File size: 4,708 Bytes
b314edd
 
 
 
 
 
 
 
1fd5d04
 
 
 
 
 
 
aff689d
1fd5d04
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
aff689d
1fd5d04
 
 
 
 
 
 
 
 
aff689d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1fd5d04
aff689d
1fd5d04
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
aff689d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
import cv2 as cv
import matplotlib.pyplot as plt
from transformers import TrOCRProcessor, VisionEncoderDecoderModel
import numpy as np
from concurrent.futures import ProcessPoolExecutor
from openai import OpenAI


def preprocess_image(image):
  gray_image = cv.cvtColor(image, cv.COLOR_BGR2GRAY)
  ret, bin_image = cv.threshold(gray_image, 127, 255, cv.THRESH_OTSU)

  bin_image = cv.copyMakeBorder(bin_image, int(0.10 * image.shape[0]), int(0.05 * image.shape[0]), int(0.05 * image.shape[1]), int(0.10 * image.shape[1]), cv.BORDER_CONSTANT, value=(255, 255, 255))
  return bin_image

#bin_image = preprocess_image(image)

def split_image_into_lines(image):
  lines = []
  while (image.shape[0] > 20):
    flag1 = 0
    flag2 = 0

    for i in range(image.shape[0]):
      if flag1 == 0:
        for j in range(image.shape[1]):
          pixel_value = image[i][j]
          if (pixel_value == 0) & (flag1 == 0):
            start = i
            flag1 = 1
            flag2 = 1
      if flag2 == 1:
        num_white_pixels = np.sum(image[i + 1] == 255)
        if (num_white_pixels > 0.98 * image.shape[1]):
          end = i + 1
          break


    line = image[int(start - 0.2 * (end - start + 1)): int(end + 1 + 0.2 * (end - start + 1))][:]
    if line.shape[0] > 20:
      line_rgb = cv.cvtColor(line, cv.COLOR_GRAY2RGB)
      lines.append(line_rgb)

    pads = 255 * np.ones((20, image.shape[1]), dtype='uint8')
    new_image = image[int(end + 2 -(0.2 * (end - start + 1))):][:]
    new_image = np.concatenate((pads, new_image))
    image = new_image


  return lines

#lines = split_image_into_lines(bin_image)


def generate_text(line):
  pixel_values = processor(images=line, return_tensors="pt").pixel_values
  generated_ids = model.generate(pixel_values, max_new_tokens=50)
  generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
  return generated_text


def get_improved_result(lines):
    with ProcessPoolExecutor() as executor:
        results = ' '.join(executor.map(generate_text, lines))
        #improve results with llm
    
        client = OpenAI()
    
        completion = client.chat.completions.create(
            model="gpt-4o",
            messages=[
                {
                    "role": "user",
    
                    "content": f"I have a string that was extracted from an image of handwritten text. The extraction process introduced minor grammatical, spelling, and punctuation errors. Please carefully review the text below and make any necessary corrections to improve readability and accuracy while preserving the original meaning. Do not change the content or style beyond necessary corrections. Return the corrected text only without adding any headings, explanations, or extra formatting. Text: {results}"
                }
            ]
        )
    
        improved_text = completion.choices[0].message.content

    return improved_text



def put_text(text, font, font_scale, color, thickness, max_width, out_image_width, top_margin):
  words = text.split(" ")
  lines = []
  current_line = ""

  for word in words:
    if cv.getTextSize(current_line + " " + word, font, font_scale, thickness)[0][0] <= (max_width * out_image_width):
      current_line += " " + word
    else:
      lines.append(current_line)
      current_line = word

  lines.append(current_line)

  out_image_height = sum([cv.getTextSize(line, font, font_scale, thickness)[0][1] for line in lines]) + 2 * top_margin + 20 * (len(lines) - 1) #20 is the gap between two consecutive lines

  out_image = 255 * (np.ones((out_image_height, out_image_width, 3), dtype=np.uint8))

  top = top_margin
  for line in lines:
    cv.putText(out_image, line.strip(), (int(((1 - max_width) * out_image_width) / 2), top), font, font_scale, 0, thickness, lineType=cv.LINE_AA)
    top += cv.getTextSize(line.strip(), font, font_scale, thickness)[0][1] + 20

  return out_image

font = cv.FONT_HERSHEY_DUPLEX
font_scale = 2
color = 0
thickness = 2
max_width = 0.9
out_image_width = 1500
top_margin = 100

#out_image = put_text(improved_text, font, font_scale, color, thickness, max_width, out_image_width, top_margin)

def predict(input_img):
    bin_image = preprocess_image(input_img)
    lines = split_image_into_lines(bin_image)
    improved_text = get_improved_result(lines)
    out_image = put_text(improved_text, font, font_scale, color, thickness, max_width, out_image_width, top_margin)
    return out_img

gradio_app = gr.Interface(
    predict,
    inputs=gr.Image(label="Image with handwritten text", sources=['upload']),
    outputs=[gr.Image(label="Output Image")],
    title="Extract Handwritten Text",
)

if __name__ == "__main__":
    gradio_app.launch()