Spaces:
Sleeping
Sleeping
Joshnicholas
commited on
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,296 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
### Stolen from https://huggingface.co/spaces/pierreguillou/tatr-demo/blob/main/app.py
|
2 |
+
|
3 |
+
import matplotlib.pyplot as plt
|
4 |
+
import matplotlib.patches as patches
|
5 |
+
from matplotlib.patches import Patch
|
6 |
+
import io
|
7 |
+
from PIL import Image, ImageDraw
|
8 |
+
import numpy as np
|
9 |
+
import csv
|
10 |
+
import pandas as pd
|
11 |
+
|
12 |
+
from torchvision import transforms
|
13 |
+
|
14 |
+
from transformers import AutoModelForObjectDetection
|
15 |
+
import torch
|
16 |
+
|
17 |
+
import easyocr
|
18 |
+
|
19 |
+
import gradio as gr
|
20 |
+
|
21 |
+
|
22 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
23 |
+
|
24 |
+
|
25 |
+
class MaxResize(object):
|
26 |
+
def __init__(self, max_size=800):
|
27 |
+
self.max_size = max_size
|
28 |
+
|
29 |
+
def __call__(self, image):
|
30 |
+
width, height = image.size
|
31 |
+
current_max_size = max(width, height)
|
32 |
+
scale = self.max_size / current_max_size
|
33 |
+
resized_image = image.resize((int(round(scale*width)), int(round(scale*height))))
|
34 |
+
|
35 |
+
return resized_image
|
36 |
+
|
37 |
+
detection_transform = transforms.Compose([
|
38 |
+
MaxResize(800),
|
39 |
+
transforms.ToTensor(),
|
40 |
+
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
|
41 |
+
])
|
42 |
+
|
43 |
+
structure_transform = transforms.Compose([
|
44 |
+
MaxResize(1000),
|
45 |
+
transforms.ToTensor(),
|
46 |
+
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
|
47 |
+
])
|
48 |
+
|
49 |
+
# load table detection model
|
50 |
+
# processor = TableTransformerImageProcessor(max_size=800)
|
51 |
+
model = AutoModelForObjectDetection.from_pretrained("microsoft/table-transformer-detection", revision="no_timm").to(device)
|
52 |
+
|
53 |
+
# load table structure recognition model
|
54 |
+
# structure_processor = TableTransformerImageProcessor(max_size=1000)
|
55 |
+
structure_model = AutoModelForObjectDetection.from_pretrained("microsoft/table-transformer-structure-recognition-v1.1-all").to(device)
|
56 |
+
|
57 |
+
# load EasyOCR reader
|
58 |
+
reader = easyocr.Reader(['en'])
|
59 |
+
|
60 |
+
|
61 |
+
# for output bounding box post-processing
|
62 |
+
def box_cxcywh_to_xyxy(x):
|
63 |
+
x_c, y_c, w, h = x.unbind(-1)
|
64 |
+
b = [(x_c - 0.5 * w), (y_c - 0.5 * h), (x_c + 0.5 * w), (y_c + 0.5 * h)]
|
65 |
+
return torch.stack(b, dim=1)
|
66 |
+
|
67 |
+
|
68 |
+
def rescale_bboxes(out_bbox, size):
|
69 |
+
width, height = size
|
70 |
+
boxes = box_cxcywh_to_xyxy(out_bbox)
|
71 |
+
boxes = boxes * torch.tensor([width, height, width, height], dtype=torch.float32)
|
72 |
+
return boxes
|
73 |
+
|
74 |
+
|
75 |
+
def outputs_to_objects(outputs, img_size, id2label):
|
76 |
+
m = outputs.logits.softmax(-1).max(-1)
|
77 |
+
pred_labels = list(m.indices.detach().cpu().numpy())[0]
