from pdf2image import convert_from_path
import cv2
import numpy as np
import numpy as np
from PIL import Image
import json
from anthropic import Anthropic, Client
import gradio as gr
def get_base64_encorded_image(image_path):
with open(image_path, "rb") as image_file:
binary_data = image_file.read()
base64_encorded_data = base64.b64encode(binary_data)
base64_string = base64_encorded_data.decode('utf-8')
return base64_string
## Process pdf
def convert_pdf_to_image(pdf_path):
# Convert PDF to images
pages = convert_from_path(pdf_path, dpi=400)
# Save images as PNG files
for i, page in enumerate(pages):
page.save(f'page_{i}.png', 'PNG')
print(f"Converted {len(pages)} pages to images.")
return pages
## Image process Subprocess - De-stamp
def destamp_image(img_path):
bgr_img = cv2.imread(img_path)
hsv_img = cv2.cvtColor(bgr_img, cv2.COLOR_BGR2HSV)
# Convert the BGR image to grayscale
gray_img = cv2.cvtColor(bgr_img, cv2.COLOR_BGR2GRAY)
# HSV ragne: (0-180, 0-255, 0-120)
# for character black color:
# H: 0-180,
# S: 0-255 ，
# V: 0-120 ，
lower_black = np.array([0,0,0])
upper_black = np.array([180,255,120])
mask = cv2.inRange(hsv_img, lower_black, upper_black)
deRed_img = ~mask # Single channel image
# imshow mask
#print(f"deRed_img shape: {deRed_img.shape}")
#show_image(deRed_img)
# thresholding -2
ret, threshold_img_2 = cv2.threshold(deRed_img, 120, 255, cv2.THRESH_BINARY | cv2.THRESH_OTSU)
#print(f' threshold-2 shape: {threshold_img_2.shape}')
#show_image(threshold_img_2)
# Desired shape: (x, y, 1)
new_shape = (threshold_img_2.shape[0], threshold_img_2.shape[1], 1)
# Resize using numpy.resize()
result_img = np.resize(threshold_img_2, new_shape)
print(f"result_img.shape: {result_img.shape}")
cv2.imshow(result_img)
#save result_img
result_filepath="result_img_0.png"
cv2.imwrite(result_filepath, result_img)
return result_filepath
def extract_image_table(image_path):
# extract table information
response = {}
response = extract_table_info(image_path)
# Get text element from response
check_response(response)
# Extract response.content[0].text
json_data = extract_json(response) #type(json_data) = "dict"
print(f"json_data: {json_data}")
return json_data
## Extract Table Information
def extract_table_info(image_path):
# Claude
client = Anthropic(api_key=my_api_key) # Pass the API key here
MODEL_NAME = "claude-3-5-sonnet-20240620"
#Do ascending sort with index of value of "代碼" for all the rows in each section. If there is "X" or "x" in "代碼", treat it as "9".
message_list = [
{
"role": "user",
"content": [
{"type": "image", "source": {"type": "base64", "media_type": "image/png", "data": get_base64_encorded_image(image_path)}},
#{"type": "text", "text": "Please extract the table information of the image, keep the context in Traditional Chinese without translation, all the alphanumeric chararacter exressed in string, give me a json dictionary of the information extracted"},
{
"type": "text",
"text": """
Please extract the table information of the image, keep the context in Traditional Chinese without translation.
if you can not recognize the value precisely, please infer it and try to make a best guess.
If you can not make the best guess, please return “UNK”.
Create a structured set of data in json format providing key information about a table.
Keep the section titles in the table as a parts of json.
Be sure to extract the information of "代碼", and save them as part of json.
All the value extracted are string, including the "代碼".
Do not do any sort operation with all the rows.
Extract the text information of each cell precisely. Do not make inference between "代碼" and "項目" if you can not extract it precisely.
JSON fields must be labelled as:
Example json structure is:
{
"table meta": [
{"企業名稱": },
{"表頭名稱": },
{"報表日期": },
{"幣別": },
...
...
...
],
"table detail": [
{
...
...
...
},
{
...
...
...
},
...
...
...
