def get_me_idx_list():
"""
get all the me_idx, no matter runnable or not
:return: list[int]
"""
cache_path = "{}/tmp/me_idx_list.pkl".format(infty_cdb_folder)
test_folder_exist_for_file_path(cache_path)
if os.path.isfile(cache_path):
return load_serialization(cache_path)
me_xlsx_path = "{}/InftyCDB-1/resources/me.xlsx".format(SHARED_FOLDER)
wb = xlrd.open_workbook(me_xlsx_path)
sheet_names = wb.sheet_names()
me_idx_list = []
ws = wb.sheet_by_index(0)
for r_idx in range(ws.nrows):
row = ws.row(r_idx)
me_idx = int(row[20].value)
me_idx_list.append(me_idx)
me_idx_list = list(set(me_idx_list))
dump_serialization(me_idx_list, cache_path)
return me_idx_list
def load_chars_by_me_idx(me_idx):
"""
given an integer of me idx, return the related chars
:param me_idx:
:return: list[dict]
"""
cached_path = get_cached_chars_path_by_me_idx(me_idx)
if not os.path.isfile(cached_path):
batch_split_chars_by_me_idx()
elem_list = load_serialization(cached_path)
# NOTE, extra processing for the path of the radical line
# check the parent not in the list, if only one, change its parent as -1 and the relation as TOP
cid_list = [elem['cid'] for elem in elem_list]
cid_list_with_bad_pid = []
for elem in elem_list:
if elem['pid'] != -1 and elem['pid'] not in cid_list:
cid_list_with_bad_pid.append(elem['cid'])
if len(cid_list_with_bad_pid) == 0:
return elem_list
elif len(cid_list_with_bad_pid) == 1:
# update the one
the_idx = -1
for i, elem in enumerate(elem_list):
if elem['pid'] != -1 and elem['pid'] not in cid_list:
the_idx = i
elem_list[the_idx]['pid'] = -1
elem_list[the_idx]['relation'] = "TOP"
return elem_list
else:
raise Exception("more than two elements")
def get_one_ltchar():
from pdfxml.pdf_util.pdf_extract import process_pdf_internal
from pdfxml.path_util import ME_RESOURCE_FOLDER
cache_path = "{}/one_ltchar.pkl".format(ME_RESOURCE_FOLDER)
if os.path.isfile(cache_path):
return load_serialization(cache_path)
pdf_file_name = "{}/NIPS_2016_6202".format(ME_RESOURCE_FOLDER)
pdf_path = "{}.pdf".format(pdf_file_name)
char_list = process_pdf_internal(pdf_path, 0)
dump_serialization(char_list[0], cache_path)
return char_list[0]
def get_all_me_idx_list():
"""
This will be the list of me_idx to run experiment with
:return:
"""
cache_path = "{}/all_me_idx_list.pkl".format(infty_cdb_tmp_folder)
if os.path.isfile(cache_path):
return load_serialization(cache_path)
xlsx_path = "{}/InftyCDB-1/resources/me.xlsx".format(SHARED_FOLDER)
elem_list = load_me_elems_xlsx(xlsx_path)
me_idx_set = set()
for elem in elem_list:
me_idx_set.add(elem['me_idx'])
me_idx_list = list(me_idx_set)
dump_serialization(me_idx_list, cache_path)
return me_idx_list
def get_l1_dist(rel_id, fea_name):
"""
:param rel_id:
:param fea_name:
:return:
"""
cache_path = get_dist_cache_path([rel_id], fea_name)
if os.path.isfile(cache_path):
return load_serialization(cache_path)
samples = get_l1_samples(rel_id, fea_name)
dist = SampleDist(samples)
dump_serialization(dist, cache_path)
return dist
def get_cid2me_idx():
"""
char id to me idx
:return:
"""
cache = "{}/cid2me_idx.pkl".format(infty_cdb_tmp_folder)
test_folder_exist_for_file_path(cache)
if os.path.isfile(cache):
return load_serialization(cache)
cid2me_idx = {}
me_xlsx_path = "{}/InftyCDB-1/resources/me.xlsx".format(SHARED_FOLDER)
