x - ul_x + 1
self.rect = Rect(Point(ul_x, ul_y), Dim(ncols, nrows))
def merge(self, other):
self.indices += other.indices
self.rect.union(other.rect)
# does one merging step
def merge_boxes(bboxes):
from gamera import graph
bboxes.sort(lambda b1, b2: b1.rect.ul_y-b2.rect.ul_y)
g = graph.Graph(graph.UNDIRECTED)
# build graph where edge means overlap of two boxes
for i in range(len(bboxes)):
g.add_node(i)
for i in range(len(bboxes)):
for j in range(i+1, len(bboxes)):
if bboxes[j].rect.ul_y > bboxes[i].rect.lr_y:
break
if bboxes[i].rect.intersects(bboxes[j].rect):
if not g.has_edge(i,j):
g.add_edge(i,j)
new_bboxes = []
for sg in g.get_subgraph_roots():
seg = [n() for n in g.BFS(sg)]
bbox = bboxes[seg[0]]
for i in range(1, len(seg)):
bbox.merge(bboxes[seg[i]])
new_bboxes.append(bbox)
return new_bboxes
# the actual plugin
from gamera.core import Dim, Rect, Point, Cc
from gamera.plugins.listutilities import median
page = self.image_copy()
ccs = page.cc_analysis()
# compute average CC size
if Ex == -1:
Ex = 2*median([c.ncols for c in ccs])
if Ey == -1:
Ey = median([c.nrows for c in ccs])
# create merged segments
bboxes = [Bbox(ccs, [i]) for i in range(len(ccs))]
for bb in bboxes:
bb.extend(Ex, Ey, page)
for i in range(iterations):
oldlen = len(bboxes)
bboxes = merge_boxes(bboxes)
if oldlen == len(bboxes):
break
seg_ccs = []
for i,bbox in enumerate(bboxes):
label = i+1
ccs_of_segment = [ccs[j] for j in bbox.indices]
for cc in ccs_of_segment:
self.highlight(cc, label)
seg_ccs.append(Cc(self, label, ccs_of_segment[0].union_rects(ccs_of_segment)))
return seg_ccs