Spaces:
Build error
Build error
First commit
Browse files- .gitattributes +1 -0
- app.py +239 -0
- packages.txt +6 -0
- postprocess.py +895 -0
- requirements.txt +9 -0
- tessdata/eng.traineddata +3 -0
- weights/structure_wts.pt +3 -0
.gitattributes
CHANGED
|
@@ -32,3 +32,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
| 32 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 33 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
| 32 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 33 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
| 35 |
+
*.traineddata filter=lfs diff=lfs merge=lfs -text
|
app.py
ADDED
|
@@ -0,0 +1,239 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import streamlit as st
|
| 2 |
+
|
| 3 |
+
import PIL
|
| 4 |
+
import numpy as np
|
| 5 |
+
import torch
|
| 6 |
+
from collections import defaultdict
|
| 7 |
+
|
| 8 |
+
import cv2
|
| 9 |
+
from doctr.io import DocumentFile
|
| 10 |
+
from doctr.models import ocr_predictor
|
| 11 |
+
from doctr.utils.visualization import visualize_page
|
| 12 |
+
|
| 13 |
+
import pytesseract
|
| 14 |
+
from pytesseract import Output
|
| 15 |
+
|
| 16 |
+
from bs4 import BeautifulSoup as bs
|
| 17 |
+
|
| 18 |
+
import sys, json
|
| 19 |
+
|
| 20 |
+
import postprocess
|
| 21 |
+
|
| 22 |
+
|
| 23 |
+
ocr_predictor = ocr_predictor('db_resnet50', 'crnn_vgg16_bn', pretrained=True)
|
| 24 |
+
structure_model = torch.hub.load('ultralytics/yolov5', 'custom', 'weights/structure_wts.pt', force_reload=True)
|
| 25 |
+
imgsz = 640
|
| 26 |
+
|
| 27 |
+
structure_class_names = [
|
| 28 |
+
'table', 'table column', 'table row', 'table column header',
|
| 29 |
+
'table projected row header', 'table spanning cell', 'no object'
|
| 30 |
+
]
|
| 31 |
+
structure_class_map = {k: v for v, k in enumerate(structure_class_names)}
|
| 32 |
+
structure_class_thresholds = {
|
| 33 |
+
"table": 0.5,
|
| 34 |
+
"table column": 0.5,
|
| 35 |
+
"table row": 0.5,
|
| 36 |
+
"table column header": 0.25,
|
| 37 |
+
"table projected row header": 0.25,
|
| 38 |
+
"table spanning cell": 0.25,
|
| 39 |
+
"no object": 10
|
| 40 |
+
}
|
| 41 |
+
|
| 42 |
+
|
| 43 |
+
def table_structure(filename):
|
| 44 |
+
image = cv2.imread(filename)
|
| 45 |
+
pred = structure_model(image, size=imgsz)
|
| 46 |
+
pred = pred.xywhn[0]
|
| 47 |
+
result = pred.cpu().numpy()
|
| 48 |
+
return result
|
| 49 |
+
|
| 50 |
+
|
| 51 |
+
def ocr(filename):
|
| 52 |
+
doc = DocumentFile.from_images(filename)
|
| 53 |
+
result = ocr_predictor(doc).export()
|
| 54 |
+
result = result['pages'][0]
|
| 55 |
+
H, W = result['dimensions']
|
| 56 |
+
ocr_res = []
|
| 57 |
+
for block in result['blocks']:
|
| 58 |
+
for line in block['lines']:
|
| 59 |
+
for word in line['words']:
|
| 60 |
+
bbox = word['geometry']
|
| 61 |
+
word_info = {
|
| 62 |
+
'bbox': [int(bbox[0][0] * W), int(bbox[0][1] * H), int(bbox[1][0] * W), int(bbox[1][1] * H)],
|
| 63 |
+
'text': word['value']
|
| 64 |
+
}
|
| 65 |
+
ocr_res.append(word_info)
|
| 66 |
+
return ocr_res
|
| 67 |
+
|
| 68 |
+
|
| 69 |
+
def convert_stucture(page_tokens, filename, structure_result):
|
| 70 |
+
image = cv2.imread(filename)
|
| 71 |
+
width = image.shape[1]
|
| 72 |
+
height = image.shape[0]
|
| 73 |
+
# print(width, height)
|
| 74 |
+
|
| 75 |
+
bboxes = []
|
| 76 |
+
scores = []
|
| 77 |
+
labels = []
|
| 78 |
+
for i, result in enumerate(structure_result):
|
| 79 |
+
class_id = int(result[5])
|
| 80 |
+
score = float(result[4])
|
| 81 |
+
min_x = result[0]
|
| 82 |
+
min_y = result[1]
|
| 83 |
+
w = result[2]
|
| 84 |
+
h = result[3]
|
| 85 |
+
|
| 86 |
+
x1 = int((min_x-w/2)*width)
|
| 87 |
+
y1 = int((min_y-h/2)*height)
|
| 88 |
+
x2 = int((min_x+w/2)*width)
|
| 89 |
+
y2 = int((min_y+h/2)*height)
|
| 90 |
+
# print(x1, y1, x2, y2)
|
| 91 |
+
|
| 92 |
+
bboxes.append([x1, y1, x2, y2])
|
| 93 |
+
scores.append(score)
|
| 94 |
+
labels.append(class_id)
|
| 95 |
+
|
| 96 |
+
table_objects = []
|
| 97 |
+
for bbox, score, label in zip(bboxes, scores, labels):
|
| 98 |
+
table_objects.append({'bbox': bbox, 'score': score, 'label': label})
|
| 99 |
+
# print('table_objects:', table_objects)
|
| 100 |
+
|
| 101 |
+
table = {'objects': table_objects, 'page_num': 0}
|
| 102 |
+
|
| 103 |
+
table_class_objects = [obj for obj in table_objects if obj['label'] == structure_class_map['table']]
|
| 104 |
+
if len(table_class_objects) > 1:
|
| 105 |
+
table_class_objects = sorted(table_class_objects, key=lambda x: x['score'], reverse=True)
|
| 106 |
+
try:
|
| 107 |
+
table_bbox = list(table_class_objects[0]['bbox'])
|
| 108 |
+
except:
|
| 109 |
+
table_bbox = (0,0,1000,1000)
|
| 110 |
+
# print('table_class_objects:', table_class_objects)
|
| 111 |
+
# print('table_bbox:', table_bbox)
|
| 112 |
+
|
| 113 |
+
tokens_in_table = [token for token in page_tokens if postprocess.iob(token['bbox'], table_bbox) >= 0.5]
|
| 114 |
+
# print('tokens_in_table:', tokens_in_table)
|
| 115 |
+
|
| 116 |
+
table_structures, cells, confidence_score = postprocess.objects_to_cells(table, table_objects, tokens_in_table, structure_class_names, structure_class_thresholds)
|
| 117 |
+
|
| 118 |
+
return table_structures, cells, confidence_score
|
| 119 |
+
|
| 120 |
+
|
| 121 |
+
def visualize_cells(filename, cells, ax):
|
| 122 |
+
image = cv2.imread(filename)
|
| 123 |
+
for i, cell in enumerate(cells):
|
| 124 |
+
bbox = cell['bbox']
|
| 125 |
+
x1 = int(bbox[0])
|
| 126 |
+
y1 = int(bbox[1])
|
| 127 |
+
x2 = int(bbox[2])
|
| 128 |
+
y2 = int(bbox[3])
|
| 129 |
+
cv2.rectangle(image, (x1, y1), (x2, y2), color=(0, 255, 0))
|
| 130 |
+
ax.image(image)
|
| 131 |
+
|
| 132 |
+
|
| 133 |
+
def pytess(cell_pil_img):
|
| 134 |
+
return ' '.join(pytesseract.image_to_data(cell_pil_img, output_type=Output.DICT, config='-c tessedit_char_blacklist=œ˜â€œï¬â™Ã©œ¢!|”?«“¥ --tessdata-dir tessdata --oem 3 --psm 6')['text']).strip()
|
| 135 |
+
|
| 136 |
+
|
| 137 |
+
def resize(pil_img, size=1800):
|
| 138 |
+
length_x, width_y = pil_img.size
|
| 139 |
+
factor = max(1, size / length_x)
|
| 140 |
+
size = int(factor * length_x), int(factor * width_y)
|
| 141 |
+
