This repository contains a small utility script for extracting structured tables from images using Tesseract OCR and a simple geometric heuristic.

It was originally developed for a finance project that required converting semi-structured table images (e.g., supplier tables in reports) into clean CSV files.

• Uses Tesseract OCR via pytesseract to recognize text and bounding boxes.
• Automatically detects the header row using configurable keyword anchors.
• Infers column boundaries from the horizontal positions of header words.
• Assigns each token to a column based on its x-coordinate.
• Groups words into rows and exports a clean CSV with predefined column names.

• Python 3.8+ (other versions may also work)
• Tesseract OCR installed on your system
• Python packages (see requirements.txt):
  • pytesseract
  • opencv-python
  • pandas

Make sure tesseract is available in your system PATH, or configure pytesseract.pytesseract.tesseract_cmd if needed.

Create a virtual environment (optional but recommended), then:

pip install -r requirements.txt

macOS (Homebrew):

brew install tesseract

Ubuntu / Debian:

sudo apt-get install tesseract-ocr

Windows: Download and install from the official project page or a trusted binary distribution, and ensure the installation path is added to your system PATH.

Run the script from the command line:

python ocr_table_extractor.py \
  --input_folder /path/to/images \
  --output_folder /path/to/output_csvs

--input_folder: directory containing input images (.png, .jpg, .jpeg, .tiff)

--output_folder: directory where per-image CSV files will be saved

For each image, the script will:

Preprocess the image (grayscale + invert) to improve OCR.

Run Tesseract OCR with image_to_data to obtain per-word bounding boxes.

Detect the header row using predefined header keywords.

Infer column boundaries from header word positions.

Assign each token to a column based on its x-center.

Group tokens into rows and output a CSV with the following columns:

Supplier Name
Industry
Mkt Cap (M)
Total Relationship Size M)
% Cost
Cost Category
% Supplier's Revenue
Source
Size Source

You can modify the list of column_names in ocr_table_extractor.py to match your own table layout.

Header detection: The function find_column_boundaries uses a dictionary of header keywords to locate the header row and compute column anchors. You can edit the header_keywords mapping to adapt the script to other table formats.

Column set: The final column order is defined in column_names inside main(). Add/remove columns or rename them as needed.

Filtering & cleaning: Confidence thresholds and text cleaning (e.g., removing row numbers from "Supplier Name") can be adjusted inside extract_data_from_image_by_coordinates.