elem_list = load_me_elems_xlsx(me_xlsx_path)
for elem in elem_list:
cid2me_idx[elem['cid']] = elem['me_idx']
dump_serialization(cid2me_idx, cache)
return cid2me_idx
def get_nycd_dist(rel_id_list, refresh=False, sample_num=me_layout_config.default_sample_num):
"""
if length of 0 or 1, will not be
:param rel_id_list:
:param refresh:
:param sample_num:
:return:
"""
cache_path = get_dist_cache_path(
rel_id_list, 'nycd_dist', sample_num=sample_num)
if os.path.isfile(cache_path) and not refresh:
return load_serialization(cache_path)
print "try to calculate {}".format(cache_path)
dist = None
if len(rel_id_list) == 0:
raise Exception("Should not be here")
elif len(rel_id_list) == 1:
dist = get_l1_nycd_dist(rel_id_list[0])
else:
n = len(rel_id_list)
mid = int(n/2)
rel_list1 = rel_id_list[0:mid]
rel_list2 = rel_id_list[mid:]
nycd_dist_1 = get_nycd_dist(rel_list1)
nycd_dist_2 = get_nycd_dist(rel_list2)
hr_dist_1 = get_hr_dist(rel_list1)
nycd_samples_1 = nycd_dist_1.gen_random(sample_num)
nycd_samples_2 = nycd_dist_2.gen_random(sample_num)
hr_samples_1 = hr_dist_1.gen_random(sample_num)
# first do nycd2 * hr1
samples = []
for nycd2, hr1, nycd1 in itertools.product(
nycd_samples_2, hr_samples_1, nycd_samples_1):
tmp = nycd2 * hr1 + nycd1
samples.append(tmp)
dist = SampleDist(samples)
dump_serialization(dist, cache_path)
return dist
def get_hr_dist(rel_id_list, refresh=False, sample_num=me_layout_config.default_sample_num, debug=False):
"""
:param rel_id_list:
:param refresh:
:param sample_num:
:return:
"""
cache_path = get_dist_cache_path(
rel_id_list, 'hr_dist', sample_num=sample_num)
if os.path.isfile(cache_path) and not refresh:
if debug:
print cache_path
return load_serialization(cache_path)
print "try to calculate {}".format(cache_path)
dist = None
if len(rel_id_list) == 0:
raise Exception("TODO")
elif len(rel_id_list) == 1:
dist = get_l1_hr_dist(rel_id_list[0])
else:
n = len(rel_id_list)
mid = int(n/2)
# NOTE, there should be symbol overlapping of the two intervals
# but at the relation level, there is no overlapping
dist1 = get_hr_dist(rel_id_list[0:mid])
dist2 = get_hr_dist(rel_id_list[mid:])
print "done dist1 and dist2"
sample1 = dist1.gen_random(sample_num)
sample2 = dist2.gen_random(sample_num)
print "done gen samples"
vp_list = itertools.product(sample1, sample2)
samples = [vp[0] * vp[1] for vp in vp_list]
dist = SampleDist(samples)
print "done create new dist"
dump_serialization(dist, cache_path)
return dist
def get_fea_val_by_fea_name(fea_name):
"""
:param fea_name:
:type fea_name: basestring
:return: dict from relative position type to list of feature values
:rtype: dict[string_wx, list[float]]
"""
assert (fea_name in fea_name2fea_func)
fea_func = fea_name2fea_func[fea_name]
cache_path = "{}/fea_val/{}.pkl".format(infty_cdb_folder, fea_name)
if os.path.isfile(cache_path):
return load_serialization(cache_path)
rel_list = prepare_parent_child_relation_list_batch()
rel2fea_val_list = {
"HORIZONTAL": [],
"RSUP": [],
"RSUB": [],
"REV_RSUB": [],
"REV_RSUP": []
}
for pr in rel_list:
if not pr['relation'] in ["HORIZONTAL", "RSUP", "RSUB"]:
continue
# adjust bbox here
p_bbox = adjust_bbox_h_gt_name(pr['pinfo']['bbox'],
pr['pinfo']['name'])
c_bbox = adjust_bbox_h_gt_name(pr['cinfo']['bbox'],
pr['cinfo']['name'])
fea_val = fea_func(p_bbox, c_bbox)