]
}
Output the json structure as a string starting with and ending with XML tags.
Do not return any narrative language. Look at the images in detail.
Do not insert and control code, like line feed, tab indent: "\n"
IF YOU COULD NOT FIND THE RIGHT INFORMATION JUST RETURN THIS VALUE “UNK”.
Example:
{
"table meta": [
{"企業名稱": "台灣水泥股份有限公司"},
{"表頭名稱": "個體資產負債表"},
{"報表日期": "民國 112 年及 111 年 12 月 31 日"},
{"幣別": "新台幣仟元"},
...
...
...
],
"table detail": [
{
"資產": [
{ "流動資產":
[
{
"代碼": "1100",
"項目": "現金及約當現金（附註四及六）",
"112年12月31日金額": "1,516,633",
"112年12月31日%": "-",
"111年12月31日金額": "4,243,295",
"111年12月31日%": "1"
},
{
"代碼": "1110",
"項目": "透過損益按公允價值衡量之金融資產（附註四、七及二六）",
"112年12月31日金額": "341,056",
"112年12月31日%": "-",
"111年12月31日金額": "259,919",
"111年12月31日%": "-"
},
{
"代碼": "1120",
"項目": "透過其他綜合損益按公允價值衡量之金融資產（附註四、八及二六）",
"112年12月31日金額": "4,333,594",
"112年12月31日%": "1",
"111年12月31日金額": "3,607,819",
"111年12月31日%": "1"
},
{
"代碼": "1150",
"項目": "應收票據及帳款淨額（附註四及九）",
"112年12月31日金額": "5,801,135",
"112年12月31日%": "2",
"111年12月31日金額": "5,319,368",
"111年12月31日%": "1"
},
{
"代碼": "1180",
"項目": "應收票據及帳款－關係人（附註四及二七）",
"112年12月31日金額": "572,118",
"112年12月31日%": "-",
"111年12月31日金額": "681,793",
"111年12月31日%": "-"
},
{
"代碼": "130X",
"項目": "存貨（附註四及十）",
"112年12月31日金額": "1,782,735",
"112年12月31日%": "1",
"111年12月31日金額": "2,321,850",
"111年12月31日%": "1"
},
{
"代碼": "1470",
"項目": "其他流動資產（附註二一及二七）",
"112年12月31日金額": "411,540",
"112年12月31日%": "-",
"111年12月31日金額": "248,683",
"111年12月31日%": "-"
},
{
"代碼": "11XX",
"項目": "流動資產總計",
"112年12月31日金額": "14,758,811",
"112年12月31日%": "4",
"111年12月31日金額": "16,682,727",
"111年12月31日%": "4"
}
]
},
{
"非流動資產": [
{
"代碼": "1517",
"項目": "透過其他綜合損益按公允價值衡量之金融資產（附註四、八及二六）",
"112年12月31日金額": "9,638,255",
"112年12月31日%": "3",
"111年12月31日金額": "7,633,603",
"111年12月31日%": "2"
},
{
"代碼": "1550",
"項目": "採用權益法之投資（附註四、五及十一）",
"112年12月31日金額": "312,351,291",
"112年12月31日%": "82",
"111年12月31日金額": "307,101,709",
"111年12月31日%": "82"
},
{
"代碼": "1600",
"項目": "不動產、廠房及設備（附註四、五、十二、十三及二八）",
"112年12月31日金額": "28,052,603",
"112年12月31日%": "7",
"111年12月31日金額": "35,583,596",
"111年12月31日%": "10"
},
{
"代碼": "1755",
"項目": "使用權資產（附註四、十五、二十、二七）",
"112年12月31日金額": "1,797,820",
"112年12月31日%": "1",
"111年12月31日金額": "1,788,972",
"111年12月31日%": "1"
},
{
"代碼": "1760",
"項目": "投資性不動產（附註四、十四及二十）",
"112年12月31日金額": "13,042,677",
"112年12月31日%": "3",
"111年12月31日金額": "2,436,675",
"111年12月31日%": "-"
},
{
"代碼": "1821",
"項目": "無形資產（附註四及二十）",
"112年12月31日金額": "58,840",
"112年12月31日%": "-",
"111年12月31日金額": "64,956",
"111年12月31日%": "-"
},
{
"代碼": "1915",
"項目": "預付設備款",
"112年12月31日金額": "600,042",
"112年12月31日%": "-",
"111年12月31日金額": "682,765",
"111年12月31日%": "-"
},
{
"代碼": "1975",
"項目": "淨確定福利資產（附註四及十八）",
"112年12月31日金額": "1,507,153",
"112年12月31日%": "-",
"111年12月31日金額": "1,526,546",
"111年12月31日%": "-"
},
{
"代碼": "1990",
"項目": "其他非流動資產（附註四、六、二一及二八）",
"112年12月31日金額": "827,628",
"112年12月31日%": "-",
"111年12月31日金額": "840,688",
"111年12月31日%": "1"
},
{
"代碼": "15XX",
"項目": "非流動資產總計",
"112年12月31日金額": "367,876,309",