pil_img = pil_img.resize(size, PIL.Image.ANTIALIAS)
|
| 142 |
+
return pil_img, factor
|
| 143 |
+
|
| 144 |
+
|
| 145 |
+
def image_smoothening(img):
|
| 146 |
+
ret1, th1 = cv2.threshold(img, 180, 255, cv2.THRESH_BINARY)
|
| 147 |
+
ret2, th2 = cv2.threshold(th1, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
|
| 148 |
+
blur = cv2.GaussianBlur(th2, (1, 1), 0)
|
| 149 |
+
ret3, th3 = cv2.threshold(blur, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
|
| 150 |
+
return th3
|
| 151 |
+
|
| 152 |
+
|
| 153 |
+
def remove_noise_and_smooth(pil_img):
|
| 154 |
+
img = cv2.cvtColor(np.array(pil_img), cv2.COLOR_RGB2GRAY)
|
| 155 |
+
filtered = cv2.adaptiveThreshold(img.astype(np.uint8), 255, cv2.ADAPTIVE_THRESH_MEAN_C, cv2.THRESH_BINARY, 41, 3)
|
| 156 |
+
kernel = np.ones((1, 1), np.uint8)
|
| 157 |
+
opening = cv2.morphologyEx(filtered, cv2.MORPH_OPEN, kernel)
|
| 158 |
+
closing = cv2.morphologyEx(opening, cv2.MORPH_CLOSE, kernel)
|
| 159 |
+
img = image_smoothening(img)
|
| 160 |
+
or_image = cv2.bitwise_or(img, closing)
|
| 161 |
+
pil_img = PIL.Image.fromarray(or_image)
|
| 162 |
+
return pil_img
|
| 163 |
+
|
| 164 |
+
|
| 165 |
+
def extract_text_from_cells(filename, cells):
|
| 166 |
+
pil_img = PIL.Image.open(filename)
|
| 167 |
+
pil_img, factor = resize(pil_img)
|
| 168 |
+
#pil_img = remove_noise_and_smooth(pil_img)
|
| 169 |
+
#display(pil_img)
|
| 170 |
+
for cell in cells:
|
| 171 |
+
bbox = [x * factor for x in cell['bbox']]
|
| 172 |
+
cell_pil_img = pil_img.crop(bbox)
|
| 173 |
+
#cell_pil_img = remove_noise_and_smooth(cell_pil_img)
|
| 174 |
+
#cell_pil_img = tess_prep(cell_pil_img)
|
| 175 |
+
cell['text'] = pytess(cell_pil_img)
|
| 176 |
+
return cells
|
| 177 |
+
|
| 178 |
+
|
| 179 |
+
def cells_to_html(cells):
|
| 180 |
+
n_cols = max(cell['column_nums'][-1] for cell in cells) + 1
|
| 181 |
+
n_rows = max(cell['row_nums'][-1] for cell in cells) + 1
|
| 182 |
+
html_code = ''
|
| 183 |
+
for r in range(n_rows):
|
| 184 |
+
r_cells = [cell for cell in cells if cell['row_nums'][0] == r]
|
| 185 |
+
r_cells.sort(key=lambda x: x['column_nums'][0])
|
| 186 |
+
r_html = ''
|
| 187 |
+
for cell in r_cells:
|
| 188 |
+
rowspan = cell['row_nums'][-1] - cell['row_nums'][0] + 1
|
| 189 |
+
colspan = cell['column_nums'][-1] - cell['column_nums'][0] + 1
|
| 190 |
+
r_html += f'<td rowspan="{rowspan}" colspan="{colspan}">{cell["text"]}</td>'
|
| 191 |
+
html_code += f'<tr>{r_html}</tr>'
|
| 192 |
+
html_code = '''<html>
|
| 193 |
+
<head>
|
| 194 |
+
<meta charset="UTF-8">
|
| 195 |
+
<style>
|
| 196 |
+
table, th, td {
|
| 197 |
+
border: 1px solid black;
|
| 198 |
+
font-size: 10px;
|
| 199 |
+
}
|
| 200 |
+
</style>
|
| 201 |
+
</head>
|
| 202 |
+
<body>
|
| 203 |
+
<table frame="hsides" rules="groups" width="100%%">
|
| 204 |
+
%s
|
| 205 |
+
</table>
|
| 206 |
+
</body>
|
| 207 |
+
</html>''' % html_code
|
| 208 |
+
soup = bs(html_code)
|
| 209 |
+
html_code = soup.prettify()
|
| 210 |
+
return html_code
|
| 211 |
+
|
| 212 |
+
|
| 213 |
+
def main():
|
| 214 |
+
|
| 215 |
+
st.set_page_config(layout="wide")
|
| 216 |
+
st.title("Table Structure Recognition Demo")
|
| 217 |
+
st.write('\n')
|
| 218 |
+
|
| 219 |
+
cols = st.beta_columns((1, 1, 1))
|
| 220 |
+
cols[0].subheader("Input page")
|
| 221 |
+
cols[1].subheader("Structure output")
|
| 222 |
+
cols[2].subheader("HTML output")
|
| 223 |
+
|
| 224 |
+
st.sidebar.title("Image upload")
|
| 225 |
+
st.set_option('deprecation.showfileUploaderEncoding', False)
|
| 226 |
+
filename = st.sidebar.file_uploader("Upload files", type=['png', 'jpeg', 'jpg'])
|
| 227 |
+
|
| 228 |
+
cols[0].image(cv2.imread(filename))
|
| 229 |
+
|
| 230 |
+
ocr_res = ocr(filename)
|
| 231 |
+
structure_result = table_structure(filename)
|
| 232 |
+
table_structures, cells, confidence_score = convert_stucture(ocr_res, filename, structure_result)
|
| 233 |
+
visualize_cells(filename, cells, cols[1])
|
| 234 |
+
|
| 235 |
+
cells = extract_text_from_cells(filename, cells)
|
| 236 |
+
html_code = cells_to_html(cells)
|
| 237 |
+
|
| 238 |
+
cols[2].html(html_code)
|
| 239 |
+
|
packages.txt
ADDED
|
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
ffmpeg
|
| 2 |
+
libsm6
|
| 3 |
+
libxext6
|
| 4 |
+
libgl1
|
| 5 |
+
tesseract-ocr-eng
|
| 6 |
+
python3-opencv
|
postprocess.py
ADDED
|
@@ -0,0 +1,895 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
Copyright (C) 2021 Microsoft Corporation
|
| 3 |
+
"""
|
| 4 |
+
from collections import defaultdict
|
| 5 |
+
|
| 6 |
+
from fitz import Rect
|
| 7 |
+
|
| 8 |
+
|
| 9 |
+
def apply_threshold(objects, threshold):
|
| 10 |
+
"""
|
| 11 |
+
Filter out objects below a certain score.
|
| 12 |
+
"""
|
| 13 |
+
return [obj for obj in objects if obj['score'] >= threshold]
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
def apply_class_thresholds(bboxes, labels, scores, class_names, class_thresholds):
|
| 17 |
+
"""
|
| 18 |
+
Filter out bounding boxes whose confidence is below the confidence threshold for
|
| 19 |
+
its associated class label.
|
| 20 |
+
"""
|
| 21 |
+
# Apply class-specific thresholds
|
| 22 |
+
indices_above_threshold = [idx for idx, (score, label) in enumerate(zip(scores, labels))
|
| 23 |
+
if score >= class_thresholds[
|
| 24 |
+
class_names[label]
|
| 25 |
+
]
|
| 26 |
+
]
|
| 27 |
+
bboxes = [bboxes[idx] for idx in indices_above_threshold]
|
| 28 |
+
scores = [scores[idx] for idx in indices_above_threshold]
|
| 29 |
+
labels = [labels[idx] for idx in indices_above_threshold]
|
| 30 |
+
|
| 31 |
+
return bboxes, scores, labels
|
| 32 |
+
|
| 33 |
+
|
| 34 |
+
def iou(bbox1, bbox2):
|
| 35 |
+
"""
|
| 36 |
+
Compute the intersection-over-union of two bounding boxes.
|
| 37 |
+
"""
|
| 38 |
+
intersection = Rect(bbox1).intersect(bbox2)
|
| 39 |
+
union = Rect(bbox1).include_rect(bbox2)
|
| 40 |
+
|
| 41 |
+
union_area = union.get_area() # getArea()
|
| 42 |
+
if union_area > 0:
|
| 43 |
+
return intersection.get_area() / union.get_area() # .getArea()
|
| 44 |
+
|
| 45 |
+
return 0
|
| 46 |
+
|
| 47 |
+
|
| 48 |
+
def iob(bbox1, bbox2):
|
| 49 |
+
"""
|
| 50 |
+
Compute the intersection area over box area, for bbox1.