rel2fea_val_list[pr['relation']].append(fea_val)
if pr['relation'] in ["RSUP", "RSUB"]:
# calculate the feature value here
rev_fea_val = fea_func(c_bbox, p_bbox)
rel2fea_val_list["REV_" + pr['relation']].append(rev_fea_val)
dump_serialization(rel2fea_val_list, cache_path)
return rel2fea_val_list
def prepare_parent_child_relation_list_one_me(me_idx):
"""
1. prepare data for different relation
2. Only keep the relation for the symbol with known x-y-z configration
3. [NOTE, not done?] normalize the bbox for chars
:param me_idx: me_idx as the identifier
:return: list of dict[parent, children, relation]
:rtype: list[dict]
"""
chars_path = get_cached_chars_path_by_me_idx(me_idx)
chars = load_serialization(chars_path)
cid2cinfo = {}
for c in chars:
cid2cinfo[c['cid']] = c
pair_list = []
for c in chars:
if c['pid'] == -1:
continue
if not cid2cinfo.has_key(c['pid']):
print 'no info for pid', c
continue
cname = c['name']
pname = cid2cinfo[c['pid']]['name']
#print cname, pname
c_adj_type = get_glyph_type(cname)
p_adj_type = get_glyph_type(pname)
adj_type_list = ['y', 'xy', 'yz', 'xyz', 'hxy', 'hxyz']
if c_adj_type not in adj_type_list:
continue
if p_adj_type not in adj_type_list:
continue
print cname, c_adj_type, pname, p_adj_type, c['relation']
pair_list.append({
'pinfo': cid2cinfo[c['pid']],
'cinfo': c,
'relation': c['relation']
})
return pair_list
def prepare_parent_child_relation_list_batch():
"""
merge all the triples
:return: :return: list of dict[parent, children, relation]
"""
cache_path = "{}/tmp/hor_sub_sup_alphanumeric.pkl".format(infty_cdb_folder)
if os.path.isfile(cache_path):
return load_serialization(cache_path)
res_list = []
# get all me_idx here
chars_folder = "{}/crop_chars".format(infty_cdb_folder)
for fname in os.listdir(chars_folder):
me_idx = int(fname[fname.rindex("_") + 1:-4])
pair_list = prepare_parent_child_relation_list_one_me(me_idx)
res_list.extend(pair_list)
print me_idx, len(pair_list)
dump_serialization(res_list, cache_path)
return res_list
def fname2shape():
"""
pre-calculate and store the shape of image file avoid overhead here
Need this because the vertical coordinate is reversed.
"""
cache_path = "{}/tmp/im_shape.pkl".format(infty_cdb_folder)
test_folder_exist_for_file_path(cache_path)
if os.path.isfile(cache_path):
return load_serialization(cache_path)
im2shape = {}
img_folder = "{}/InftyCDB-1/Images".format(SHARED_FOLDER)
for fname in os.listdir(img_folder):
if not fname.endswith("png"):
continue
fpath = "{}/InftyCDB-1/Images/{}".format(SHARED_FOLDER, fname)
im = imread(fpath)
im2shape[fname] = im.shape
dump_serialization(im2shape, cache_path)
return im2shape
def get_pid2cid_list():
"""
:return: map from parent char id to list of children
"""
cache_path = 'pid2cid_list.pkl'
if os.path.isfile(cache_path):
return load_serialization(cache_path)
cid2info = load_char_map()
pid2cid_list = {}
for cid, info in cid2info.items():
if info['pid'] == -1:
continue
if not pid2cid_list.has_key(info['pid']):
pid2cid_list[info['pid']] = []
pid2cid_list[info['pid']].append(cid)
dump_serialization(pid2cid_list, cache_path)
return pid2cid_list
def process_pdf_lines(fname, page_num='all', do_adjust=False):
"""
:param fname: file path to the PDF file
:param page_num: default to extract all
:return:
:rtype: list(list(LTChar))
"""
