sxandie's picture
Update app.py
e43f981 verified
import os
os.system('pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cpu')
import glob, fitz
import PIL
import re
import torch
import cv2
import pytesseract
import pandas as pd
import numpy as np
import gradio as gr
from PIL import Image
from tqdm import tqdm
from difflib import SequenceMatcher
from itertools import groupby
from scipy import ndimage
from scipy.ndimage import interpolation as inter
from datasets import load_metric
from datasets import load_dataset
from datasets.features import ClassLabel
from transformers import AutoProcessor
from PIL import Image, ImageDraw, ImageFont
from transformers import AutoModelForTokenClassification
from transformers.data.data_collator import default_data_collator
from datasets import Features, Sequence, ClassLabel, Value, Array2D, Array3D
from transformers import LayoutLMv3ForTokenClassification,LayoutLMv3FeatureExtractor,LayoutLMv3ImageProcessor
import io
# import paddleocr
# from paddleocr import PaddleOCR
auth_token = os.environ.get("HUGGING_FACE_HUB_TOKEN")
import warnings
# Ignore warning messages
warnings.filterwarnings("ignore")
id2label= {0: 'others', 1: 'issuer_name', 2: 'issuer_addr', 3: 'issuer_cap', 4: 'issuer_city', 5: 'issuer_prov', 6: 'issuer_state', 7: 'issuer_tel', 8: 'issuer_id', 9: 'issuer_fax', 10: 'issuer_vat', 11: 'issuer_contact', 12: 'issuer_contact_email', 13: 'issuer_contact_phone', 14: 'receiver_name', 15: 'receiver_addr', 16: 'receiver_cap', 17: 'receiver_city', 18: 'receiver_prov', 19: 'receiver_state', 20: 'receiver_tel', 21: 'receiver_fax', 22: 'receiver_vat', 23: 'receiver_id', 24: 'receiver_contact', 25: 'dest_name', 26: 'dest_addr', 27: 'dest_cap', 28: 'dest_city', 29: 'dest_prov', 30: 'dest_state', 31: 'dest_tel', 32: 'dest_fax', 33: 'dest_vat', 34: 'doc_type', 35: 'doc_nr', 36: 'doc_date', 37: 'order_nr', 38: 'order_date', 39: 'service_order', 40: 'shipment_nr', 41: 'client_reference', 42: 'client_vat', 43: 'client_id', 44: 'client_code', 45: 'time', 46: 'notes', 47: 'client_tel', 48: 'art_code', 49: 'ref_code', 50: 'order_reason', 51: 'order_ref', 52: 'order_ref_date', 53: 'detail_desc', 54: 'lot_id', 55: 'lot_qty', 56: 'detail_um', 57: 'detail_qty', 58: 'detail_tare', 59: 'detail_grossw', 60: 'detail_packages', 61: 'detail_netw', 62: 'detail_origin', 63: 'payment_bank', 64: 'payment_terms', 65: 'tot_qty', 66: 'tot_grossw', 67: 'tot_netw', 68: 'tot_volume', 69: 'shipment_reason', 70: 'package_type', 71: 'transport_respons', 72: 'transport_vectors', 73: 'transport_terms', 74: 'transport_datetime', 75: 'return_plt', 76: 'nonreturn_plt', 77: 'dest_signature', 78: 'driver_signature', 79: 'transport_signature', 80: 'page', 81: 'varieta', 82: 'raccolta', 83: 'detail_volume'}
custom_config = r'--oem 3 --psm 6'
lang='eng'
#Google Vision OCR
from google.cloud import vision_v1p3beta1 as vision
from google.cloud import vision_v1p3beta1 as vision
from google.cloud import vision
# os.environ["GOOGLE_APPLICATION_CREDENTIALS"] = "test-apikey.json"
processor = AutoProcessor.from_pretrained("microsoft/layoutlmv3-base", apply_ocr=False)
model = AutoModelForTokenClassification.from_pretrained("sxandie/doc-ai-information-extraction",use_auth_token=auth_token)
from tabulate import tabulate
def print_df(df):
print(tabulate(df, headers = df.columns, tablefmt = 'psql'))
def process_image_pytesseract(image,width,height):
width, height = image.size
feature_extractor = LayoutLMv3ImageProcessor(apply_ocr=True,lang=lang)
encoding_feature_extractor = feature_extractor(image, return_tensors="pt",truncation=True)
words, boxes = encoding_feature_extractor.words, encoding_feature_extractor.boxes
