From 0a8ac5d6cd0bf37fa15253475b4e8bee4b396e2d Mon Sep 17 00:00:00 2001
From: Luis Salamanca <luis.salamanca@sdsc.ethz.ch>
Date: Tue, 25 Feb 2020 11:12:21 +0100
Subject: [PATCH] Changes for improving, hopefully, the xml correction
---
src/python/def_classes.py | 11 +++++---
src/python/preproc_docs.py | 56 ++++++++++++++++++++++++++------------
2 files changed, 45 insertions(+), 22 deletions(-)
diff --git a/src/python/def_classes.py b/src/python/def_classes.py
index 2439079d..210e2ab5 100644
--- a/src/python/def_classes.py
+++ b/src/python/def_classes.py
@@ -40,7 +40,7 @@ class Document:
limit_year_f1 = 1950
limit_id_f3 = 110001491
- flag_end_run = 1
+ flag_end_run = 0 # 0, we run the command python pdf2txt.py to convert to xml, 1, we juts run it like pdf2txt.py
name_inpdf = '00_rawpdfs'
name_inmeta = '01_rawmeta'
name_inmeta_corr = '03_correctedmeta'
@@ -61,6 +61,8 @@ class Document:
self.name_wo_ext = os.path.splitext(self.name_file)[0]
self.folder_database = folder_database
self.flag_tar = flag_tar
+ self.flag_forcecomp = False # Only update the xml corrected if it is too old
+ self.flag_order = 1 # To indicate that we will reevaluate the order of textlines to reorder them
self._meta_corr_ext()
if flag_type == 1:
@@ -285,7 +287,7 @@ class Document:
#####
# Central vertical line and horizontal lines, through Hough transform
coord_vert_def, coord_horz, flag_2col, sections_page = preproc_docs.find_mainHorandCentral_Hough(np.copy(imarray), coord_textline, bbox_page,
- flag_2col, flag_central)
+ self.flag_type, flag_central)
print('Page %d: flag_2col %d' % (ind_page, flag_2col))
#####
@@ -307,7 +309,7 @@ class Document:
#####
# Order the textlines, taken all them together, in order to later merge
# in a single textbox textlines that so far form different textboxes
- flag_order = 1
+ flag_order = self.flag_order
if self.flag_type == 2:
flag_order = 0
set_of_blocks, centrall_ord = preproc_docs.order_textl(rescale_factor, coord_vert_def, coord_horz,
@@ -1244,7 +1246,8 @@ class Document:
def xml(self):
if not hasattr(self, "_XML_main_corr") or self._XML_main_corr is None:
# the instance does not have a pointer to the xml
- self._XML_main_corr = self.load_or_compute_xml()
+ force_compute = self.flag_forcecomp
+ self._XML_main_corr = self.load_or_compute_xml(force_compute=force_compute)
return self._XML_main_corr
diff --git a/src/python/preproc_docs.py b/src/python/preproc_docs.py
index da1904e5..f4cb0c13 100644
--- a/src/python/preproc_docs.py
+++ b/src/python/preproc_docs.py
@@ -71,7 +71,7 @@ def find_mainHorandCentral_Hough(img, coord, dim_bbox_page,
width_resc = WIDTH_CHAR * rescale_factor[0,1]
height_resc = HEIGHT_CHAR * rescale_factor[0,1]
- gap_central = int(4 * width_resc)
+ gap_central = int(2 * width_resc)
top_bbox_red = 0 #int(height_resc/2)
for ind in range(coord.shape[1]):
@@ -128,7 +128,7 @@ def find_mainHorandCentral_Hough(img, coord, dim_bbox_page,
# regions of 1 or 2 columns. For documents of type 1, always to 1
dim_page = img_aux.shape
flag_2col, sections_page = n_cols_perblock(coord, coord_horz, dim_page, gap_central)
- print(flag_2col, sections_page)
+ print(flag_2col)#, sections_page)
if flag_type == 1:
flag_2col = 1
aux_sections_page = copy.deepcopy(sections_page)
@@ -155,19 +155,26 @@ def find_mainHorandCentral_Hough(img, coord, dim_bbox_page,