|
| 51 |
+
"""
|
| 52 |
+
intersection = Rect(bbox1).intersect(bbox2)
|
| 53 |
+
|
| 54 |
+
bbox1_area = Rect(bbox1).get_area() # .getArea()
|
| 55 |
+
if bbox1_area > 0:
|
| 56 |
+
return intersection.get_area() / bbox1_area # getArea()
|
| 57 |
+
|
| 58 |
+
return 0
|
| 59 |
+
|
| 60 |
+
|
| 61 |
+
def objects_to_cells(table, objects_in_table, tokens_in_table, class_map, class_thresholds):
|
| 62 |
+
"""
|
| 63 |
+
Process the bounding boxes produced by the table structure recognition model
|
| 64 |
+
and the token/word/span bounding boxes into table cells.
|
| 65 |
+
|
| 66 |
+
Also return a confidence score based on how well the text was able to be
|
| 67 |
+
uniquely slotted into the cells detected by the table model.
|
| 68 |
+
"""
|
| 69 |
+
|
| 70 |
+
table_structures = objects_to_table_structures(table, objects_in_table, tokens_in_table, class_map,
|
| 71 |
+
class_thresholds)
|
| 72 |
+
|
| 73 |
+
# Check for a valid table
|
| 74 |
+
if len(table_structures['columns']) < 1 or len(table_structures['rows']) < 1:
|
| 75 |
+
cells = []#None
|
| 76 |
+
confidence_score = 0
|
| 77 |
+
else:
|
| 78 |
+
cells, confidence_score = table_structure_to_cells(table_structures, tokens_in_table, table['bbox'])
|
| 79 |
+
|
| 80 |
+
return table_structures, cells, confidence_score
|
| 81 |
+
|
| 82 |
+
|
| 83 |
+
def objects_to_table_structures(table_object, objects_in_table, tokens_in_table, class_names, class_thresholds):
|
| 84 |
+
"""
|
| 85 |
+
Process the bounding boxes produced by the table structure recognition model into
|
| 86 |
+
a *consistent* set of table structures (rows, columns, supercells, headers).
|
| 87 |
+
This entails resolving conflicts/overlaps, and ensuring the boxes meet certain alignment
|
| 88 |
+
conditions (for example: rows should all have the same width, etc.).
|
| 89 |
+
"""
|
| 90 |
+
|
| 91 |
+
page_num = table_object['page_num']
|
| 92 |
+
|
| 93 |
+
table_structures = {}
|
| 94 |
+
|
| 95 |
+
columns = [obj for obj in objects_in_table if class_names[obj['label']] == 'table column']
|
| 96 |
+
rows = [obj for obj in objects_in_table if class_names[obj['label']] == 'table row']
|
| 97 |
+
headers = [obj for obj in objects_in_table if class_names[obj['label']] == 'table column header']
|
| 98 |
+
supercells = [obj for obj in objects_in_table if class_names[obj['label']] == 'table spanning cell']
|
| 99 |
+
for obj in supercells:
|
| 100 |
+
obj['subheader'] = False
|
| 101 |
+
subheaders = [obj for obj in objects_in_table if class_names[obj['label']] == 'table projected row header']
|
| 102 |
+
for obj in subheaders:
|
| 103 |
+
obj['subheader'] = True
|
| 104 |
+
supercells += subheaders
|
| 105 |
+
for obj in rows:
|
| 106 |
+
obj['header'] = False
|
| 107 |
+
for header_obj in headers:
|
| 108 |
+
if iob(obj['bbox'], header_obj['bbox']) >= 0.5:
|
| 109 |
+
obj['header'] = True
|
| 110 |
+
|
| 111 |
+
for row in rows:
|
| 112 |
+
row['page'] = page_num
|
| 113 |
+
|
| 114 |
+
for column in columns:
|
| 115 |
+
column['page'] = page_num
|
| 116 |
+
|
| 117 |
+
#Refine table structures
|
| 118 |
+
rows = refine_rows(rows, tokens_in_table, class_thresholds['table row'])
|
| 119 |
+
columns = refine_columns(columns, tokens_in_table, class_thresholds['table column'])
|
| 120 |
+
|
| 121 |
+
# Shrink table bbox to just the total height of the rows
|
| 122 |
+
# and the total width of the columns
|
| 123 |
+
row_rect = Rect()
|
| 124 |
+
for obj in rows:
|
| 125 |
+
row_rect.include_rect(obj['bbox'])
|
| 126 |
+
column_rect = Rect()
|
| 127 |
+
for obj in columns:
|
| 128 |
+
column_rect.include_rect(obj['bbox'])
|
| 129 |
+
table_object['row_column_bbox'] = [column_rect[0], row_rect[1], column_rect[2], row_rect[3]]
|
| 130 |
+
table_object['bbox'] = table_object['row_column_bbox']
|
| 131 |
+
|
| 132 |
+
# Process the rows and columns into a complete segmented table
|
| 133 |
+
columns = align_columns(columns, table_object['row_column_bbox'])
|
| 134 |
+
rows = align_rows(rows, table_object['row_column_bbox'])
|
| 135 |
+
|
| 136 |
+
table_structures['rows'] = rows
|
| 137 |
+
table_structures['columns'] = columns
|
| 138 |
+
table_structures['headers'] = headers
|
| 139 |
+
table_structures['supercells'] = supercells
|
| 140 |
+
|
| 141 |
+
if len(rows) > 0 and len(columns) > 1:
|
| 142 |
+
table_structures = refine_table_structures(table_object['bbox'], table_structures, tokens_in_table, class_thresholds)
|
| 143 |
+
|
| 144 |
+
return table_structures
|
| 145 |
+
|
| 146 |
+
|
| 147 |
+
def refine_rows(rows, page_spans, score_threshold):
|
| 148 |
+
"""
|
| 149 |
+
Apply operations to the detected rows, such as
|
| 150 |
+
thresholding, NMS, and alignment.
|
| 151 |
+
"""
|
| 152 |
+
|
| 153 |
+
#MODIFY
|
| 154 |
+
rows = [obj for obj in rows if obj['score'] >= score_threshold or obj['header']]
|
| 155 |
+
###
|
| 156 |
+
|
| 157 |
+
rows = nms_by_containment(rows, page_spans, overlap_threshold=0.5)
|
| 158 |
+
# remove_objects_without_content(page_spans, rows) # TODO
|
| 159 |
+
if len(rows) > 1:
|
| 160 |
+
rows = sort_objects_top_to_bottom(rows)
|
| 161 |
+
|
| 162 |
+
return rows
|
| 163 |
+
|
| 164 |
+
|
| 165 |
+
def refine_columns(columns, page_spans, score_threshold):
|
| 166 |
+
"""
|
| 167 |
+
Apply operations to the detected columns, such as
|
| 168 |
+
thresholding, NMS, and alignment.
|
| 169 |
+
"""
|
| 170 |
+
|
| 171 |
+
#MODIFY
|
| 172 |
+
columns = [obj for obj in columns if obj['score'] >= score_threshold]
|
| 173 |
+
###
|
| 174 |
+
|
| 175 |
+
columns = nms_by_containment(columns, page_spans, overlap_threshold=0.5)
|
| 176 |
+
# remove_objects_without_content(page_spans, columns) # TODO
|
| 177 |
+
if len(columns) > 1:
|
| 178 |
+
columns = sort_objects_left_to_right(columns)
|
| 179 |
+
|
| 180 |
+
return columns
|
| 181 |
+
|
| 182 |
+
|
| 183 |
+
def nms_by_containment(container_objects, package_objects, overlap_threshold=0.5):
|
| 184 |
+
"""
|
| 185 |
+
Non-maxima suppression (NMS) of objects based on shared containment of other objects.
|
| 186 |
+
"""
|
| 187 |
+
container_objects = sort_objects_by_score(container_objects)
|
| 188 |
+
num_objects = len(container_objects)
|
| 189 |
+
suppression = [False for obj in container_objects]
|
| 190 |
+
|
| 191 |
+
packages_by_container, _, _ = slot_into_containers(container_objects, package_objects, overlap_threshold=overlap_threshold,
|
| 192 |
+
unique_assignment=True, forced_assignment=False)
|
| 193 |
+
|
| 194 |
+
for object2_num in range(1, num_objects):
|
| 195 |
+
object2_packages = set(packages_by_container[object2_num])
|
| 196 |
+
if len(object2_packages) == 0:
|
| 197 |
+
suppression[object2_num] = True
|
| 198 |
+
for object1_num in range(object2_num):
|
| 199 |
+
if not suppression[object1_num]:
|
| 200 |
+
object1_packages = set(packages_by_container[object1_num])
|
| 201 |
+
if len(object2_packages.intersection(object1_packages)) > 0:
|
| 202 |
+
suppression[object2_num] = True
|
| 203 |
+
|
| 204 |
+
final_objects = [obj for idx, obj in enumerate(container_objects) if not suppression[idx]]
|
| 205 |
+
return final_objects
|
| 206 |
+
|
| 207 |
+
|
| 208 |
+
def slot_into_containers(container_objects, package_objects, overlap_threshold=0.5,
|
| 209 |
+
unique_assignment=True, forced_assignment=False):
|
| 210 |
+
"""
|
| 211 |
+
Slot a collection of objects into the container they occupy most (the container which holds the largest fraction of the object).
