This project focuses on comparing three OCR tools—EasyOCR, TesseractOCR, and Keras-OCR—with potential improvements through preprocessing techniques applied using OpenCV. The goal is to assess the performance of these OCR models and explore how preprocessing can enhance accuracy, especially for a diverse vision dataset of images.
Since the images in the dataset are quite different, it was decided to categorize them into subsets based on common features. Because applying a single preprocessing technique to all images would not be optimal results, so a manual labeling tool was created to assign appropriate labels to each image, allowing targeted preprocessing. The labeled images are stored in image_labels.json, which streamlines the process by associating each image with its corresponding preprocessing method. It is manual tool since there is a not large dataset we have. However, it is still good tool to speed up the process.
The following preprocessing techniques are applied based on the assigned labels in the dataset:
- Same Back Stylewriting:
preprocess_image_same_back_stylewriting
Enhances contrast using CLAHE (Contrast Limited Adaptive Histogram Equalization).
- Handwriting:
preprocess_handwritten_image()
Reduces noise using median blurring, converts to LAB color space for contrast adjustment (CLAHE on the L channel), and converts back to BGR format.
- White on Black:
preprocess_white_on_black_image()
Enhances contrast, sharpens the image, and inverts colors to make white text on a black background more visible.
- Blur-High-Contrast:
preprocess_blur_high_contrast_image()
Adjusts brightness and contrast and applies sharpening filters to improve clarity.
- White Background:
preprocess_white_background_image()
Converts the image to grayscale and applies simple thresholding (binarization) to make text on white backgrounds clearer.
Processed images are saved in the processed_images/ folder, and the appropriate preprocessing technique is applied based on the image's label.
! The list of labels can be expaneded if needed depending on the dataset. Also it is possible to apply some labels for one image.
After preprocessing, the OCR tools are applied:
-
EasyOCR
easy-ocr.py: GPU is recommended for faster processing. -
Keras-OCR
keras-ocr.py: Also works better with GPU acceleration. -
Tesseract-OCR
tesseract-ocr.py: Requires additional files (already included in the project) for setup.
The results from each OCR tool are stored in their respective text files within the results/ folder. These text files contain metrics such as similarity, accuracy, precision, recall, and F1-score for each image, along with the overall results at the end.
A ground truth file ground-truth.txt is provided for evaluating OCR output accuracy by comparing the original text with the results produced by each OCR tool.
For detailed analysis, the results.py script can be run to generate plots showing performance dynamics for each image. This script helps visualize how different preprocessing techniques and OCR tools compare in accuracy across the dataset.
- Install the required dependencies using:
pip install -r requirements.txt
- Run the labeling tool to manually label the images:
python labeling-tool.py
- Run the preprocessing script to apply preprocessing techniques based on image labels:
python preprocessing.py
- Once the images are preprocessed, run the OCR tools:
python easy-ocr.py
python keras-ocr.py
python tesseract-ocr.py
-
Check the results in the results/ folder. Each OCR tool will output results, including metrics like similarity, accuracy, precision, recall, and F1-score.
-
For further analysis, run the following command to generate performance plots:
python results.py
EasyOCR showed the highest accuracy among the three OCR tools accrding to plots below. The x axis - is number of image, y - metric measure. Preprocessing improved accuracy by around 10-12%, though primarily for EasyOCR. It is because of complexities of dataset and needs to improve preprocessing.
Tesseract OCR Documentation: https://github.com/UB-Mannheim/tesseract/wiki
EasyOCR: A Comprehensive Guide: https://medium.com/@adityamahajan.work/easyocr-a-comprehensive-guide-5ff1cb850168
KesarOCR documentation: https://keras-ocr.readthedocs.io/en/latest/