# TODO, cache the informatin here?
from pdfxml.path_util import get_tmp_path
tmp_pdf_path = get_tmp_path(fname)
pdf_lines_cache = "{}.pdf_line.{}.pkl".format(tmp_pdf_path, page_num)
if os.path.isfile(pdf_lines_cache):
return load_serialization(pdf_lines_cache)
line_list = []
char_list = []
def print_layout(l):
""" get all the chars
"""
for e in l:
if isinstance(e, LTTextLineHorizontal):
#print "try recursively text line"
print_layout(e)
line_list.append(copy.copy(char_list))
while len(char_list) > 0:
char_list.pop()
if isinstance(e, LTTextBoxHorizontal):
#print "try recursively text box"
print_layout(e)
if isinstance(e, LTChar) or isinstance(e, LTAnno):
char_list.append(e)
fp = open(fname, 'rb')
parser = PDFParser(fp)
document = PDFDocument(parser)
if not document.is_extractable:
raise PDFTextExtractionNotAllowed
rsrcmgr = PDFResourceManager()
device = PDFDevice(rsrcmgr)
interpreter = PDFPageInterpreter(rsrcmgr, device)
# Set parameters for analysis.
laparams = LAParams()
# Create a PDF page aggregator object.
device = PDFPageAggregator(rsrcmgr, laparams=laparams)
interpreter = PDFPageInterpreter(rsrcmgr, device)
for i, page in enumerate(PDFPage.create_pages(document)):
process_mark = (page_num == 'all' or page_num == i)
if process_mark:
interpreter.process_page(page)
layout = device.get_result()
print_layout(layout)
if page_num == i:
break
if do_adjust:
for line in line_list:
adjust_basedon_glyph_ratio(line, fname, page_num)
# adjust based on crop bbox
crop_bbox = get_pdf_page_bbox_abandon(fname, page_num)
for line in line_list:
for char in line:
if isinstance(char, LTChar):
adjust_element_bbox(char, crop_bbox)
dump_serialization(line_list, pdf_lines_cache)
return line_list
def pdf_extract_lines(pdf_path, pid=0, force_single=False):
"""
each line is a list of LTChar
:param pdf_path:
:param pid:
:return:
"""
tmp_pdf_path = get_tmp_path(pdf_path)
cache_path = "{}.pdfbox_merge_line.{}.pkl".format(tmp_pdf_path, pid)
if os.path.isfile(cache_path):
return load_serialization(cache_path)
char_list_list = pdf_extract_lines_raw(pdf_path, pid)