return words,boxes
def create_bounding_box5(vertices, width_scale, height_scale):
# Get the x, y coordinates
x1 = int(vertices[0].x * width_scale)
y1 = int(vertices[0].y * height_scale)
x2 = int(vertices[2].x * width_scale)
y2 = int(vertices[2].y * height_scale)
# Validate x1 < x2
if x1 > x2:
x1, x2 = x2, x1
# Validate y1 < y2
if y1 > y2:
y1, y2 = y2, y1
# Return valid bounding box
return [x1, y1, x2, y2]
#Google Vision OCR
def process_image_GoogleVision(image, width, height):
inference_image = [image.convert("RGB")]
client = vision.ImageAnnotatorClient()
with io.BytesIO() as output:
image.save(output, format='JPEG')
content = output.getvalue()
image = vision.Image(content=content)
response = client.text_detection(image=image)
texts = response.text_annotations
# Get the bounding box vertices and remove the first item
bboxes = [text.bounding_poly.vertices[1:] for text in texts]
# Create the list of words and boxes
words = [text.description for text in texts]
boxes = [create_bounding_box5(bbox, 1000/width, 1000/height) for bbox in bboxes]
return words,boxes
def generate_unique_colors(id2label):
# Generate unique colors
label_ints = np.random.choice(len(PIL.ImageColor.colormap), len(id2label), replace=False)
label_color_pil = list(PIL.ImageColor.colormap.values())
label_color = [label_color_pil[i] for i in label_ints]
color = {}
for k, v in id2label.items():
if v[:2] == '':
color['o'] = label_color[k]
else:
color[v[0:]] = label_color[k]
return color
def create_bounding_box1(bbox_data, width_scale: float, height_scale: float):
xs = []
ys = []
for x, y in bbox_data:
xs.append(x)
ys.append(y)
left = int(max(0, min(xs) * width_scale))
top = int(max(0, min(ys) * height_scale))
right = int(min(1000, max(xs) * width_scale))
bottom = int(min(1000, max(ys) * height_scale))
return [left, top, right, bottom]
def unnormalize_box(bbox, width, height):
return [
width * (bbox[0] / 1000),
height * (bbox[1] / 1000),
width * (bbox[2] / 1000),
height * (bbox[3] / 1000),
]
def iob_to_label(label):
return id2label.get(label, 'others')
def process_image(image):
custom_config = r'--oem 3 --psm 6'
# lang='eng+deu+ita+chi_sim'
lang='eng'
width, height = image.size
feature_extractor = LayoutLMv3FeatureExtractor(apply_ocr=True)
encoding_feature_extractor = feature_extractor(image, return_tensors="pt",truncation=True)
words, boxes = encoding_feature_extractor.words, encoding_feature_extractor.boxes
custom_config = r'--oem 3 --psm 6'
# encode
inference_image = [image.convert("RGB")]
encoding = processor(inference_image , truncation=True, return_offsets_mapping=True, return_tensors="pt", padding="max_length", stride =128, max_length=512, return_overflowing_tokens=True)
offset_mapping = encoding.pop('offset_mapping')
overflow_to_sample_mapping = encoding.pop('overflow_to_sample_mapping')
# change the shape of pixel values
x = []
for i in range(0, len(encoding['pixel_values'])):
x.append(encoding['pixel_values'][i])
x = torch.stack(x)
encoding['pixel_values'] = x
# forward pass
outputs = model(**encoding)
# get predictions
predictions = outputs.logits.argmax(-1).squeeze().tolist()
token_boxes = encoding.bbox.squeeze().tolist()
# only keep non-subword predictions
preds = []
l_words = []
bboxes = []
token_section_num = []
if (len(token_boxes) == 512):
predictions = [predictions]
token_boxes = [token_boxes]
for i in range(0, len(token_boxes)):
for j in range(0, len(token_boxes[i])):
unnormal_box = unnormalize_box(token_boxes[i][j], width, height)
if (np.asarray(token_boxes[i][j]).shape != (4,)):
continue
elif (token_boxes[i][j] == [0, 0, 0, 0] or token_boxes[i][j] == 0):
#print('zero found!')