|
| 212 |
+
"""
|
| 213 |
+
best_match_scores = []
|
| 214 |
+
|
| 215 |
+
container_assignments = [[] for container in container_objects]
|
| 216 |
+
package_assignments = [[] for package in package_objects]
|
| 217 |
+
|
| 218 |
+
if len(container_objects) == 0 or len(package_objects) == 0:
|
| 219 |
+
return container_assignments, package_assignments, best_match_scores
|
| 220 |
+
|
| 221 |
+
match_scores = defaultdict(dict)
|
| 222 |
+
for package_num, package in enumerate(package_objects):
|
| 223 |
+
match_scores = []
|
| 224 |
+
package_rect = Rect(package['bbox'])
|
| 225 |
+
package_area = package_rect.get_area() # getArea()
|
| 226 |
+
for container_num, container in enumerate(container_objects):
|
| 227 |
+
container_rect = Rect(container['bbox'])
|
| 228 |
+
intersect_area = container_rect.intersect(package['bbox']).get_area() # getArea()
|
| 229 |
+
overlap_fraction = intersect_area / package_area
|
| 230 |
+
match_scores.append({'container': container, 'container_num': container_num, 'score': overlap_fraction})
|
| 231 |
+
|
| 232 |
+
sorted_match_scores = sort_objects_by_score(match_scores)
|
| 233 |
+
|
| 234 |
+
best_match_score = sorted_match_scores[0]
|
| 235 |
+
best_match_scores.append(best_match_score['score'])
|
| 236 |
+
if forced_assignment or best_match_score['score'] >= overlap_threshold:
|
| 237 |
+
container_assignments[best_match_score['container_num']].append(package_num)
|
| 238 |
+
package_assignments[package_num].append(best_match_score['container_num'])
|
| 239 |
+
|
| 240 |
+
if not unique_assignment: # slot package into all eligible slots
|
| 241 |
+
for match_score in sorted_match_scores[1:]:
|
| 242 |
+
if match_score['score'] >= overlap_threshold:
|
| 243 |
+
container_assignments[match_score['container_num']].append(package_num)
|
| 244 |
+
package_assignments[package_num].append(match_score['container_num'])
|
| 245 |
+
else:
|
| 246 |
+
break
|
| 247 |
+
|
| 248 |
+
return container_assignments, package_assignments, best_match_scores
|
| 249 |
+
|
| 250 |
+
|
| 251 |
+
def sort_objects_by_score(objects, reverse=True):
|
| 252 |
+
"""
|
| 253 |
+
Put any set of objects in order from high score to low score.
|
| 254 |
+
"""
|
| 255 |
+
if reverse:
|
| 256 |
+
sign = -1
|
| 257 |
+
else:
|
| 258 |
+
sign = 1
|
| 259 |
+
return sorted(objects, key=lambda k: sign*k['score'])
|
| 260 |
+
|
| 261 |
+
|
| 262 |
+
def remove_objects_without_content(page_spans, objects):
|
| 263 |
+
"""
|
| 264 |
+
Remove any objects (these can be rows, columns, supercells, etc.) that don't
|
| 265 |
+
have any text associated with them.
|
| 266 |
+
"""
|
| 267 |
+
for obj in objects[:]:
|
| 268 |
+
object_text, _ = extract_text_inside_bbox(page_spans, obj['bbox'])
|
| 269 |
+
if len(object_text.strip()) == 0:
|
| 270 |
+
objects.remove(obj)
|
| 271 |
+
|
| 272 |
+
|
| 273 |
+
def extract_text_inside_bbox(spans, bbox):
|
| 274 |
+
"""
|
| 275 |
+
Extract the text inside a bounding box.
|
| 276 |
+
"""
|
| 277 |
+
bbox_spans = get_bbox_span_subset(spans, bbox)
|
| 278 |
+
bbox_text = extract_text_from_spans(bbox_spans, remove_integer_superscripts=True)
|
| 279 |
+
|
| 280 |
+
return bbox_text, bbox_spans
|
| 281 |
+
|
| 282 |
+
|
| 283 |
+
def get_bbox_span_subset(spans, bbox, threshold=0.5):
|
| 284 |
+
"""
|
| 285 |
+
Reduce the set of spans to those that fall within a bounding box.
|
| 286 |
+
|
| 287 |
+
threshold: the fraction of the span that must overlap with the bbox.
|
| 288 |
+
"""
|
| 289 |
+
span_subset = []
|
| 290 |
+
for span in spans:
|
| 291 |
+
if overlaps(span['bbox'], bbox, threshold):
|
| 292 |
+
span_subset.append(span)
|
| 293 |
+
return span_subset
|
| 294 |
+
|
| 295 |
+
|
| 296 |
+
def overlaps(bbox1, bbox2, threshold=0.5):
|
| 297 |
+
"""
|
| 298 |
+
Test if more than "threshold" fraction of bbox1 overlaps with bbox2.
|
| 299 |
+
"""
|
| 300 |
+
rect1 = Rect(list(bbox1))
|
| 301 |
+
area1 = rect1.get_area() # .getArea()
|
| 302 |
+
if area1 == 0:
|
| 303 |
+
return False
|
| 304 |
+
return rect1.intersect(list(bbox2)).get_area()/area1 >= threshold # getArea()
|
| 305 |
+
|
| 306 |
+
|
| 307 |
+
def extract_text_from_spans(spans, join_with_space=True, remove_integer_superscripts=True):
|
| 308 |
+
"""
|
| 309 |
+
Convert a collection of page tokens/words/spans into a single text string.
|
| 310 |
+
"""
|
| 311 |
+
|
| 312 |
+
if join_with_space:
|
| 313 |
+
join_char = " "
|
| 314 |
+
else:
|
| 315 |
+
join_char = ""
|
| 316 |
+
spans_copy = spans[:]
|
| 317 |
+
|
| 318 |
+
if remove_integer_superscripts:
|
| 319 |
+
for span in spans:
|
| 320 |
+
flags = span['flags']
|
| 321 |
+
if flags & 2**0: # superscript flag
|
| 322 |
+
if is_int(span['text']):
|
| 323 |
+
spans_copy.remove(span)
|
| 324 |
+
else:
|
| 325 |
+
span['superscript'] = True
|
| 326 |
+
|
| 327 |
+
if len(spans_copy) == 0:
|
| 328 |
+
return ""
|
| 329 |
+
|
| 330 |
+
spans_copy.sort(key=lambda span: span['span_num'])
|
| 331 |
+
spans_copy.sort(key=lambda span: span['line_num'])
|
| 332 |
+
spans_copy.sort(key=lambda span: span['block_num'])
|
| 333 |
+
|
| 334 |
+
# Force the span at the end of every line within a block to have exactly one space
|
| 335 |
+
# unless the line ends with a space or ends with a non-space followed by a hyphen
|
| 336 |
+
line_texts = []
|
| 337 |
+
line_span_texts = [spans_copy[0]['text']]
|
| 338 |
+
for span1, span2 in zip(spans_copy[:-1], spans_copy[1:]):
|
| 339 |
+
if not span1['block_num'] == span2['block_num'] or not span1['line_num'] == span2['line_num']:
|
| 340 |
+
line_text = join_char.join(line_span_texts).strip()
|
| 341 |
+
if (len(line_text) > 0
|
| 342 |
+
and not line_text[-1] == ' '
|
| 343 |
+
and not (len(line_text) > 1 and line_text[-1] == "-" and not line_text[-2] == ' ')):
|
| 344 |
+
if not join_with_space:
|
| 345 |
+
line_text += ' '
|
| 346 |
+
line_texts.append(line_text)
|
| 347 |
+
line_span_texts = [span2['text']]
|
| 348 |
+
else:
|
| 349 |
+
line_span_texts.append(span2['text'])
|
| 350 |
+
line_text = join_char.join(line_span_texts)
|
| 351 |
+
line_texts.append(line_text)
|
| 352 |
+
|
| 353 |
+
return join_char.join(line_texts).strip()
|
| 354 |
+
|
| 355 |
+
|
| 356 |
+
def sort_objects_left_to_right(objs):
|
| 357 |
+
"""
|
| 358 |
+
Put the objects in order from left to right.
