Image Segmentation and Optical Character Recognition

This project was done by Chandan R T as a part of E9 241: Digital Image Processing course offered at IISc Bangalore in fall of 2022 GitHub repository link: https://github.com/chanrt/image-segmentation-ocr

Instructions:

1. The input image must be placed inside the 'inputs' folder.
2. The main() function in main.py must be called with the name of the input image
3. All necessary corrections can be turned off or on, inside the main() function of main.py
4. The main() function will display and return the predicted string of characters
5. Parameters can be further tweaked in the settings.py file

How to carry out rigorous evaluation:

1. The evaluator() function in evaluation.py must be called with the name of the input image
2. The input image must be present in the 'inputs' folder
3. The transcribed text of the input image must be present in the 'expected_outputs' folder, with the same file name as the input image
4. The evaluator() function will display the Levenshtein distance between the program output and the expected output, as well as the error

---

Example

Input image:

Program output: in thss classic text george pblya l18871985j offers something unique a set ofstratekies foi solving mathelnatical problems the heutistic theotetical appioach based on a deep analysis of the methods and wles of discovery and invention ptoved an inspiration to a geneiation of teachers and students yet the lessons ate uttetiy practical pblya ptslliantly demonstiates how the true mathematician learns to diaw unexpected analo gies tackle ptobiems ftom unusual angles and extiact a little more information fiom the data tiaditional mathematics gan often seem just a piocess of dcy rigotous dedugtion holv to solbe it wondeifully conveys its challenge and excitement as a problemsolmng activity

Evaluations: Levenshtein distance: 47 Error: 6.79 %

---

Models: there are four trained models here

1. A Dense Neural Network trained on handwritten data
2. A Dense Neural Network trained on printed characters data
3. A Convolutional Neural Network trained on handwritten data
4. A Convolutional Neural Network trained on printed characters data CNNs generally perform better You can select the desired model in the arguments of character_recognizer() function in main.py

Datasets:

1. Hand-written characters dataset: Cohen, G., Afshar, S., Tapson, J., & van Schaik, A. (2017). EMNIST: an extension of MNIST to handwritten letters
2. Printed characters dataset: T. E. de Campos, B. R. Babu and M. Varma (2009). Character recognition in natural images These datasets are not included in the repository due to space concerns. You can download these datasets from the links given in the dataset_links.txt file inside the 'data' folder