continue
# if bbox is available in the list, just we need to update text
elif (unnormal_box not in bboxes):
preds.append(predictions[i][j])
l_words.append(processor.tokenizer.decode(encoding["input_ids"][i][j]))
bboxes.append(unnormal_box)
token_section_num.append(i)
else:
# we have to update the word
_index = bboxes.index(unnormal_box)
if (token_section_num[_index] == i):
# check if they're in a same section or not (documents with more than 512 tokens will divide to seperate
# parts, so it's possible to have a word in both of the pages and we have to control that repetetive words
# HERE: because they're in a same section, so we can merge them safely
l_words[_index] = l_words[_index] + processor.tokenizer.decode(encoding["input_ids"][i][j])
else:
continue
return bboxes, preds, l_words, image
def process_image_encoding(model, processor, image, words, boxes,width,height):
# encode
inference_image = [image.convert("RGB")]
encoding = processor(inference_image ,words,boxes=boxes, truncation=True, return_offsets_mapping=True, return_tensors="pt",
padding="max_length", stride =128, max_length=512, return_overflowing_tokens=True)
offset_mapping = encoding.pop('offset_mapping')
overflow_to_sample_mapping = encoding.pop('overflow_to_sample_mapping')
# change the shape of pixel values
x = []
for i in range(0, len(encoding['pixel_values'])):
x.append(encoding['pixel_values'][i])
x = torch.stack(x)
encoding['pixel_values'] = x
# forward pass
outputs = model(**encoding)
# get predictions
predictions = outputs.logits.argmax(-1).squeeze().tolist()
token_boxes = encoding.bbox.squeeze().tolist()
# only keep non-subword predictions
preds = []
l_words = []
bboxes = []
token_section_num = []
if (len(token_boxes) == 512):
predictions = [predictions]
token_boxes = [token_boxes]
for i in range(0, len(token_boxes)):
for j in range(0, len(token_boxes[i])):
unnormal_box = unnormalize_box(token_boxes[i][j], width, height)
if (np.asarray(token_boxes[i][j]).shape != (4,)):
continue
elif (token_boxes[i][j] == [0, 0, 0, 0] or token_boxes[i][j] == 0):
#print('zero found!')
continue
# if bbox is available in the list, just we need to update text
elif (unnormal_box not in bboxes):
preds.append(predictions[i][j])
l_words.append(processor.tokenizer.decode(encoding["input_ids"][i][j]))
bboxes.append(unnormal_box)
token_section_num.append(i)
else:
# we have to update the word
_index = bboxes.index(unnormal_box)
if (token_section_num[_index] == i):
# check if they're in a same section or not (documents with more than 512 tokens will divide to seperate
# parts, so it's possible to have a word in both of the pages and we have to control that repetetive words
# HERE: because they're in a same section, so we can merge them safely
l_words[_index] = l_words[_index] + processor.tokenizer.decode(encoding["input_ids"][i][j])
else:
continue
return bboxes, preds, l_words, image
def process_form_(json_df):
labels = [x['LABEL'] for x in json_df]
texts = [x['TEXT'] for x in json_df]
cmb_list = []
for i, j in enumerate(labels):
cmb_list.append([labels[i], texts[i]])
grouper = lambda l: [[k] + sum((v[1::] for v in vs), []) for k, vs in groupby(l, lambda x: x[0])]
list_final = grouper(cmb_list)
lst_final = []
for x in list_final:
json_dict = {}
json_dict[x[0]] = (' ').join(x[1:])
lst_final.append(json_dict)
return lst_final
def createExcel(maindf, detailsdf, pdffile):
outputPath = f'{pdffile}.xlsx'
with pd.ExcelWriter(outputPath, engine='xlsxwriter') as writer:
maindf.to_excel(writer, sheet_name='headers', index=False)
detailsdf.to_excel(writer, sheet_name='details', index=False)
worksheet1 = writer.sheets["headers"]
for idx, col in enumerate(maindf):
series = maindf[col]
max_len = max((
series.astype(str).map(len).max(),
len(str(series.name))
)) + 1
worksheet1.set_column(idx, idx, max_len)
worksheet2 = writer.sheets["details"]
for idx, col in enumerate(detailsdf):
series = detailsdf[col]
max_len = max((
series.astype(str).map(len).max(),
len(str(series.name))
)) + 1
worksheet2.set_column(idx, idx, max_len)
return outputPath
def visualize_image(final_bbox, final_preds, l_words, image,label2color):
draw = ImageDraw.Draw(image)
font = ImageFont.load_default()
json_df = []
for ix, (prediction, box) in enumerate(zip(final_preds, final_bbox)):
if prediction is not None:
predicted_label = iob_to_label(prediction).lower()
if predicted_label not in ["others"]:
draw.rectangle(box, outline=label2color[predicted_label])
draw.text((box[0]+10, box[1]-10), text=predicted_label, fill=label2color[predicted_label], font=font)
json_dict = {}
json_dict['TEXT'] = l_words[ix]
json_dict['LABEL'] = label2color[predicted_label]
json_df.append(json_dict)
return image, json_df
def rotate_image(image):
extracted_text = pytesseract.image_to_string(image)
# check if the image contains any text
if not extracted_text:
print("The image does not contain any text.")