|
78 |
+
pred_scores = list(m.values.detach().cpu().numpy())[0]
|
79 |
+
pred_bboxes = outputs['pred_boxes'].detach().cpu()[0]
|
80 |
+
pred_bboxes = [elem.tolist() for elem in rescale_bboxes(pred_bboxes, img_size)]
|
81 |
+
|
82 |
+
objects = []
|
83 |
+
for label, score, bbox in zip(pred_labels, pred_scores, pred_bboxes):
|
84 |
+
class_label = id2label[int(label)]
|
85 |
+
if not class_label == 'no object':
|
86 |
+
objects.append({'label': class_label, 'score': float(score),
|
87 |
+
'bbox': [float(elem) for elem in bbox]})
|
88 |
+
|
89 |
+
return objects
|
90 |
+
|
91 |
+
|
92 |
+
def fig2img(fig):
|
93 |
+
"""Convert a Matplotlib figure to a PIL Image and return it"""
|
94 |
+
buf = io.BytesIO()
|
95 |
+
fig.savefig(buf)
|
96 |
+
buf.seek(0)
|
97 |
+
image = Image.open(buf)
|
98 |
+
return image
|
99 |
+
|
100 |
+
|
101 |
+
def visualize_detected_tables(img, det_tables):
|
102 |
+
plt.imshow(img, interpolation="lanczos")
|
103 |
+
fig = plt.gcf()
|
104 |
+
fig.set_size_inches(20, 20)
|
105 |
+
ax = plt.gca()
|
106 |
+
|
107 |
+
for det_table in det_tables:
|
108 |
+
bbox = det_table['bbox']
|
109 |
+
|
110 |
+
if det_table['label'] == 'table':
|
111 |
+
facecolor = (1, 0, 0.45)
|
112 |
+
edgecolor = (1, 0, 0.45)
|
113 |
+
alpha = 0.3
|
114 |
+
linewidth = 2
|
115 |
+
hatch='//////'
|
116 |
+
elif det_table['label'] == 'table rotated':
|
117 |
+
facecolor = (0.95, 0.6, 0.1)
|
118 |
+
edgecolor = (0.95, 0.6, 0.1)
|
119 |
+
alpha = 0.3
|
120 |
+
linewidth = 2
|
121 |
+
hatch='//////'
|
122 |
+
else:
|
123 |
+
continue
|
124 |
+
|
125 |
+
rect = patches.Rectangle(bbox[:2], bbox[2]-bbox[0], bbox[3]-bbox[1], linewidth=linewidth,
|
126 |
+
edgecolor='none',facecolor=facecolor, alpha=0.1)
|
127 |
+
ax.add_patch(rect)
|
128 |
+
rect = patches.Rectangle(bbox[:2], bbox[2]-bbox[0], bbox[3]-bbox[1], linewidth=linewidth,
|
129 |
+
edgecolor=edgecolor,facecolor='none',linestyle='-', alpha=alpha)
|
130 |
+
ax.add_patch(rect)
|
131 |
+
rect = patches.Rectangle(bbox[:2], bbox[2]-bbox[0], bbox[3]-bbox[1], linewidth=0,
|
132 |
+
edgecolor=edgecolor,facecolor='none',linestyle='-', hatch=hatch, alpha=0.2)
|
133 |
+
ax.add_patch(rect)
|
134 |
+
|
135 |
+
plt.xticks([], [])
|
136 |
+
plt.yticks([], [])
|
137 |
+
|
138 |
+
legend_elements = [Patch(facecolor=(1, 0, 0.45), edgecolor=(1, 0, 0.45),
|
139 |
+
label='Table', hatch='//////', alpha=0.3),
|
140 |
+
Patch(facecolor=(0.95, 0.6, 0.1), edgecolor=(0.95, 0.6, 0.1),
|
141 |
+
label='Table (rotated)', hatch='//////', alpha=0.3)]
|
142 |
+
plt.legend(handles=legend_elements, bbox_to_anchor=(0.5, -0.02), loc='upper center', borderaxespad=0,
|
143 |
+
fontsize=10, ncol=2)
|
144 |
+
plt.gcf().set_size_inches(10, 10)
|
145 |
+
plt.axis('off')
|
146 |
+
|
147 |
+
return fig
|
148 |
+
|
149 |
+
|
150 |
+
def detect_and_crop_table(image):
|
151 |
+
# prepare image for the model
|
152 |
+
# pixel_values = processor(image, return_tensors="pt").pixel_values
|
