Introduction

I have recently completed a project at work, the creation of a custom ChatGPT chatbot. I will break the project into two parts, the first part will scan a folder of PDF files into a dataframe and the second part will pass the data to OpenAI API. This entire project was completed in python.

Project outline

PDFs can be easily scanned in python with the pypdf module. It is easily installed and easily run, but I have found that the quality of the scan to be lacking. Pypdf also seems to have some issues with PDFs that created from scanned documents, not directly created from a text document. For this reason, I have found an alternative method.

The first step is to convert all the PDFs in a directory to PNG images. This can be achieved with the convert from path function from the pdf2image library with the poppler application. The poppler program can be downloaded here and unzipped into its own directory. There is no Windows version on the poppler site, but I found a repo with a nearly updated Windows version here. You will need to copy the directory path into your code. We can create a for loop to open each PDF file one at a time. It’s important to remember to change the ‘’ to’/’ for Windows users when referring to directory positions.

The second step is to then scan through the PNG images with OCR. For this task, we can use Tesseract. Tesseract is a Google project that is easy to use. Like Poppler, you will need to download the application separately. You will also need to install the helper python package pytessseract. The Tesseract application can be found here. I have my program to save the data in a CSV file, but you can store it anyway you want. I decided to save each PDF file as a separate CSV file and assigning each row as a different PNG file or page of the PDF. This was to ensure that may data is easily organized

The next stages require getting into the Langchain library. These steps will be included in the follow-up to this post as both post, are quite lengthy and each can stand alone.

Converting PDF to PNG with Poppler

Again, prior to running this code, you will need to install the Poppler Application. You also need to copy the directory to the location of the Poppler bin folder. The rest of this section is pretty simple, I’ve created a loop to go through every filename that ends with ‘.pdf’ in a specific PDF folder. I also save the PNG file with the page number included into the title. If the results from the OCR scans are inaccurate, you can adjust the resolution of the PNG files with the parameter ‘dpi = 300’ passed to the convert from path function. The default value is 100. Fair warning, increasing the resolution will slow down the entire process and can potentially add additional artifacts into the OCR scan.

import os
import pandas as pd
from PIL import Image
from pdf2image import convert_from_path

poppler_path = 'C:/Program Files/poppler-23.08.0/Library/bin'
for pdf_file in [f for f in os.listdir('//Desktop/PDF') if f.endswith('.pdf')]:
  images = convert_from_path(pdf_path = '//Desktop/PDF/' + pdf_file, poppler_path = poppler_path)

  for count, img in enumerate(images):
    img_name = f"{pdf_file[:-4]}_page_{count}.png"
    img.save('//Desktop/PDF/' + img_name, "PNG")

OCR from PNG files

The Tesseract application is required for the next stage. Since every PNG from every PDF will need to go through the process, I’ve recreated the first section and included the Tesseract functions into the same loop. I’ve also included a step to delete each PNG file after it has been scanned, since it will no longer be needed. The final stage is to save all the returned data as a CSV file. I have found that it is useful to specify the encoding used in saving the CSV.

import os
import pandas as pd
from PIL import Image
from pdf2image import convert_from_path
import pytesseract

poppler_path = 'C:/Program Files/poppler-23.08.0/Library/bin'
pytesseract.pytesseract.tesseract_cmd = '//Tesseract-OCR/tesseract.exe'

for pdf_file in [f for f in os.listdir('//Desktop/PDF') if f.endswith('.pdf')]:
  images = convert_from_path(pdf_path = '//Desktop/PDF' + pdf_file, poppler_path = poppler_path)
  extracted_text = []
  for count, img in enumerate(images):
    img_name = f"{pdf_file[:-4]}_page_{count}.png"
    img.save('//Desktop/PDF' + img_name, "PNG")
    
    extracted_data.append(pytesseract.image_to_string(Image.open('C:/Users/Mark/Desktop/PDF' + img_name)))
    os.remove('//Desktop/PDF' + img_name)
    
  df = pd.DataFrame(extracted_text)
  df.to_csv('//Desktop/PDF' + pdf_name[:-4] + '.csv', encoding = 'utf-8-sig')

Conclusion

We are finally able to create a usable CSV file from a OCR scanned PDF file. The first step was to convert the pdf into PNG files with Poppler. Each png is then scanned with Tesseract. And the returned values are stored in a CSV file. By why would you want to go through all the steps in the first place? Well, we will need to proceed with the next post about creating the ChatGPT chatbot.