return None
elif extracted_text.isspace():
print("The image contains only spaces.")
return None
text = pytesseract.image_to_osd(image)
angle = int(re.search('(?<=Rotate: )\d+', text).group(0))
angle = 360 - angle
rotated = ndimage.rotate(image, angle)
data = Image.fromarray(rotated)
return data
# correct the skewness of images
def correct_skew(image, delta=1, limit=5):
def determine_score(arr, angle):
data = inter.rotate(arr, angle, reshape=False, order=0)
histogram = np.sum(data, axis=1, dtype=float)
score = np.sum((histogram[1:] - histogram[:-1]) ** 2, dtype=float)
return histogram, score
# Convert the PIL Image object to a numpy array
image = np.asarray(image.convert('L'), dtype=np.uint8)
# Apply thresholding
thresh = cv2.threshold(image, 0, 255, cv2.THRESH_BINARY_INV + cv2.THRESH_OTSU)[1]
scores = []
angles = np.arange(-limit, limit + delta, delta)
for angle in angles:
histogram, score = determine_score(thresh, angle)
scores.append(score)
best_angle = angles[scores.index(max(scores))]
(h, w) = image.shape[:2]
center = (w // 2, h // 2)
M = cv2.getRotationMatrix2D(center, best_angle, 1.0)
corrected = cv2.warpAffine(image, M, (w, h), flags=cv2.INTER_CUBIC, \
borderMode=cv2.BORDER_REPLICATE)
return best_angle, corrected
def removeBorders(img):
result = img.copy()
if len(result.shape) == 2:
# if the input image is grayscale, convert it to BGR format
result = cv2.cvtColor(result, cv2.COLOR_GRAY2BGR)
gray = cv2.cvtColor(result, cv2.COLOR_BGR2GRAY) # convert to grayscale
thresh = cv2.threshold(gray, 0, 255, cv2.THRESH_BINARY_INV + cv2.THRESH_OTSU)[1]
# Remove horizontal lines
horizontal_kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (40,1))
remove_horizontal = cv2.morphologyEx(thresh, cv2.MORPH_OPEN, horizontal_kernel, iterations=2)
cnts = cv2.findContours(remove_horizontal, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
cnts = cnts[0] if len(cnts) == 2 else cnts[1]
for c in cnts:
cv2.drawContours(result, [c], -1, (255,255,255), 5)
# Remove vertical lines
vertical_kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (1,40))
remove_vertical = cv2.morphologyEx(thresh, cv2.MORPH_OPEN, vertical_kernel, iterations=2)
cnts = cv2.findContours(remove_vertical, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
cnts = cnts[0] if len(cnts) == 2 else cnts[1]
for c in cnts:
cv2.drawContours(result, [c], -1, (255,255,255), 5)
return result
def color2label_except(label2color, excluded_labels):
"""
Inversely maps colors to labels based on the provided label2color dictionary,
excluding the specified labels.
Args:
label2color (dict): Dictionary mapping labels to colors.
excluded_labels (list): List of labels to exclude.
Returns:
dict: Dictionary mapping colors to labels, excluding the specified labels.