lines_vert = probabilistic_hough_line(edges_cent, theta = theta, line_length = int(2 * width_resc),
line_gap = int(width_resc))
if len(lines_vert):
- lines_vert = lines_vert.reshape((-1,4)) + np.array([0,in_dim,0,in_dim]).reshape((1,4))
+ lines_vert = np.concatenate(lines_vert, axis = 0).reshape((-1,4)) + np.array([0,in_dim,0,in_dim]).reshape((1,4))
aux_lines_vert.append(lines_vert)
sections_page[ikd][2] = 2
#print(lines_vert,img_aux.shape)
-
+ #print(flag_2col, sections_page)
if flag_central:
- lines_vert = np.transpose(np.concatenate(aux_lines_vert, axis = 0).reshape((-1,4)))
- #lines_vert = np.transpose(np.asarray(lines_vert).reshape((len(lines_vert),4)))
- lines_vert = np.concatenate((np.minimum(lines_vert[1,:],lines_vert[3,:]).reshape((1,lines_vert.shape[1])),
- np.minimum(lines_vert[0,:],lines_vert[2,:]).reshape((1,lines_vert.shape[1])),
- np.maximum(lines_vert[1,:],lines_vert[3,:]).reshape((1,lines_vert.shape[1])),
- np.maximum(lines_vert[0,:],lines_vert[2,:]).reshape((1,lines_vert.shape[1])))).astype(np.int32)
+ if len(aux_lines_vert):
+ lines_vert = np.transpose(np.concatenate(aux_lines_vert, axis = 0).reshape((-1,4)))
+ #lines_vert = np.transpose(np.asarray(lines_vert).reshape((len(lines_vert),4)))
+ lines_vert = np.concatenate((np.minimum(lines_vert[1,:],lines_vert[3,:]).reshape((1,lines_vert.shape[1])),
+ np.minimum(lines_vert[0,:],lines_vert[2,:]).reshape((1,lines_vert.shape[1])),
+ np.maximum(lines_vert[1,:],lines_vert[3,:]).reshape((1,lines_vert.shape[1])),
+ np.maximum(lines_vert[0,:],lines_vert[2,:]).reshape((1,lines_vert.shape[1])))).astype(np.int32)
+ else:
+ flag_central = 0
+ flag_2col = 0
+ coord_vert_def = np.zeros((4,1))
+
+
else:
sum_img_aux_in = np.sum(img_aux_in, axis = 0)
sum_img_aux_in = sum_img_aux_in[int(2*img_aux.shape[1]/5):int(3*img_aux.shape[1]/5)]
@@ -185,13 +192,14 @@ def find_mainHorandCentral_Hough(img, coord, dim_bbox_page,
ind_central = int((ind_min_start + ind_min_end)/2)
coord_vert_def = np.array([1, ind_central - int(width_resc/2),
img_aux_in.shape[0], ind_central + int(width_resc/2)])
-
+ coord_vert_def = clean_vert_line(coord_vert_def, width_resc, ind_central)
#lines_horz = transform_coord(lines_horz, dim_page = img_aux.shape, invert_xy = True)
#lines_vert = transform_coord(lines_vert, dim_page = img_aux.shape, invert_xy = True)
# First clean the vertical from unexpected outliers
if flag_central:
+ img_cent = np.copy(img_prev)
sum_rows = np.sum(img_cent, axis = 0)/255
ind_central = int(2*img.shape[1]/5) + np.argmax(sum_rows[int(2*img.shape[1]/5):int(3*img.shape[1]/5)])
ind_valid = np.intersect1d(np.argwhere([(ind_central - gap_central) < aux_l1 < (ind_central + gap_central) for aux_l1 in lines_vert[1,:]]),
@@ -203,7 +211,7 @@ def find_mainHorandCentral_Hough(img, coord, dim_bbox_page,
else:
coord_vert_def = np.array([0, img_aux.shape[1]/2 - width_resc, height_resc, img_aux.shape[1]/2 + width_resc])
- coord_vert_def = clean_vert_line(coord_vert_def, width_resc, ind_central)
+ coord_vert_def = clean_vert_line(coord_vert_def, width_resc, ind_central)
#ind_central = np.mean(coord_vert_def[[1,3]])
else:
@@ -214,7 +222,6 @@ def find_mainHorandCentral_Hough(img, coord, dim_bbox_page,
else:
coord_horz_def = coord_horz
- # And now, just iterate over the horizontal lines, merging them if required.
