LipiNet

A deep learning model for classifying handwritten Devanagari characters and digits using a custom CNN architecture with scaffold injection, residual blocks, channel attention, and linear capsule routing.

---

Overview

This project implements a high-accuracy model for the Devanagari Handwritten Character Dataset (DHCD), which contains 46 classes (36 consonants + 10 digits). The model leverages:

• Dual-path stem (standard + horizontal stroke scaffold)
• Dense residual blocks for feature extraction
• Squeeze-and-Excitation (SE) channel attention
• Linear Capsule Routing for class-discriminative scores
• Gated fusion of capsule and dense head outputs
• Cosine-annealing learning rate schedule and AdamW optimizer

---

Setup

1. Clone the Repository

git clone <your-repo-url>
cd <your-repo-directory>

2. Install Dependencies

pip install -r requirements.txt

Requirements:

• Python 3.8+
• TensorFlow 2.x
• PyTorch
• NumPy

3. Download Dataset

The model expects the DHCD dataset to be available at:

https://archive.ics.uci.edu/dataset/389/devanagari+handwritten+character+dataset

Or, manually download and extract the dataset to ./data/DHCD/.

---

Project Structure

.
├── data/
│   └── DHCD/
│       ├── Train/
│       └── Test/
├── results/
│   └── (saved model checkpoints)
├── README.md
└── train.py

---

Training

Configuration

• num_classes: 46 (36 consonants + 10 digits)
• image_size: 32x32 (grayscale)
• batch_size: 64
• epochs: 100
• lr: 5e-4 (peak learning rate)
• weight_decay: 1e-4 (AdamW L2 regularization)
• label_smoothing: 0.1
• val_split: 0.1

Run Training

python train.py

• Models are saved to ./results/ as .keras files.
• Training uses cosine-annealing LR schedule and early stopping (patience=15).

---

Model Architecture

Key Components

1. Stem:

   • Dual-path: Standard 3×3 conv + horizontal (1×5) conv for Devanagari stroke detection.
   • Channel attention (SE block) for feature recalibration.

2. Encoder:

   • 3 stages of dense residual blocks (64→128→256 channels).
   • Scaffold injection: Horizontal stroke features are pooled and added to each encoder stage.

3. Decoder Head:

   • Linear Capsule Routing: Projects multi-scale features into capsule space.
   • Dense Head: Parallel classification head.
   • Gated Fusion: Learns to blend capsule and dense outputs.

4. Optimization:

   • AdamW optimizer with weight decay.
   • Cosine-annealing LR schedule (decays to 1e-6).
   • Label smoothing for better calibration.

---

Results

Model	Test Accuracy	Macro F1	Parameters	Test Loss
LipiNet	99.75%	99.75%	(3,863,880) ~3.9M	0.703

Note: Results are run with seeds [42] for reproducibility.

---