"""
# Filter out excluded labels from label2color dictionary
filtered_label2color = {label: color for label, color in label2color.items() if label not in excluded_labels}
# Invert the filtered label2color dictionary to create color2label mapping
return {v: k for k, v in filtered_label2color.items()}
def add_dataframe(df_main,labels_repeating,label2color):
col_name_map =color2label_except(label2color,labels_repeating)
columns = list(col_name_map.values())
data = {col:[] for col in columns}
for i in df_main:
for k, v in i.items():
if k in col_name_map:
data[col_name_map[k]].append(v)
# join the list of strings for each column and convert to a dataframe
for col in columns:
data[col] = [' '.join(data[col])]
df_upper = pd.DataFrame(data)
key_value_pairs = []
for col in df_upper.columns:
key_value_pairs.append({'key': col, 'value': df_upper[col][0]})
df_key_value = pd.DataFrame(key_value_pairs)
# Extract the value from the containertype column
# container_quantity = int(df_key_value[df_key_value['key'] == 'containertype']['value'].str.split("x").str[0])
# # Add a new row to the DataFrame
# df_key_value = df_key_value.append({'key': 'containerquantity', 'value': container_quantity}, ignore_index=True)
# # Extract the desired value from the containertype column
# df_key_value.loc[df_key_value['key'] == 'containertype', 'value'] = df_key_value.loc[df_key_value['key'] == 'containertype', 'value'].str.split("x").str[1]
return df_key_value
import statistics
def id2label_row(s, id2label):
if s in id2label.values():
return s
return id2label[s]
def dist_height(y1,y2):
return abs(int(y1)- int(y2))
def mergeBoxes(df):
xmin, ymin, xmax, ymax = [], [], [], []
for i in range(df.shape[0]):
box = df['bbox_column'].iloc[i]
xmin.append(box[0])
ymin.append(box[1])
xmax.append(box[2])
ymax.append(box[3])
return [min(xmin), min(ymin), max(xmax), max(ymax)]
def transform_dataset(df, merge_labels):
df_temp = df.iloc[merge_labels] # a duplicate df with only concerned rows
df_temp.reset_index(drop = True, inplace = True)
text = ' '.join(df_temp['scr_column'])
bbox = mergeBoxes(df_temp)
retain_index = merge_labels[0] #the first index is parent row
df['scr_column'].iloc[retain_index] = text
df['bbox_column'].iloc[retain_index] = bbox
# keeping the first & removing rest
df = df.loc[~df.index.isin(merge_labels[1:]), :]
df.reset_index(drop = True, inplace = True)
return df
def box_overlap(box1, box2, horizontal_vertical):
# Extract coordinates of box1
x1_box1, y1_box1, x2_box1, y2_box1 = box1
# Extract coordinates of box2
x1_box2, y1_box2, x2_box2, y2_box2 = box2
# Check if boxes overlap horizontally and vertically
if horizontal_vertical == "H":
if x1_box1 <= x2_box2 and x2_box1 >= x1_box2:
return True
else:
return False
if horizontal_vertical == "V":
if y1_box1 <= y2_box2 and y2_box1 >= y1_box2:
return True
else:
return False
def horizonatal_merging(df, font_length, perform_overlapping =False, x_change = 0, y_change = 0):
fat_df = df.copy()
for i in range(df.shape[0]):
box = fat_df['bbox_column'].iloc[i]
fat_df['bbox_column'].iloc[i] = [box[0]-x_change, box[1]-y_change, box[2]+x_change, box[3] + y_change]
if perform_overlapping == True:
redundant_rows = []
for i in range(fat_df.shape[0]):
box_i = fat_df.bbox_column[i]
indices2merge = []
for j in range(i+1, fat_df.shape[0]):
if fat_df.preds_column[j] == fat_df.preds_column[i]: # if labels are same
box_j = fat_df.bbox_column[j]
if abs(box_i[1]-box_j[3])<font_length*1.5: # if the boxes are at height within 50% more range of font size
# Check if boxes overlap horizontally
if box_overlap(box_i, box_j, 'H'):
indices2merge.append(j)
df.scr_column[i] += df.scr_column[j]
