Tausia01/bangla-compound-character-classification

This project was built for my Neural Networks course, where the goal was to classify 120 different Bangla handwritten compound characters. To tackle this, I used a convolutional neural network (CNN) with a multi-stage self-attention mechanism to better focus on important spatial features during classification. # Bangla Compound Character Classification *Using CNN with Multi-Level Self-Attention on the MatriVasha Dataset* This is my final project for the Neural Networks course. The goal was to classify 120 different **Bangla handwritten compound characters** using a custom-built deep learning pipeline that combines a CNN backbone with attention mechanisms. I also included heatmap visualizations to show where the model was focusing during classification. --- ## Overview This project is split into two main stages: 1. **Pretraining**: A convolutional neural network (CNN) is trained on CIFAR-100 (grayscale) to learn general image features. 2. **Finetuning**: The pretrained model is extended with **multi-level self-attention modules** and fine-tuned on the MatriVasha dataset of Bangla handwritten compound characters. Due to **resource limitations** (running on Google Colab), training was done for only a few epochs — but even then, the model produced solid results and insightful attention maps. --- ## Dataset This project uses the **MatriVasha** dataset, a large-scale handwritten Bangla compound character dataset containing 120 classes collected from male and female contributors. - 🔗 Download the dataset on Mendeley Data **Citation**: Ferdous, Jannatul; Karmaker, Suvrajit; Rabby, AKM Shahariar Azad; Hossain, Syed Akhter (2021), *“MatriVasha: Bangla Handwritten Compound Character Dataset and Recognition”*, Mendeley Data, V1, doi.org > Dataset was manually preprocessed by merging male/female folders, resizing to 128×128, inverting pixel intensities, and normalizing to [-1, 1]. --- ## Model Architecture ### 📌 Stage 1: `CustomCNN` (Pretraining) - 4-layer convolutional network - Trained on grayscale CIFAR-100 - Used `AdamW` optimizer and Xavier initialization ### 📌 Stage 2: `CNNEncoderWithAttention` (Finetuning) - Self-attention modules applied on three CNN layers: - 32×128×128 features - 64×64×64 features - 128×32×32 features - Each att …