153 |
+
pixel_values = detection_transform(image).unsqueeze(0).to(device)
|
154 |
+
|
155 |
+
# forward pass
|
156 |
+
with torch.no_grad():
|
157 |
+
outputs = model(pixel_values)
|
158 |
+
|
159 |
+
# postprocess to get detected tables
|
160 |
+
id2label = model.config.id2label
|
161 |
+
id2label[len(model.config.id2label)] = "no object"
|
162 |
+
detected_tables = outputs_to_objects(outputs, image.size, id2label)
|
163 |
+
|
164 |
+
# visualize
|
165 |
+
# fig = visualize_detected_tables(image, detected_tables)
|
166 |
+
# image = fig2img(fig)
|
167 |
+
|
168 |
+
# crop first detected table out of image
|
169 |
+
cropped_table = image.crop(detected_tables[0]["bbox"])
|
170 |
+
|
171 |
+
return cropped_table
|
172 |
+
|
173 |
+
|
174 |
+
def recognize_table(image):
|
175 |
+
# prepare image for the model
|
176 |
+
# pixel_values = structure_processor(images=image, return_tensors="pt").pixel_values
|
177 |
+
pixel_values = structure_transform(image).unsqueeze(0).to(device)
|
178 |
+
|
179 |
+
# forward pass
|
180 |
+
with torch.no_grad():
|
181 |
+
outputs = structure_model(pixel_values)
|
182 |
+
|
183 |
+
# postprocess to get individual elements
|
184 |
+
id2label = structure_model.config.id2label
|
185 |
+
id2label[len(structure_model.config.id2label)] = "no object"
|
186 |
+
cells = outputs_to_objects(outputs, image.size, id2label)
|
187 |
+
|
188 |
+
# visualize cells on cropped table
|
189 |
+
draw = ImageDraw.Draw(image)
|
190 |
+
|
191 |
+
for cell in cells:
|
192 |
+
draw.rectangle(cell["bbox"], outline="red")
|
193 |
+
|
194 |
+
return image, cells
|
195 |
+
|
196 |
+
|
197 |
+
def get_cell_coordinates_by_row(table_data):
|
198 |
+
# Extract rows and columns
|
199 |
+
rows = [entry for entry in table_data if entry['label'] == 'table row']
|
200 |
+
columns = [entry for entry in table_data if entry['label'] == 'table column']
|
201 |
+
|
202 |
+
# Sort rows and columns by their Y and X coordinates, respectively
|
203 |
+
rows.sort(key=lambda x: x['bbox'][1])
|
204 |
+
columns.sort(key=lambda x: x['bbox'][0])
|
205 |
+
|
206 |
+
# Function to find cell coordinates
|
207 |
+
def find_cell_coordinates(row, column):
|
208 |
+
cell_bbox = [column['bbox'][0], row['bbox'][1], column['bbox'][2], row['bbox'][3]]
|
209 |
+
return cell_bbox
|
210 |
+
|
211 |
+
# Generate cell coordinates and count cells in each row
|
212 |
+
cell_coordinates = []
|
213 |
+
|
214 |
+
for row in rows:
|
215 |
+
row_cells = []
|
216 |
+
for column in columns:
|
217 |
+
cell_bbox = find_cell_coordinates(row, column)
|
218 |
+
row_cells.append({'column': column['bbox'], 'cell': cell_bbox})
|
219 |
+
|
220 |
+
# Sort cells in the row by X coordinate
|
221 |
+
row_cells.sort(key=lambda x: x['column'][0])
|
222 |
+
|
223 |
+
# Append row information to cell_coordinates
|
224 |
+
cell_coordinates.append({'row': row['bbox'], 'cells': row_cells, 'cell_count': len(row_cells)})
|
225 |