box_i = fat_df.bbox_column[j] # finding the next connected word
#once we have all indices that belong to a particular category
# merging the boundong boxes, keeping them in 1st note/row.
if len(indices2merge)!=0:
df['bbox_column'].iloc[i] = mergeBoxes(df.loc[indices2merge])
redundant_rows.extend(indices2merge)
#now since all the transformation is done, lets remove the redundant rows
return df.drop(redundant_rows)
def mergeLabelsExtensive_repeating(df_grouped, repeating_label):
df_grouped.reset_index(inplace = True, drop = True)
# this function merges same label entities together in a single instance.
df_grouped = df_grouped[df_grouped['preds_column'].isin(repeating_label)]
font_length =0
count = 0
while count<5 and count<df_grouped.shape[0]:
box_i = df_grouped['bbox_column'].iloc[count] # box of current label contains [x1,y1,x3,y3]
font_length += box_i[3]-box_i[1]
count +=1
font_length = font_length/5
df_grouped = horizonatal_merging(df_grouped, font_length, True, 30, 0)
return df_grouped
def group_labels_wrt_height(df):
"""
This function groups the labels based on the height of the bounding box.
"""
#sorting the lines based on heights using column 'y_axis'
df = df.sort_values(by='y_axis')
df.reset_index(inplace = True, drop = True)
print("entering: group_labels_wrt_height ")
final_yaxis = []
final_scr = []
final_pred = []
current_group = []
current_scr = []
current_pred = []
# Iterate through the column values
for i, (value,scr,preds ) in enumerate(zip(df['y_axis'], df['scr_column'], df['preds_column'])):
if i == 0:
# Start a new group with the first value
current_group.append(value)
current_scr.append(scr)
current_pred.append(preds)
else:
# Check if the difference between the current value and the previous value is <= 20
if abs(value - df['y_axis'][i - 1]) <= 35:
# Add the value to the current group
current_group.append(value)
current_scr.append(scr)
current_pred.append(preds)
else:
# Start a new group with the current value
final_yaxis.append(current_group)
final_scr.append(current_scr)
final_pred.append(current_pred)
current_group = [value]
current_scr = [scr]
current_pred = [preds]
# Add the last group
final_yaxis.append(current_group)
final_scr.append(current_scr)
final_pred.append(current_pred)
final_grouped_df = pd.DataFrame({'y_axis': final_yaxis, 'scr_column': final_scr, 'preds_column': final_pred})
print("Grouped df after sorting based on height")
print_df(final_grouped_df)
return final_grouped_df
# searches the set of labels in the whole range
def search_labelSet_height_range(df, d, keyList):
print("search_labelSet_height_range")
keyDict = dict.fromkeys(keyList, []) #stores the required information as dictonary, then coverted to df
print("Dataframe from extraction is going to happen: ")
for i in range(df.shape[0]): # search df for right-bottom y axis value and check if it lies within the range d.
box = df['bbox_column'].iloc[i]
if dist_height(box[1], d)<50:
key = df['preds_column'].iloc[i]
keyDict[key] = df['scr_column'].iloc[i]
return keyDict
def clean_colText(df, column):
for i in range(df.shape[0]):
df[column].iloc[i] = df[column].iloc[i].replace('[', '').replace('|', '').replace('+', '')
return df
def find_repeatingLabels(df, labels_repeating):
print("In find_repeatingLabels: ")
row2drop = [] # dropping the rows that have been covered in previous dataframe
for i in range(df.shape[0]):
df['preds_column'].iloc[i] = id2label_row(df['preds_column'].iloc[i], id2label)
if df['preds_column'].iloc[i] not in labels_repeating:
row2drop.append(i)
df.drop(index = row2drop, inplace = True)
df = clean_colText(df, 'scr_column')
print("removing non-tabular labels.")
df = mergeLabelsExtensive_repeating(df,labels_repeating)
print('after merging non-tabular labels: ')
labels_repeating = list(set(list(df["preds_column"])))
print("labels_repeating in this document are: ",labels_repeating)
# adding extra column that contains the Y-axis information (Height)
df['y_axis'] = np.NaN
for i in range(df.shape[0]):
box = df['bbox_column'].iloc[i]
df['y_axis'].iloc[i] = box[1]
print("After adding y-axis data in the dataframes: ")
df = mergeLabelsExtensive(df)
print("aftermerging the df extensively")
print("Grouping the labels wrt heights: ")
grouped_df = group_labels_wrt_height(df)
#once labels are grouped, now we will create dictionaries for labels and values occuring in single line
row_dicts = [] # will contains each row of df as single dictionary.