return coord_vert_def.astype(int), coord_horz_def.astype(int), flag_2col, sections_page
def n_cols_perblock(coord_textl, coord_horz_def, dim_page, gap_central):
@@ -685,7 +692,7 @@ def order_textl(rescale_factor, coord_vert_def, coord_horz, list_allcoords_textl
height_resc = HEIGHT_CHAR * rescale_factor[0,1]
widht_resc = WIDTH_CHAR * rescale_factor[0,1]
- gap_central = 3 * widht_resc
+ gap_central = int(2 * widht_resc)
gap_row = height_resc/2
sections_page_mat = np.array([]).reshape((3,0))
@@ -725,21 +732,31 @@ def order_textl(rescale_factor, coord_vert_def, coord_horz, list_allcoords_textl
-1 * np.ones(array_coords_coll.shape[1]).reshape((1,array_coords_coll.shape[1]))))
not_visited = 1
- toprow = 0
count_b = 0
+ toprow = 0
set_of_blocks = dict()
array_coords_centrall_ord = np.array([]).reshape((4,0))
+ #for i_col in range(sections_page_mat.shape[1]):
while not_visited:
-
+ #if array_coords_centrall.size > 3:
+ #toprow = sections_page_mat[0,i_col]
+ #bottomrow = np.min(array_coords_centrall[0,:])
+ #bottomrow = sections_page_mat[1,i_col]
+ #if i_col < (sections_page_mat.shape[1] - 1):
if array_coords_centrall.size > 3:
bottomrow = np.min(array_coords_centrall[0,:])
ind_bottomrow = np.argmin(array_coords_centrall[0,:])
array_coords_centrall_ord = np.concatenate((array_coords_centrall_ord,
array_coords_centrall[:,ind_bottomrow].reshape((4,1))), axis = 1)
+ aux_ch_coords = array_coords_centrall[:,ind_bottomrow]
+ ind_horz_cent = aux_ch_coords[2] - aux_ch_coords[0]
+ ind_val = np.intersect1d(np.argwhere(sections_page_mat[0,:] <= ind_horz_cent),
+ np.argwhere(sections_page_mat[1,:] > ind_horz_cent))
array_coords_centrall = np.delete(array_coords_centrall,ind_bottomrow,1)
else:
bottomrow = 10000
not_visited = 0
+ ind_val = sections_page_mat.shape[1] - 1
ind_textl_proc = np.intersect1d(np.argwhere((array_coords_textl[2,:] - gap_row) >= toprow),
np.argwhere((array_coords_textl[2,:] - gap_row) < bottomrow))
@@ -755,9 +772,11 @@ def order_textl(rescale_factor, coord_vert_def, coord_horz, list_allcoords_textl
array_coords_coll[:,ind_lines_proc].reshape(5,len(ind_lines_proc))),
axis = 1)
- n_cols = identify_n_cols(coord_cat[:4,:], central_line, gap_central)
- flag_2col_aux = n_cols - 1
+ #n_cols = identify_n_cols(coord_cat[:4,:], central_line, gap_central)
+ #flag_2col_aux = n_cols - 1
+ flag_2col_aux = int(sections_page_mat[2,ind_val] - 1)
+
if coord_cat.size > 0:
if flag_order:
flag_col = 1
@@ -822,6 +841,7 @@ def order_textl(rescale_factor, coord_vert_def, coord_horz, list_allcoords_textl
toprow = np.copy(bottomrow)
set_of_blocks[count_b] = order_coords
count_b += 1
+
return set_of_blocks, array_coords_centrall_ord
--
GitLab