# TODO, do another round of line merging
# still use our column line detection model to find the region.
fontname2space = pdf_extract_fontname2space(pdf_path, pid)
word_info_list = pdf_extract_words(pdf_path, pid)
res_char_list_list = []
if not force_single and is_double_column(pdf_path, pid):
# split the current list into three parts
# detect the center split, create two column
# outside of the double column,
# within the double column
page_size = get_pdf_page_size(pdf_path, pid)
page_width = page_size['width']
out_char_list_list = []
left_char_list_list = []
right_char_list_list = []
from pdfxml.pdf_util.layout_util import get_char_list_bbox
for char_list in char_list_list:
bbox = get_char_list_bbox(char_list)
if bbox.left() < bbox.right() < page_width / 2:
left_char_list_list.append(char_list)
elif bbox.right() > bbox.left() > page_width / 2:
right_char_list_list.append(char_list)
else:
out_char_list_list.append(char_list)
# before mering do the word_info_filter
word_info_list = word_info_filter(char_list_list, word_info_list)
new_out_char_list_list = merging_lines(out_char_list_list,
fontname2space, word_info_list,
pdf_path, pid)
new_left_char_list_list = merging_lines(left_char_list_list,
fontname2space, word_info_list,
pdf_path, pid)
new_right_char_list_list = merging_lines(right_char_list_list,
fontname2space,
word_info_list, pdf_path, pid)
# not in the vertical range of the double dolumn
# center on the left part,
# center on the right part,
char_list_list = []
char_list_list.extend(new_out_char_list_list)
char_list_list.extend(new_left_char_list_list)
char_list_list.extend(new_right_char_list_list)
res_char_list_list = char_list_list
else:
# before mering do the word_info_filter
word_info_list = word_info_filter(char_list_list, word_info_list)
# single column, then just go on merging the lines
new_char_list_list = merging_lines(char_list_list, fontname2space,
word_info_list, pdf_path, pid)
res_char_list_list = new_char_list_list
dump_serialization(res_char_list_list, cache_path)
return res_char_list_list
def get_char2extend_for_one(me_idx, debug=False):
"""
based on the horizontal grouping from construct_hierarchy
try to make assessment on the correction of the bbox
:param me_idx: the idx of ME
:type me_idx: int
:param debug:
:return: a dict, code 2 upper_list and code 2 lower_list
'code2upper_ratio': use height to adjust upper
'code2upper_ratio_hor': use the width to adjust upper, because the error rate for the flat sign such as minus is hard to estimate.
'code2lower_ratio': use height to adjust lower
'code2lower_ratio_hor': use the width to adjust lower
"""
code2name = get_code2name()
cached_path = "{}/char2extend_ratio/{}.pkl".format(infty_cdb_folder,
me_idx)
if not debug and os.path.isfile(cached_path):
# if not debug and the file exist
return load_serialization(cached_path)
code2upper_ratio, code2upper_ratio_hor, \
code2lower_ratio, code2lower_ratio_hor = {}, {}, {}, {}
res = {
'code2upper_ratio': code2upper_ratio,
'code2upper_ratio_hor': code2upper_ratio_hor,
'code2lower_ratio': code2lower_ratio,
'code2lower_ratio_hor': code2lower_ratio_hor
}
def update_dict(d, k, v):
if not d.has_key(k):
d[k] = []
d[k].append(v)
# change to inftyCDBME
try:
struct_info = construct_hierarchy_by_me_idx(me_idx)
except Exception as e:
print "failed for me_idx {} {}".format(me_idx, str(e))
return res
cid2info = struct_info.cid2chars
for group in struct_info.hor_groups:
# each group is a list of cid
if upper_exist(group, cid2info):
upper_line = get_upper_line(group, cid2info)
if debug:
print 'ascender_line: {}'.format(upper_line)
print_line(group, cid2info)
for cid in group:
r = get_upper_ratio(cid2info[cid], upper_line)
r_hor = get_upper_ratio_hor(cid2info[cid], upper_line)
update_dict(code2upper_ratio, cid2info[cid]['code'], r)
update_dict(code2upper_ratio_hor, cid2info[cid]['code'], r_hor)
if debug:
print cid, code2name[cid2info[cid]['code']], \
"code2upper_ratio", r, \
"code2upper_ratio_hor", r_hor
if lower_exist(group, cid2info):
lower_line = get_lower_line(group, cid2info)
if debug:
print 'descender_line: {}'.format(lower_line)
print_line(group, cid2info)
for cid in group:
r = get_lower_ratio(cid2info[cid], lower_line)
r_hor = get_lower_ratio_hor(cid2info[cid], lower_line)
update_dict(code2lower_ratio, cid2info[cid]['code'], r)
update_dict(code2lower_ratio_hor, cid2info[cid]['code'], r_hor)
if debug:
print cid, code2name[cid2info[cid]['code']], \
"code2lower_ratio", r, \
"code2lower_ratio_hor", r_hor
dump_serialization(res, cached_path)
return res