for _, row in grouped_df.iterrows():
row_dict = {}
for preds, scr in zip(row['preds_column'], row['scr_column']):
row_dict[preds] = scr
row_dicts.append(row_dict)
#creating new
final_df = pd.DataFrame(columns=labels_repeating)
for d in row_dicts:
final_df = final_df.append(d, ignore_index=True)
final_df = final_df.fillna('')
return final_df
def mergeImageVertical(images):
# pick the image which is the smallest, and resize the others to match it (can be arbitrary image shape here)
min_shape = sorted( [(np.sum(i.size), i.size ) for i in images])[0][1]
imgs_comb = np.hstack([i.resize(min_shape) for i in images])
# for a vertical stacking it is simple: use vstack
imgs_comb = np.vstack([i.resize(min_shape) for i in images])
imgs_comb = Image.fromarray(imgs_comb)
return imgs_comb
def perform_erosion(img):
# Check if the image is already in grayscale
if len(img.shape) == 2:
gray = img
else:
# Convert the image to grayscale
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
# Define the kernel for erosion and dilation
kernel = np.ones((3, 3), np.uint8)
# Perform erosion followed by dilation
erosion = cv2.erode(gray, kernel, iterations=1)
dilation = cv2.dilate(erosion, kernel, iterations=1)
# Double the size of the image
double_size = cv2.resize(gray, None, fx=2, fy=2, interpolation=cv2.INTER_LINEAR)
# Perform erosion on the doubled image
double_erosion = cv2.erode(double_size, kernel, iterations=1)
return double_erosion
def remove_leading_trailing_special_characters(input_string):
cleaned_string = re.sub(r'^[^A-Za-z0-9]+|[^A-Za-z0-9]+$', '', str(input_string))
return cleaned_string
def clean_dataframe(df):
# Apply the remove_leading_trailing_special_characters function to all string columns
for column in df.select_dtypes(include='object').columns:
df[column] = df[column].apply(remove_leading_trailing_special_characters)
# Remove rows with all NaN or blank values
df = df.fillna('') # Replace NaN values with blank
return df
def mergeLabelsExtensive(df_grouped):
i = 0
while i < df_grouped.shape[0]:
merge_labels = [i] # collects indices whose data has been merged, so we need to delete it now.
label = df_grouped['preds_column'].iloc[i]
box1 = df_grouped['bbox_column'].iloc[i]
for j in range(i+1, df_grouped.shape[0]):
box2 = df_grouped['bbox_column'].iloc[j]
if label == df_grouped['preds_column'].iloc[j] and dist_height(box1[3], box2[3])<20: # which are in the vicinity of 20 pixels.
merge_labels.append(j)
print_df(df_grouped)
df_grouped = transform_dataset(df_grouped, merge_labels)
i = i+1
return df_grouped
def multilabelsHandle(df, thermo_details):
# Since 0 is assigned to 'others' and these values are not so important. We delete these values.
df = df[df.preds_column != 0]
df.reset_index(drop=True, inplace=True)
for i in range(df.shape[0]):
df['preds_column'].iloc[i] = id2label.get(df['preds_column'].iloc[i])
df['preds_column'].unique()
df_grouped = df.copy() #stores the index of relevant labels.
df_grouped.shape[0]
for i in range(df.shape[0]):
if df['preds_column'].iloc[i] not in thermo_details:
df_grouped.drop(i, inplace = True)
df_grouped.reset_index(drop=True, inplace=True)
keyList = df_grouped['preds_column'].unique()
df_grouped = mergeLabelsExtensive(df_grouped)
# extract the height of boxes
df_grouped = extract_yaxis(df_grouped)
shipment_labels = ['delivery_name','delivery_address','contact_phone']
# shipment
heights_shipment = get_heights(df_grouped, shipment_labels)