|
| 359 |
+
"""
|
| 360 |
+
return sorted(objs, key=lambda k: k['bbox'][0] + k['bbox'][2])
|
| 361 |
+
|
| 362 |
+
|
| 363 |
+
def sort_objects_top_to_bottom(objs):
|
| 364 |
+
"""
|
| 365 |
+
Put the objects in order from top to bottom.
|
| 366 |
+
"""
|
| 367 |
+
return sorted(objs, key=lambda k: k['bbox'][1] + k['bbox'][3])
|
| 368 |
+
|
| 369 |
+
|
| 370 |
+
def align_columns(columns, bbox):
|
| 371 |
+
"""
|
| 372 |
+
For every column, align the top and bottom boundaries to the final
|
| 373 |
+
table bounding box.
|
| 374 |
+
"""
|
| 375 |
+
try:
|
| 376 |
+
for column in columns:
|
| 377 |
+
column['bbox'][1] = bbox[1]
|
| 378 |
+
column['bbox'][3] = bbox[3]
|
| 379 |
+
except Exception as err:
|
| 380 |
+
print("Could not align columns: {}".format(err))
|
| 381 |
+
pass
|
| 382 |
+
|
| 383 |
+
return columns
|
| 384 |
+
|
| 385 |
+
|
| 386 |
+
def align_rows(rows, bbox):
|
| 387 |
+
"""
|
| 388 |
+
For every row, align the left and right boundaries to the final
|
| 389 |
+
table bounding box.
|
| 390 |
+
"""
|
| 391 |
+
try:
|
| 392 |
+
for row in rows:
|
| 393 |
+
row['bbox'][0] = bbox[0]
|
| 394 |
+
row['bbox'][2] = bbox[2]
|
| 395 |
+
except Exception as err:
|
| 396 |
+
print("Could not align rows: {}".format(err))
|
| 397 |
+
pass
|
| 398 |
+
|
| 399 |
+
return rows
|
| 400 |
+
|
| 401 |
+
|
| 402 |
+
def refine_table_structures(table_bbox, table_structures, page_spans, class_thresholds):
|
| 403 |
+
"""
|
| 404 |
+
Apply operations to the detected table structure objects such as
|
| 405 |
+
thresholding, NMS, and alignment.
|
| 406 |
+
"""
|
| 407 |
+
rows = table_structures["rows"]
|
| 408 |
+
columns = table_structures['columns']
|
| 409 |
+
|
| 410 |
+
#columns = fill_column_gaps(columns, table_bbox)
|
| 411 |
+
#rows = fill_row_gaps(rows, table_bbox)
|
| 412 |
+
|
| 413 |
+
# Process the headers
|
| 414 |
+
headers = table_structures['headers']
|
| 415 |
+
headers = apply_threshold(headers, class_thresholds["table column header"])
|
| 416 |
+
headers = nms(headers)
|
| 417 |
+
headers = align_headers(headers, rows)
|
| 418 |
+
|
| 419 |
+
# Process supercells
|
| 420 |
+
supercells = [elem for elem in table_structures['supercells'] if not elem['subheader']]
|
| 421 |
+
subheaders = [elem for elem in table_structures['supercells'] if elem['subheader']]
|
| 422 |
+
supercells = apply_threshold(supercells, class_thresholds["table spanning cell"])
|
| 423 |
+
subheaders = apply_threshold(subheaders, class_thresholds["table projected row header"])
|
| 424 |
+
supercells += subheaders
|
| 425 |
+
# Align before NMS for supercells because alignment brings them into agreement
|
| 426 |
+
# with rows and columns first; if supercells still overlap after this operation,
|
| 427 |
+
# the threshold for NMS can basically be lowered to just above 0
|
| 428 |
+
supercells = align_supercells(supercells, rows, columns)
|
| 429 |
+
supercells = nms_supercells(supercells)
|
| 430 |
+
|
| 431 |
+
header_supercell_tree(supercells)
|
| 432 |
+
|
| 433 |
+
table_structures['columns'] = columns
|
| 434 |
+
table_structures['rows'] = rows
|
| 435 |
+
table_structures['supercells'] = supercells
|
| 436 |
+
table_structures['headers'] = headers
|
| 437 |
+
|
| 438 |
+
return table_structures
|
| 439 |
+
|
| 440 |
+
|
| 441 |
+
def nms(objects, match_criteria="object2_overlap", match_threshold=0.05, keep_metric="score", keep_higher=True):
|
| 442 |
+
"""
|
| 443 |
+
A customizable version of non-maxima suppression (NMS).
|
| 444 |
+
|
| 445 |
+
Default behavior: If a lower-confidence object overlaps more than 5% of its area
|
| 446 |
+
with a higher-confidence object, remove the lower-confidence object.
|
| 447 |
+
|
| 448 |
+
objects: set of dicts; each object dict must have a 'bbox' and a 'score' field
|
| 449 |
+
match_criteria: how to measure how much two objects "overlap"
|
| 450 |
+
match_threshold: the cutoff for determining that overlap requires suppression of one object
|
| 451 |
+
keep_metric: which metric to use to determine the object to keep
|
| 452 |
+
keep_higher: if True, keep the object with the higher metric; otherwise, keep the lower
|
| 453 |
+
"""
|
| 454 |
+
if len(objects) == 0:
|
| 455 |
+
return []
|
| 456 |
+
|
| 457 |
+
if keep_metric=="score":
|
| 458 |
+
objects = sort_objects_by_score(objects, reverse=keep_higher)
|
| 459 |
+
elif keep_metric=="area":
|
| 460 |
+
objects = sort_objects_by_area(objects, reverse=keep_higher)
|
| 461 |
+
|
| 462 |
+
num_objects = len(objects)
|
| 463 |
+
suppression = [False for obj in objects]
|
| 464 |
+
|
| 465 |
+
for object2_num in range(1, num_objects):
|
| 466 |
+
object2_rect = Rect(objects[object2_num]['bbox'])
|
| 467 |
+
object2_area = object2_rect.get_area() # .getArea()
|
| 468 |
+
for object1_num in range(object2_num):
|
| 469 |
+
if not suppression[object1_num]:
|
| 470 |
+
object1_rect = Rect(objects[object1_num]['bbox'])
|
| 471 |
+
object1_area = object1_rect.get_area() # .getArea()
|
| 472 |
+
intersect_area = object1_rect.intersect(object2_rect).get_area() # .getArea()
|
| 473 |
+
try:
|
| 474 |
+
if match_criteria=="object1_overlap":
|
| 475 |
+
metric = intersect_area / object1_area
|
| 476 |
+
elif match_criteria=="object2_overlap":
|
| 477 |
+
metric = intersect_area / object2_area
|
| 478 |
+
elif match_criteria=="iou":
|
| 479 |
+
metric = intersect_area / (object1_area + object2_area - intersect_area)
|
| 480 |
+
if metric >= match_threshold:
|
| 481 |
+
suppression[object2_num] = True
|
| 482 |
+
break
|
| 483 |
+
except Exception:
|
| 484 |
+
# Intended to recover from divide-by-zero
|
| 485 |
+
pass
|
| 486 |
+
|
| 487 |
+
return [obj for idx, obj in enumerate(objects) if not suppression[idx]]
|
| 488 |
+
|
| 489 |
+
|
| 490 |
+
def align_headers(headers, rows):
|
| 491 |
+
"""
|
| 492 |
+
Adjust the header boundary to be the convex hull of the rows it intersects
|
| 493 |
+
at least 50% of the height of.
|
| 494 |
+
|
| 495 |
+
For now, we are not supporting tables with multiple headers, so we need to
|
| 496 |
+
eliminate anything besides the top-most header.
