👁️ Extracting Text from Images: A Step-by-Step Guide to Tesseract OCR with Python

September 3, 2022 by dumira

From reading license plates automatically to scanning paper invoices into database fields, Optical Character Recognition (OCR) is the bridge connecting visual media to digital data strings.

While enterprise cloud solutions (like Google Cloud Vision or AWS Textract) charge per request, you can run high-accuracy, enterprise-ready OCR locally for free using Google’s open-source powerhouse: Tesseract OCR.

In this guide, we will break down the structural pipeline behind computer vision text extraction and build a complete step-by-step Python implementation using pytesseract and OpenCV.

🏗️ The OCR Processing Pipeline

Taking a raw picture from a camera and throwing it straight into an OCR engine usually yields terrible accuracy. Images suffer from poor lighting, noise, shadows, and weird angles. To get clean text extraction, every professional pipeline follows a three-stage workflow:

  1. Image Ingestion & Preprocessing: Converting the image to grayscale, blurring out digital artifacts, and thresholding it to pure black-and-white (binarization).
  2. Text Detection & Segmentation: Tesseract locates where structural text blocks, paragraphs, and individual word lines live on the geometric canvas.
  3. Character Recognition: The neural network engine interprets the distinct pixel contours into real Unicode text characters.

🛠️ Step-by-Step Implementation Guide

Let’s write a modular script that loads an image containing text, sanitizes it so the text pops out for the computer, extracts the text string, and isolates specific word boundary locations.

Step 1: Install the System Tesseract Engine

Unlike standard Python packages, pytesseract is simply a wrapper. It requires the actual Tesseract engine binary compiled on your operating system.

  • Windows: Download and run the executable installer from the UB-Mannheim Tesseract Repo.
  • Mac (via Homebrew): brew install tesseract
  • Linux (Ubuntu/Debian): sudo apt-get install tesseract-ocr

Step 2: Install Python Libraries

Now, open your terminal and pull in the Python wrappers along with OpenCV for advanced image processing:

Bash

pip install pytesseract opencv-python pillow

Step 3: Load and Preprocess the Image

Create a file named ocr_processor.py. We will start by loading our image and stripping out color complexity using OpenCV. This makes it significantly easier for Tesseract to differentiate text borders from background artifacts.

Python

import cv2
import pytesseract
from PIL import Image

# NOTE for Windows Users Only: You must explicitly point python to your tesseract.exe path
# pytesseract.pytesseract.tesseract_cmd = r'C:\Program Files\Tesseract-OCR\tesseract.exe'

# 1. Load the raw image
image_path = "sample_receipt.png"
image = cv2.imread(image_path)

# 2. Preprocessing: Convert to Grayscale
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)

# 3. Preprocessing: Apply Thresholding (Binarization)
# This turns the image into stark black text on a pure white background
processed_img = cv2.threshold(gray, 0, 255, cv2.THRESH_BINARY | cv2.THRESH_OTSU)[1]

# Save the preprocessed image to disk to visually inspect the sanity of the text
cv2.imwrite("sanitized_preprocessed.png", processed_img)

Step 4: Extract Global Text Strings

With our image successfully sanitized into high-contrast pixels, extracting the text string requires just a single line of execution code.

Python

# 4. Pass the processed image matrix into Tesseract
extracted_text = pytesseract.image_to_string(processed_img)

print("📝 --- EXTRACTED TEXT RESULT --- 📝")
print(extracted_text)
print("---------------------------------")

Step 5: Advanced Usage – Getting Bounding Box Data

Sometimes, simply dumping a raw string isn’t enough. If you are building automated forms processing or PDF highlighting engines, you need to know exactly where on the visual canvas each word resides.

We can extract structural dictionary maps detailing the geometric coordinates ($x$, $y$, width, height) of every matched element:

Python

# 5. Extract structural layout dataframe data from the image
data = pytesseract.image_to_data(processed_img, output_type=pytesseract.Output.DICT)

n_boxes = len(data['text'])
for i in range(n_boxes):
    # Filter out empty space detections and low-confidence guesses
    if int(data['conf'][i]) > 60:
        (x, y, w, h) = (data['left'][i], data['top'][i], data['width'][i], data['height'][i])
        
        # Draw a visual green rectangle bounding box around the word onto the original color image
        cv2.rectangle(image, (x, y), (x + w, y + h), (0, 255, 0), 2)

# Save the diagnostic image output showing text box detections
cv2.imwrite("bounding_boxes_result.png", image)
print("🎯 Geometric bounding boxes drawn and saved to bounding_boxes_result.png!")

📈 Pro-Tips for Maximizing Tesseract Accuracy

If you notice Tesseract skipping words or returning gibberish text symbols, adjust these settings inside your code logic:

  • Tesseract Page Segmentation Modes (PSM): Tesseract analyzes structural layouts differently depending on your data structure. You can pass explicit configs like config='--psm 6'. Mode 6 assumes a uniform single block of text, while Mode 4 handles multi-column tables, and Mode 11 hunts for sparse text scattered randomly across the page.
  • Resizing Resolution: Tesseract expects character heights to be roughly 30 to 40 pixels tall. If you are feeding in tiny images or giant 4K mobile scans, use OpenCV’s cv2.resize() function to scale down or scale up before passing the matrix array into the library.
  • Denoising: If the original snapshot is grainy, call cv2.medianBlur(img, 3) to average out high-frequency sensor noise before extracting tokens.