# now segregating the other repeating values in df like measiure, weight, volume etc.
# they will be containeed within the heights, as they act as boudaries.
df_labelSet = pd.DataFrame(columns= thermo_details)
for i in range(len(heights_shipment)):
if i == len(heights_shipment)-1:
new_df = search_labelSet_between_h1_h2(df_grouped, heights_shipment[i], 5000, keyList)
else:
new_df = search_labelSet_between_h1_h2(df_grouped, heights_shipment[i], heights_shipment[i+1], keyList)
df_labelSet = df_labelSet.append(new_df, ignore_index=True)
return df_labelSet
def completepreprocess(pdffile,ocr_type):
myDataFrame = pd.DataFrame()
myDataFrame2 = pd.DataFrame()
merge_pages=[]
doc = fitz.open(pdffile)
for i in range(0, len(doc)):
page = doc.load_page(i)
zoom = 2
mat = fitz.Matrix(zoom, zoom)
pix = page.get_pixmap(matrix = mat, dpi = 300)
image = Image.frombytes("RGB", [pix.width, pix.height], pix.samples)
ro_image = rotate_image(image)
if ro_image is None:
return None
angle, skewed_image = correct_skew(ro_image)
if skewed_image is None:
return None
remove_border = removeBorders(skewed_image)
image = Image.fromarray(remove_border)
width,height=image.size
label2color = generate_unique_colors(id2label)
width,height=image.size
if ocr_type == "GoogleVisionOCR":
words, boxes = process_image_GoogleVision(image, width, height)
else:
words, boxes = process_image_pytesseract(image, width, height)
bbox, preds, words, image = process_image_encoding(model, processor, image, words, boxes,width,height)
im, df_visualize = visualize_image(bbox, preds, words, image,label2color)
df_main = process_form_(df_visualize)
bbox_column = bbox
preds_column = preds
scr_column = words
# dictionary of lists
dict = {'bbox_column': bbox_column, 'preds_column': preds_column, 'scr_column': scr_column}
df_single_page = pd.DataFrame(dict)
labels_repeating = ['art_code', 'ref_code', 'detail_desc','lot_id','detail_qty','detail_um','detail_tare','detail_grossw','detail_netw','detail_origin','varieta','raccolta']
df_repeating_page = find_repeatingLabels(df_single_page, labels_repeating)
myDataFrame2= myDataFrame2.append(df_repeating_page,sort=False)
df1=add_dataframe(df_main,labels_repeating,label2color).astype(str)
myDataFrame= myDataFrame.append(df1,sort=False).reset_index(drop = True)
myDataFrame['value'].apply(len)
row2drop = []
for i in range(myDataFrame.shape[0]):
if len( myDataFrame['value'].iloc[i]) ==0:
row2drop.append(i)
myDataFrame.drop(index = row2drop, inplace = True)
myDataFrame.reset_index(drop = True, inplace = True)
myDataFrame = myDataFrame[myDataFrame["value"].notnull()]
myDataFrame.drop_duplicates(subset=["key"],inplace=True)
myDataFrame2 = myDataFrame2.loc[:, ~(myDataFrame2.apply(lambda x: all(isinstance(val, list) and len(val) == 0 for val in x)))]
merge_pages.append(im)
im2=mergeImageVertical(merge_pages)
myDataFrame2 = clean_dataframe(myDataFrame2)
myDataFrame = clean_dataframe(myDataFrame)
myDataFrame = myDataFrame[myDataFrame['key'] != 'others']
output_excel_path = createExcel(myDataFrame, myDataFrame2, pdffile.name)
return im2,myDataFrame,myDataFrame2,output_excel_path
title = "Interactive demo: Document Information Extraction model PDF/Images"
description = "Upload your own document, or use the one given below at the left corner. Results will show up in a few seconds. The annotated image can be opened in a new window for a better view."
css = """.output_image, .input_image {height: 600px !important}"""
examples = [["sample_doc.pdf"]]
iface = gr.Interface(
fn=completepreprocess,
inputs=[
gr.components.File(label="PDF"),
gr.components.Dropdown(label="Select the OCR", choices=["Pytesseract","GoogleVisionOCR"]),
],
outputs=[
gr.components.Image(type="pil", label="annotated image"),
"dataframe",
"dataframe"
#gr.File(label="Excel output")
],
title=title,
description=description,
examples=examples,
css=css
)
iface.launch(inline=True, debug=True)