|
| 497 |
+
"""
|
| 498 |
+
|
| 499 |
+
aligned_headers = []
|
| 500 |
+
|
| 501 |
+
for row in rows:
|
| 502 |
+
row['header'] = False
|
| 503 |
+
|
| 504 |
+
header_row_nums = []
|
| 505 |
+
for header in headers:
|
| 506 |
+
for row_num, row in enumerate(rows):
|
| 507 |
+
row_height = row['bbox'][3] - row['bbox'][1]
|
| 508 |
+
min_row_overlap = max(row['bbox'][1], header['bbox'][1])
|
| 509 |
+
max_row_overlap = min(row['bbox'][3], header['bbox'][3])
|
| 510 |
+
overlap_height = max_row_overlap - min_row_overlap
|
| 511 |
+
if overlap_height / row_height >= 0.5:
|
| 512 |
+
header_row_nums.append(row_num)
|
| 513 |
+
|
| 514 |
+
if len(header_row_nums) == 0:
|
| 515 |
+
return aligned_headers
|
| 516 |
+
|
| 517 |
+
header_rect = Rect()
|
| 518 |
+
if header_row_nums[0] > 0:
|
| 519 |
+
header_row_nums = list(range(header_row_nums[0]+1)) + header_row_nums
|
| 520 |
+
|
| 521 |
+
last_row_num = -1
|
| 522 |
+
for row_num in header_row_nums:
|
| 523 |
+
if row_num == last_row_num + 1:
|
| 524 |
+
row = rows[row_num]
|
| 525 |
+
row['header'] = True
|
| 526 |
+
header_rect = header_rect.include_rect(row['bbox'])
|
| 527 |
+
last_row_num = row_num
|
| 528 |
+
else:
|
| 529 |
+
# Break as soon as a non-header row is encountered.
|
| 530 |
+
# This ignores any subsequent rows in the table labeled as a header.
|
| 531 |
+
# Having more than 1 header is not supported currently.
|
| 532 |
+
break
|
| 533 |
+
|
| 534 |
+
header = {'bbox': list(header_rect)}
|
| 535 |
+
aligned_headers.append(header)
|
| 536 |
+
|
| 537 |
+
return aligned_headers
|
| 538 |
+
|
| 539 |
+
|
| 540 |
+
def align_supercells(supercells, rows, columns):
|
| 541 |
+
"""
|
| 542 |
+
For each supercell, align it to the rows it intersects 50% of the height of,
|
| 543 |
+
and the columns it intersects 50% of the width of.
|
| 544 |
+
Eliminate supercells for which there are no rows and columns it intersects 50% with.
|
| 545 |
+
"""
|
| 546 |
+
aligned_supercells = []
|
| 547 |
+
|
| 548 |
+
for supercell in supercells:
|
| 549 |
+
supercell['header'] = False
|
| 550 |
+
row_bbox_rect = None
|
| 551 |
+
col_bbox_rect = None
|
| 552 |
+
intersecting_header_rows = set()
|
| 553 |
+
intersecting_data_rows = set()
|
| 554 |
+
for row_num, row in enumerate(rows):
|
| 555 |
+
row_height = row['bbox'][3] - row['bbox'][1]
|
| 556 |
+
supercell_height = supercell['bbox'][3] - supercell['bbox'][1]
|
| 557 |
+
min_row_overlap = max(row['bbox'][1], supercell['bbox'][1])
|
| 558 |
+
max_row_overlap = min(row['bbox'][3], supercell['bbox'][3])
|
| 559 |
+
overlap_height = max_row_overlap - min_row_overlap
|
| 560 |
+
if 'span' in supercell:
|
| 561 |
+
overlap_fraction = max(overlap_height/row_height,
|
| 562 |
+
overlap_height/supercell_height)
|
| 563 |
+
else:
|
| 564 |
+
overlap_fraction = overlap_height / row_height
|
| 565 |
+
if overlap_fraction >= 0.5:
|
| 566 |
+
if 'header' in row and row['header']:
|
| 567 |
+
intersecting_header_rows.add(row_num)
|
| 568 |
+
else:
|
| 569 |
+
intersecting_data_rows.add(row_num)
|
| 570 |
+
|
| 571 |
+
# Supercell cannot span across the header boundary; eliminate whichever
|
| 572 |
+
# group of rows is the smallest
|
| 573 |
+
supercell['header'] = False
|
| 574 |
+
if len(intersecting_data_rows) > 0 and len(intersecting_header_rows) > 0:
|
| 575 |
+
if len(intersecting_data_rows) > len(intersecting_header_rows):
|
| 576 |
+
intersecting_header_rows = set()
|
| 577 |
+
else:
|
| 578 |
+
intersecting_data_rows = set()
|
| 579 |
+
if len(intersecting_header_rows) > 0:
|
| 580 |
+
supercell['header'] = True
|
| 581 |
+
elif 'span' in supercell:
|
| 582 |
+
continue # Require span supercell to be in the header
|
| 583 |
+
intersecting_rows = intersecting_data_rows.union(intersecting_header_rows)
|
| 584 |
+
# Determine vertical span of aligned supercell
|
| 585 |
+
for row_num in intersecting_rows:
|
| 586 |
+
if row_bbox_rect is None:
|
| 587 |
+
row_bbox_rect = Rect(rows[row_num]['bbox'])
|
| 588 |
+
else:
|
| 589 |
+
row_bbox_rect = row_bbox_rect.include_rect(rows[row_num]['bbox'])
|
| 590 |
+
if row_bbox_rect is None:
|
| 591 |
+
continue
|
| 592 |
+
|
| 593 |
+
intersecting_cols = []
|
| 594 |
+
for col_num, col in enumerate(columns):
|
| 595 |
+
col_width = col['bbox'][2] - col['bbox'][0]
|
| 596 |
+
supercell_width = supercell['bbox'][2] - supercell['bbox'][0]
|
| 597 |
+
min_col_overlap = max(col['bbox'][0], supercell['bbox'][0])
|
| 598 |
+
max_col_overlap = min(col['bbox'][2], supercell['bbox'][2])
|
| 599 |
+
overlap_width = max_col_overlap - min_col_overlap
|
| 600 |
+
if 'span' in supercell:
|
| 601 |
+
overlap_fraction = max(overlap_width/col_width,
|
| 602 |
+
overlap_width/supercell_width)
|
| 603 |
+
# Multiply by 2 effectively lowers the threshold to 0.25
|
| 604 |
+
if supercell['header']:
|
| 605 |
+
overlap_fraction = overlap_fraction * 2
|
| 606 |
+
else:
|
| 607 |
+
overlap_fraction = overlap_width / col_width
|
| 608 |
+
if overlap_fraction >= 0.5:
|
| 609 |
+
intersecting_cols.append(col_num)
|
| 610 |
+
if col_bbox_rect is None:
|
| 611 |
+
col_bbox_rect = Rect(col['bbox'])
|
| 612 |
+
else:
|
| 613 |
+
col_bbox_rect = col_bbox_rect.include_rect(col['bbox'])
|
| 614 |
+
if col_bbox_rect is None:
|
| 615 |
+
continue
|
| 616 |
+
|
| 617 |
+
supercell_bbox = list(row_bbox_rect.intersect(col_bbox_rect))
|
| 618 |
+
supercell['bbox'] = supercell_bbox
|
| 619 |
+
|
| 620 |
+
# Only a true supercell if it joins across multiple rows or columns
|
| 621 |
+
if (len(intersecting_rows) > 0 and len(intersecting_cols) > 0
|
| 622 |
+
and (len(intersecting_rows) > 1 or len(intersecting_cols) > 1)):
|
| 623 |
+
supercell['row_numbers'] = list(intersecting_rows)
|
| 624 |
+
supercell['column_numbers'] = intersecting_cols
|
| 625 |
+
aligned_supercells.append(supercell)
|
| 626 |
+
|
| 627 |
+
# A span supercell in the header means there must be supercells above it in the header
|
| 628 |
+
if 'span' in supercell and supercell['header'] and len(supercell['column_numbers']) > 1:
|
| 629 |
+
for row_num in range(0, min(supercell['row_numbers'])):
|
| 630 |
+
new_supercell = {'row_numbers': [row_num], 'column_numbers': supercell['column_numbers'],
|
| 631 |
+
'score': supercell['score'], 'propagated': True}
|
| 632 |
+
new_supercell_columns = [columns[idx] for idx in supercell['column_numbers']]
|
| 633 |
+
new_supercell_rows = [rows[idx] for idx in supercell['row_numbers']]
|
| 634 |
+
bbox = [min([column['bbox'][0] for column in new_supercell_columns]),
|
| 635 |
+
min([row['bbox'][1] for row in new_supercell_rows]),
|
| 636 |
+
max([column['bbox'][2] for column in new_supercell_columns]),
|
| 637 |
+
max([row['bbox'][3] for row in new_supercell_rows])]
|
| 638 |
+
new_supercell['bbox'] = bbox
|
| 639 |
+
aligned_supercells.append(new_supercell)
|
| 640 |
+
|
| 641 |
+
return aligned_supercells
|
| 642 |
+
|
| 643 |
+
|
| 644 |
+
def nms_supercells(supercells):
|
| 645 |
+
"""
|
| 646 |
+
A NMS scheme for supercells that first attempts to shrink supercells to
|
| 647 |
+
resolve overlap.
|
| 648 |
+
If two supercells overlap the same (sub)cell, shrink the lower confidence
|
| 649 |
+
supercell to resolve the overlap. If shrunk supercell is empty, remove it.
