Handwritten Text Recognition using EMNIST Datasets

This project demonstrates handwritten text recognition using deep learning on two different datasets from the EMNIST (Extended MNIST) dataset collection:

• EMNIST ByClass (62 classes: digits + uppercase + lowercase letters)
• EMNIST Balanced (47 classes: balanced mix of digits and letters)

Both models share the same CNN-based architecture and are evaluated on accuracy, loss, precision, recall, F1-score, and ROC-AUC metrics.

---

🧠 Model Architecture

• Deep Convolutional Neural Network (CNN)
• Input size: 28x28 grayscale images
• Optimizer: Adam
• Loss function: Sparse Categorical Crossentropy
• Metrics: Accuracy, Top-5 Accuracy, Precision, Recall, F1-Score, ROC-AUC

---

📊 Dataset Overview

EMNIST ByClass

• Training Samples: 697,931
• Test Samples: 116,322
• Total Samples: 814,253
• Unique Classes: 62 (includes digits, uppercase & lowercase letters)

EMNIST Balanced

• Training Samples: 112,799
• Test Samples: 18,799
• Total Samples: 131,598
• Unique Classes: 47 (balanced distribution)

---

🧪 Model Evaluation

EMNIST ByClass Model

• Test Accuracy: 87.38%
• Test Loss: 0.3392
• Top-5 Accuracy: 99.75%
• Precision (macro): 0.7781
• Recall (macro): 0.7402
• F1-Score (macro): 0.7301
• ROC-AUC (macro): 0.9965
• Epochs: 30
• Model Parameters: 502,686

EMNIST Balanced Model

• Test Accuracy: 88.52%
• Test Loss: 0.3216
• Top-5 Accuracy: 99.52%
• Precision (macro): 0.8951
• Recall (macro): 0.8855
• F1-Score (macro): 0.8828
• ROC-AUC (macro): 0.9972
• Epochs: 20
• Model Parameters: 498,831

---

🆚 Comparison & Recommendation

Metric	EMNIST ByClass	EMNIST Balanced
Accuracy	87.38%	88.52% ✅
Top-5 Accuracy	99.75% ✅	99.52%
Precision (macro)	0.7781	0.8951 ✅
Recall (macro)	0.7402	0.8855 ✅
F1-Score (macro)	0.7301	0.8828 ✅
ROC-AUC (macro)	0.9965	0.9972 ✅
Training Time	30 epochs	20 epochs ✅
Classes	62 (More) ✅	47

Verdict:

• Use EMNIST ByClass if you need full coverage of characters (digits + upper/lowercase letters).
• Use EMNIST Balanced if you want higher accuracy and better performance on a more focused dataset.

---

🖊️ Canvas App

A custom Python-based Tkinter Canvas GUI is included, allowing you to draw characters or words, and the model will predict the handwritten text in real-time using the trained model.

• Uses the trained model to classify drawn characters

---

✨ Features

• Real-time Drawing: Draw characters directly on the canvas.
• Image Loading: Load handwritten images from your files.
• Character Segmentation: Automatically detects and separates individual characters from a word or sentence.
• Confidence Scoring: Displays model prediction confidence for each recognized character.
• Visual Feedback: Shows segmented character images alongside predictions for better understanding.
• Custom Label Mapping: Supports digits (0-9), uppercase letters (A-Z), and select lowercase letters (a, b, d, e, f, g, h, n, q, r, t).

---

🚀 Quick Start

1. Install dependencies

   pip install -r requirements.txt
   

✨ Features

• Real-time Drawing: Draw characters directly on the canvas.
• Image Loading: Load handwritten images from your files.
• Character Segmentation: Automatically detects and separates individual characters from a word or sentence.
• Confidence Scoring: Displays model prediction confidence for each recognized character.
• Visual Feedback: Shows segmented character images alongside predictions for better understanding.
• Custom Label Mapping: Supports digits (0-9), uppercase letters (A-Z), and select lowercase letters (a, b, d, e, f, g, h, n, q, r, t).

---

🚀 Quick Start

1. Install dependencies

   pip install -r requirements.txt

2. Place your trained EMNIST model Put your .h5 model file here:

   model/emnist_balanced_cnn_best.h5
   

3. Run the application

   python handwriting_app.py

4. Start drawing or load an image Use the canvas or load an image to recognize handwritten text!

---

🛠️ Tech Stack

• GUI: Tkinter (Python built-in)
• ML Framework: TensorFlow / Keras
• Image Processing: OpenCV, PIL
• Model: CNN trained on EMNIST Balanced dataset
• Supported Characters: 0-9, A-Z, a, b, d, e, f, g, h, n, q, r, t

---

📊 Model Support

This canvas app is designed for CNN models trained on the EMNIST Balanced dataset.

• Supported Character Classes: 47
• Character Types: Digits, Uppercase Letters, and select Lowercase Letters
• Optimized for recognition in real-time use cases with clean segmentation and confidence visualization.

---

---