README (3).md

✍️ Printed Word-Level Script Identification Multi-class (14-Class Model)

initial model for Printed document word-level script separation across 13 Indic languages + English.
The model is designed to classify word images into their respective script categories. i.e. Assamese, Bengali, English, Gujarati, Hindi, Kannada, Malayalam, Manipuri, Marathi, Punjabi, Tamil, Telugu, Urdu, odia

📊 Dataset Overview

• Training samples: ~650560
• Validation samples: ~95909
• (Test set used for evaluation)

---

⚙️ Training Setup

• Model: ResNet-18
• Preprocessing: Custom binarization function applied for improved feature extraction
• Input size: 224 × 224 RGB
• Optimizer: Adam
• Loss function: CrossEntropyLoss
• Epochs: model trained up to 35th epoch (weights shared)

---

📈 Results & Evaluation

The model was evaluated on the test set.
Accompanying this README, you will find PNG visualizations for:

• Confusion Matrix
• Per-class Precision, Recall, F1-Score
• Support vs Correct Predictions per class
• Top Misclassifications

These provide a detailed breakdown of model performance across all 14 classes.

---

📂 Included Files

• model_weights/ → Trained ResNet-18 weights
• wt_35_test_report/ → Evaluation visualizations (confusion matrix, metrics, misclassifications, etc.)
• test.py → Script used to run evaluation

---

🗂️ Class Labels

The model predicts among 14 classes:

Assamese, Bengali, English, Gujarati, Hindi, Kannada, Malayalam, Manipuri, Marathi, Punjabi, Tamil, Telugu, Urdu, Odia

---

📝 Note

This is an initial baseline model trained.
Further improvements can be made by training on the complete dataset and tuning hyperparameters.