|
| 650 |
+
"""
|
| 651 |
+
|
| 652 |
+
supercells = sort_objects_by_score(supercells)
|
| 653 |
+
num_supercells = len(supercells)
|
| 654 |
+
suppression = [False for supercell in supercells]
|
| 655 |
+
|
| 656 |
+
for supercell2_num in range(1, num_supercells):
|
| 657 |
+
supercell2 = supercells[supercell2_num]
|
| 658 |
+
for supercell1_num in range(supercell2_num):
|
| 659 |
+
supercell1 = supercells[supercell1_num]
|
| 660 |
+
remove_supercell_overlap(supercell1, supercell2)
|
| 661 |
+
if ((len(supercell2['row_numbers']) < 2 and len(supercell2['column_numbers']) < 2)
|
| 662 |
+
or len(supercell2['row_numbers']) == 0 or len(supercell2['column_numbers']) == 0):
|
| 663 |
+
suppression[supercell2_num] = True
|
| 664 |
+
|
| 665 |
+
return [obj for idx, obj in enumerate(supercells) if not suppression[idx]]
|
| 666 |
+
|
| 667 |
+
|
| 668 |
+
def header_supercell_tree(supercells):
|
| 669 |
+
"""
|
| 670 |
+
Make sure no supercell in the header is below more than one supercell in any row above it.
|
| 671 |
+
The cells in the header form a tree, but a supercell with more than one supercell in a row
|
| 672 |
+
above it means that some cell has more than one parent, which is not allowed. Eliminate
|
| 673 |
+
any supercell that would cause this to be violated.
|
| 674 |
+
"""
|
| 675 |
+
header_supercells = [supercell for supercell in supercells if 'header' in supercell and supercell['header']]
|
| 676 |
+
header_supercells = sort_objects_by_score(header_supercells)
|
| 677 |
+
|
| 678 |
+
for header_supercell in header_supercells[:]:
|
| 679 |
+
ancestors_by_row = defaultdict(int)
|
| 680 |
+
min_row = min(header_supercell['row_numbers'])
|
| 681 |
+
for header_supercell2 in header_supercells:
|
| 682 |
+
max_row2 = max(header_supercell2['row_numbers'])
|
| 683 |
+
if max_row2 < min_row:
|
| 684 |
+
if (set(header_supercell['column_numbers']).issubset(
|
| 685 |
+
set(header_supercell2['column_numbers']))):
|
| 686 |
+
for row2 in header_supercell2['row_numbers']:
|
| 687 |
+
ancestors_by_row[row2] += 1
|
| 688 |
+
for row in range(0, min_row):
|
| 689 |
+
if not ancestors_by_row[row] == 1:
|
| 690 |
+
supercells.remove(header_supercell)
|
| 691 |
+
break
|
| 692 |
+
|
| 693 |
+
|
| 694 |
+
def table_structure_to_cells(table_structures, table_spans, table_bbox):
|
| 695 |
+
"""
|
| 696 |
+
Assuming the row, column, supercell, and header bounding boxes have
|
| 697 |
+
been refined into a set of consistent table structures, process these
|
| 698 |
+
table structures into table cells. This is a universal representation
|
| 699 |
+
format for the table, which can later be exported to Pandas or CSV formats.
|
| 700 |
+
Classify the cells as header/access cells or data cells
|
| 701 |
+
based on if they intersect with the header bounding box.
|
| 702 |
+
"""
|
| 703 |
+
columns = table_structures['columns']
|
| 704 |
+
rows = table_structures['rows']
|
| 705 |
+
supercells = table_structures['supercells']
|
| 706 |
+
cells = []
|
| 707 |
+
subcells = []
|
| 708 |
+
|
| 709 |
+
# Identify complete cells and subcells
|
| 710 |
+
for column_num, column in enumerate(columns):
|
| 711 |
+
for row_num, row in enumerate(rows):
|
| 712 |
+
column_rect = Rect(list(column['bbox']))
|
| 713 |
+
row_rect = Rect(list(row['bbox']))
|
| 714 |
+
cell_rect = row_rect.intersect(column_rect)
|
| 715 |
+
header = 'header' in row and row['header']
|
| 716 |
+
cell = {'bbox': list(cell_rect), 'column_nums': [column_num], 'row_nums': [row_num],
|
| 717 |
+
'header': header}
|
| 718 |
+
|
| 719 |
+
cell['subcell'] = False
|
| 720 |
+
for supercell in supercells:
|
| 721 |
+
supercell_rect = Rect(list(supercell['bbox']))
|
| 722 |
+
if (supercell_rect.intersect(cell_rect).get_area() # .getArea()
|
| 723 |
+
/ cell_rect.get_area()) > 0.5: # getArea()
|
| 724 |
+
cell['subcell'] = True
|
| 725 |
+
break
|
| 726 |
+
|
| 727 |
+
if cell['subcell']:
|
| 728 |
+
subcells.append(cell)
|
| 729 |
+
else:
|
| 730 |
+
#cell_text = extract_text_inside_bbox(table_spans, cell['bbox'])
|
| 731 |
+
#cell['cell_text'] = cell_text
|
| 732 |
+
cell['subheader'] = False
|
| 733 |
+
cells.append(cell)
|
| 734 |
+
|
| 735 |
+
for supercell in supercells:
|
| 736 |
+
supercell_rect = Rect(list(supercell['bbox']))
|
| 737 |
+
cell_columns = set()
|
| 738 |
+
cell_rows = set()
|
| 739 |
+
cell_rect = None
|
| 740 |
+
header = True
|
| 741 |
+
for subcell in subcells:
|
| 742 |
+
subcell_rect = Rect(list(subcell['bbox']))
|
| 743 |
+
subcell_rect_area = subcell_rect.get_area() # .getArea()
|
| 744 |
+
if (subcell_rect.intersect(supercell_rect).get_area() # .getArea()
|
| 745 |
+
/ subcell_rect_area) > 0.5:
|
| 746 |
+
if cell_rect is None:
|
| 747 |
+
cell_rect = Rect(list(subcell['bbox']))
|
| 748 |
+
else:
|
| 749 |
+
cell_rect.include_rect(Rect(list(subcell['bbox'])))
|
| 750 |
+
cell_rows = cell_rows.union(set(subcell['row_nums']))
|
| 751 |
+
cell_columns = cell_columns.union(set(subcell['column_nums']))
|
| 752 |
+
# By convention here, all subcells must be classified
|
| 753 |
+
# as header cells for a supercell to be classified as a header cell;
|
| 754 |
+
# otherwise, this could lead to a non-rectangular header region
|
| 755 |
+
header = header and 'header' in subcell and subcell['header']
|
| 756 |
+
if len(cell_rows) > 0 and len(cell_columns) > 0:
|
| 757 |
+
cell = {'bbox': list(cell_rect), 'column_nums': list(cell_columns), 'row_nums': list(cell_rows),
|
| 758 |
+
'header': header, 'subheader': supercell['subheader']}
|
| 759 |
+
cells.append(cell)
|
| 760 |
+
|
| 761 |
+
# Compute a confidence score based on how well the page tokens
|
| 762 |
+
# slot into the cells reported by the model
|
| 763 |
+
_, _, cell_match_scores = slot_into_containers(cells, table_spans)
|
| 764 |
+
try:
|
| 765 |
+
mean_match_score = sum(cell_match_scores) / len(cell_match_scores)
|
| 766 |
+
min_match_score = min(cell_match_scores)
|
| 767 |
+
confidence_score = (mean_match_score + min_match_score)/2
|
| 768 |
+
except:
|
| 769 |
+
confidence_score = 0
|
| 770 |
+
|
| 771 |
+
# Dilate rows and columns before final extraction
|
| 772 |
+
#dilated_columns = fill_column_gaps(columns, table_bbox)
|
| 773 |
+
dilated_columns = columns
|
| 774 |
+
#dilated_rows = fill_row_gaps(rows, table_bbox)
|
| 775 |
+
dilated_rows = rows
|
| 776 |
+
for cell in cells:
|
| 777 |
+
column_rect = Rect()
|
| 778 |
+
for column_num in cell['column_nums']:
|
| 779 |
+
column_rect.include_rect(list(dilated_columns[column_num]['bbox']))
|
| 780 |
+
row_rect = Rect()
|
| 781 |
+
for row_num in cell['row_nums']:
|
| 782 |
+
row_rect.include_rect(list(dilated_rows[row_num]['bbox']))
|
| 783 |
+
cell_rect = column_rect.intersect(row_rect)
|
| 784 |
+
cell['bbox'] = list(cell_rect)
|
| 785 |
+
|