"112年12月31日%": "96",
"111年12月31日金額": "357,659,510",
"111年12月31日%": "96"
}]
},
{
"代碼": "1XXX",
"項目": "資產總計",
"112年12月31日金額": "382,635,120",
"112年12月31日%": "100",
"111年12月31日金額": "374,342,237",
"111年12月31日%": "100"
}
]
},
{
"負債": [
...
...
...
]
},
...
...
...
]
}
"""
}
]
}
]
# Update how the API is called
response = client.messages.create(
model=MODEL_NAME,
max_tokens=8192, # limit the amount of response information
messages=message_list,
temperature=0.7,
extra_headers={"anthropic-beta": "max-tokens-3-5-sonnet-2024-07-15"} # Changed to a dictionary
)
tokens = response.usage.output_tokens
print(f"Generated Tokens: {tokens}")
#print(f"Response: {response}")
return response
## Check Response
def check_response(response):
# Check the type and content of the response
print(type(response.content))
print(response.content)
# Assuming the text content is in the first element of the list
if isinstance(response.content, list) and response.content:
content_text = response.content[0].text
#print(json.dumps(content_text, sort_keys=True, indent=4))
else:
print("Unexpected response format. Unable to extract text.")
return None
## Extract Json data
def extract_json(response):
response_text = response.content[0].text # Access the 'text' attribute of the TextBlock object
# Try to find the start and end of the JSON object more robustly
# skip
json_start = response_text.find("")+6 # Skip the tag
json_end = response_text.rfind("") # Include the closing brace
# Check if valid start and end indices were found
if json_start >= 0 and json_end > json_start:
try:
return json.loads(response_text[json_start:json_end])
except json.JSONDecodeError as e:
print(f"Error decoding JSON: {e}")
print(f"Problematic JSON string: {response_text[json_start+1:json_end]}")
return {response_text[json_start+1:json_end]}
else:
print("Could not find valid JSON object in response.")
return
## Convert json to Dataframe
## Convert to csv
## Process PDF
def pipeline(pdf_path):
pages = convert_pdf_to_image(pdf_path)
print(f"pages: {pages}")
destamp_img = destamp_image("page_0.png")
response = {}
response = extract_table_info(destamp_img)
check_response(response)
json_data = extract_json(response)
return len(pages), destamp_img, json_data
## Gradio Interface
title = "Demo: Financial Statement(PDF) information Extraction - Traditional Chinese"
description = """Demo pdf, either editable or scanned image, information extraction for Traditional Chinese without OCR"""
examples = ['text_pdf.pdf', 'image_pdf.pdf']
pdf_file = gr.File(label="Upload PDF", type="filepath")
pages = gr.File(label="Pages", type="filepath")
num_pages = gr.Number(label="Number of Pages")
destamp_img = gr.Image(type="numpy", label="De-stamped Image")
json_data = gr.JSON(label="JSON Data")
app = gr.Interface(fn=pipeline,
inputs=pdf_file,
outputs=[num_pages, destamp_img, json_data],
title=title,
description=description,
examples=examples)
app.queue()
app.launch(debug=True, share=True)
#app.launch()