+
|
226 |
+
# Sort rows from top to bottom
|
227 |
+
cell_coordinates.sort(key=lambda x: x['row'][1])
|
228 |
+
|
229 |
+
return cell_coordinates
|
230 |
+
|
231 |
+
|
232 |
+
def apply_ocr(cell_coordinates, cropped_table):
|
233 |
+
# let's OCR row by row
|
234 |
+
data = dict()
|
235 |
+
max_num_columns = 0
|
236 |
+
for idx, row in enumerate(cell_coordinates):
|
237 |
+
row_text = []
|
238 |
+
for cell in row["cells"]:
|
239 |
+
# crop cell out of image
|
240 |
+
cell_image = np.array(cropped_table.crop(cell["cell"]))
|
241 |
+
# apply OCR
|
242 |
+
result = reader.readtext(np.array(cell_image))
|
243 |
+
if len(result) > 0:
|
244 |
+
text = " ".join([x[1] for x in result])
|
245 |
+
row_text.append(text)
|
246 |
+
|
247 |
+
if len(row_text) > max_num_columns:
|
248 |
+
max_num_columns = len(row_text)
|
249 |
+
|
250 |
+
data[str(idx)] = row_text
|
251 |
+
|
252 |
+
# pad rows which don't have max_num_columns elements
|
253 |
+
# to make sure all rows have the same number of columns
|
254 |
+
for idx, row_data in data.copy().items():
|
255 |
+
if len(row_data) != max_num_columns:
|
256 |
+
row_data = row_data + ["" for _ in range(max_num_columns - len(row_data))]
|
257 |
+
data[str(idx)] = row_data
|
258 |
+
|
259 |
+
# write to csv
|
260 |
+
with open('output.csv','w') as result_file:
|
261 |
+
wr = csv.writer(result_file, dialect='excel')
|
262 |
+
|
263 |
+
for row, row_text in data.items():
|
264 |
+
wr.writerow(row_text)
|
265 |
+
|
266 |
+
# return as Pandas dataframe
|
267 |
+
df = pd.read_csv('output.csv')
|
268 |
+
|
269 |
+
return df, data
|
270 |
+
|
271 |
+
|
272 |
+
def process_pdf(image):
|
273 |
+
cropped_table = detect_and_crop_table(image)
|
274 |
+
|
275 |
+
image, cells = recognize_table(cropped_table)
|
276 |
+
|
277 |
+
cell_coordinates = get_cell_coordinates_by_row(cells)
|
278 |
+
|
279 |
+
df, data = apply_ocr(cell_coordinates, image)
|
280 |
+
|
281 |
+
return image, df, data
|
282 |
+
|
283 |
+
|
284 |
+
title = "Demo: table detection & recognition with Table Transformer (TATR)."
|
285 |
+
description = """Demo for table extraction with the Table Transformer. First, table detection is performed on the input image using https://huggingface.co/microsoft/table-transformer-detection,
|
286 |
+
after which the detected table is extracted and https://huggingface.co/microsoft/table-transformer-structure-recognition-v1.1-all is leveraged to recognize the individual rows, columns and cells. OCR is then performed per cell, row by row."""
|
287 |
+
examples = [['image.png'], ['mistral_paper.png']]
|
288 |
+
|
289 |
+
app = gr.Interface(fn=process_pdf,
|
290 |
+
inputs=gr.Image(type="pil"),
|
291 |
+
outputs=[gr.Image(type="pil", label="Detected table"), gr.Dataframe(label="Table as CSV"), gr.JSON(label="Data as JSON")],
|
292 |
+
title=title,
|
293 |
+
description=description,
|
294 |
+
examples=examples)
|
295 |
+
app.queue()
|
296 |
+
app.launch(debug=True)
|