| 786 |
+
span_nums_by_cell, _, _ = slot_into_containers(cells, table_spans, overlap_threshold=0.001,
|
| 787 |
+
unique_assignment=True, forced_assignment=False)
|
| 788 |
+
|
| 789 |
+
for cell, cell_span_nums in zip(cells, span_nums_by_cell):
|
| 790 |
+
cell_spans = [table_spans[num] for num in cell_span_nums]
|
| 791 |
+
# TODO: Refine how text is extracted; should be character-based, not span-based;
|
| 792 |
+
# but need to associate
|
| 793 |
+
# cell['cell_text'] = extract_text_from_spans(cell_spans, remove_integer_superscripts=False) # TODO
|
| 794 |
+
cell['spans'] = cell_spans
|
| 795 |
+
|
| 796 |
+
# Adjust the row, column, and cell bounding boxes to reflect the extracted text
|
| 797 |
+
num_rows = len(rows)
|
| 798 |
+
rows = sort_objects_top_to_bottom(rows)
|
| 799 |
+
num_columns = len(columns)
|
| 800 |
+
columns = sort_objects_left_to_right(columns)
|
| 801 |
+
min_y_values_by_row = defaultdict(list)
|
| 802 |
+
max_y_values_by_row = defaultdict(list)
|
| 803 |
+
min_x_values_by_column = defaultdict(list)
|
| 804 |
+
max_x_values_by_column = defaultdict(list)
|
| 805 |
+
for cell in cells:
|
| 806 |
+
min_row = min(cell["row_nums"])
|
| 807 |
+
max_row = max(cell["row_nums"])
|
| 808 |
+
min_column = min(cell["column_nums"])
|
| 809 |
+
max_column = max(cell["column_nums"])
|
| 810 |
+
for span in cell['spans']:
|
| 811 |
+
min_x_values_by_column[min_column].append(span['bbox'][0])
|
| 812 |
+
min_y_values_by_row[min_row].append(span['bbox'][1])
|
| 813 |
+
max_x_values_by_column[max_column].append(span['bbox'][2])
|
| 814 |
+
max_y_values_by_row[max_row].append(span['bbox'][3])
|
| 815 |
+
for row_num, row in enumerate(rows):
|
| 816 |
+
if len(min_x_values_by_column[0]) > 0:
|
| 817 |
+
row['bbox'][0] = min(min_x_values_by_column[0])
|
| 818 |
+
if len(min_y_values_by_row[row_num]) > 0:
|
| 819 |
+
row['bbox'][1] = min(min_y_values_by_row[row_num])
|
| 820 |
+
if len(max_x_values_by_column[num_columns-1]) > 0:
|
| 821 |
+
row['bbox'][2] = max(max_x_values_by_column[num_columns-1])
|
| 822 |
+
if len(max_y_values_by_row[row_num]) > 0:
|
| 823 |
+
row['bbox'][3] = max(max_y_values_by_row[row_num])
|
| 824 |
+
for column_num, column in enumerate(columns):
|
| 825 |
+
if len(min_x_values_by_column[column_num]) > 0:
|
| 826 |
+
column['bbox'][0] = min(min_x_values_by_column[column_num])
|
| 827 |
+
if len(min_y_values_by_row[0]) > 0:
|
| 828 |
+
column['bbox'][1] = min(min_y_values_by_row[0])
|
| 829 |
+
if len(max_x_values_by_column[column_num]) > 0:
|
| 830 |
+
column['bbox'][2] = max(max_x_values_by_column[column_num])
|
| 831 |
+
if len(max_y_values_by_row[num_rows-1]) > 0:
|
| 832 |
+
column['bbox'][3] = max(max_y_values_by_row[num_rows-1])
|
| 833 |
+
for cell in cells:
|
| 834 |
+
row_rect = Rect()
|
| 835 |
+
column_rect = Rect()
|
| 836 |
+
for row_num in cell['row_nums']:
|
| 837 |
+
row_rect.include_rect(list(rows[row_num]['bbox']))
|
| 838 |
+
for column_num in cell['column_nums']:
|
| 839 |
+
column_rect.include_rect(list(columns[column_num]['bbox']))
|
| 840 |
+
cell_rect = row_rect.intersect(column_rect)
|
| 841 |
+
if cell_rect.get_area() > 0: # getArea()
|
| 842 |
+
cell['bbox'] = list(cell_rect)
|
| 843 |
+
pass
|
| 844 |
+
|
| 845 |
+
return cells, confidence_score
|
| 846 |
+
|
| 847 |
+
|
| 848 |
+
def remove_supercell_overlap(supercell1, supercell2):
|
| 849 |
+
"""
|
| 850 |
+
This function resolves overlap between supercells (supercells must be
|
| 851 |
+
disjoint) by iteratively shrinking supercells by the fewest grid cells
|
| 852 |
+
necessary to resolve the overlap.
|
| 853 |
+
Example:
|
| 854 |
+
If two supercells overlap at grid cell (R, C), and supercell #1 is less
|
| 855 |
+
confident than supercell #2, we eliminate either row R from supercell #1
|
| 856 |
+
or column C from supercell #1 by comparing the number of columns in row R
|
| 857 |
+
versus the number of rows in column C. If the number of columns in row R
|
| 858 |
+
is less than the number of rows in column C, we eliminate row R from
|
| 859 |
+
supercell #1. This resolves the overlap by removing fewer grid cells from
|
| 860 |
+
supercell #1 than if we eliminated column C from it.
|
| 861 |
+
"""
|
| 862 |
+
common_rows = set(supercell1['row_numbers']).intersection(set(supercell2['row_numbers']))
|
| 863 |
+
common_columns = set(supercell1['column_numbers']).intersection(set(supercell2['column_numbers']))
|
| 864 |
+
|
| 865 |
+
# While the supercells have overlapping grid cells, continue shrinking the less-confident
|
| 866 |
+
# supercell one row or one column at a time
|
| 867 |
+
while len(common_rows) > 0 and len(common_columns) > 0:
|
| 868 |
+
# Try to shrink the supercell as little as possible to remove the overlap;
|
| 869 |
+
# if the supercell has fewer rows than columns, remove an overlapping column,
|
| 870 |
+
# because this removes fewer grid cells from the supercell;
|
| 871 |
+
# otherwise remove an overlapping row
|
| 872 |
+
if len(supercell2['row_numbers']) < len(supercell2['column_numbers']):
|
| 873 |
+
min_column = min(supercell2['column_numbers'])
|
| 874 |
+
max_column = max(supercell2['column_numbers'])
|
| 875 |
+
if max_column in common_columns:
|
| 876 |
+
common_columns.remove(max_column)
|
| 877 |
+
supercell2['column_numbers'].remove(max_column)
|
| 878 |
+
elif min_column in common_columns:
|
| 879 |
+
common_columns.remove(min_column)
|
| 880 |
+
supercell2['column_numbers'].remove(min_column)
|
| 881 |
+
else:
|
| 882 |
+
supercell2['column_numbers'] = []
|
| 883 |
+
common_columns = set()
|
| 884 |
+
else:
|
| 885 |
+
min_row = min(supercell2['row_numbers'])
|
| 886 |
+
max_row = max(supercell2['row_numbers'])
|
| 887 |
+
if max_row in common_rows:
|
| 888 |
+
common_rows.remove(max_row)
|
| 889 |
+
supercell2['row_numbers'].remove(max_row)
|
| 890 |
+
elif min_row in common_rows:
|
| 891 |
+
common_rows.remove(min_row)
|
| 892 |
+
supercell2['row_numbers'].remove(min_row)
|
| 893 |
+
else:
|
| 894 |
+
supercell2['row_numbers'] = []
|
| 895 |
+
common_rows = set()
|
requirements.txt
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
-e git+https://github.com/mindee/doctr.git#egg=python-doctr[tf]
|
| 2 |
+
streamlit>=0.65.0
|
| 3 |
+
PyMuPDF>=1.16.0,!=1.18.11,!=1.18.12,!=1.19.5
|
| 4 |
+
tf2onnx==1.13.0
|
| 5 |
+
Pillow==9.0.1
|
| 6 |
+
pytesseract==0.3.10
|
| 7 |
+
torch==1.12.0
|
| 8 |
+
torchvision==0.13.0
|
| 9 |
+
numpy==1.21.6
|
tessdata/eng.traineddata
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:8280aed0782fe27257a68ea10fe7ef324ca0f8d85bd2fd145d1c2b560bcb66ba
|
| 3 |
+
size 15400601
|
weights/structure_wts.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:46121ab2f4aba48a7d38624c861658ffeaacd0f305e95efcf66cb017e588b700
|
| 3 |
+
size 14371957
|