Tamil Handwritten Character Recognition

This repository implements multiple deep learning architectures for recognizing Tamil handwritten characters.
The dataset consists of 156 character classes, preprocessed into 64×64 grayscale images.

The project benchmarks different neural network architectures — Custom CNN, DenseNet, GoogLeNet (HCCR-inspired), and Capsule CNN — and compares their performance using consistent preprocessing, training, and evaluation strategies.

---

Results Summary

Model	Accuracy (%)
CNN	96.94
DenseNet	96.87
HCCR	96.14
CapsNet	93.10

Training Curves

CNN

DenseNet

HCCR

CapsNet CNN

Features

• Custom Dataset Loader – Uses CSVs mapping filenames to labels.
• Image Preprocessing – Resizing, grayscale conversion, normalization, and augmentations (rotation, affine transforms, scaling).
• Models Implemented
  • implementation_cnn.py → Custom 5-block CNN
  • densenetmodel.py → DenseNet-121 (adapted for 64×64 grayscale)
  • hccrmodel.py → GoogLeNet (HCCR-inspired with Inception blocks)
  • capsnetcnn.py → Capsule CNN with squashing nonlinearity
• Training Utilities
  • Optimizer: Adam (lr=1e-3 or 1e-4)
  • Loss: CrossEntropyLoss
  • Learning Rate Scheduler: ReduceLROnPlateau – halves learning rate when validation loss plateaus
  • Early Stopping: Stops training after 10 patience epochs
• Evaluation & Visualization
  • Accuracy and classification reports (precision, recall, F1)
  • Confusion matrix with top-10 confused classes heatmap
  • Loss & accuracy curves across epochs
  • Prediction samples with confidence scores

---

Tech Stack

• Frameworks: PyTorch, Torchvision
• Data & Evaluation: NumPy, Pandas, Scikit-learn
• Visualization: Matplotlib, Seaborn
• Progress Tracking: tqdm

---

Training & Evaluation Workflow

Data Loading

• Reads FileNames and Ground Truth from CSV files.
• Custom HandwritingDataset class wraps image loading and transforms.

Training

• Each script defines a train_model() function.
• Training settings:
  • Epochs: 200
  • Batch size: 32
  • Optimizer: Adam
  • Scheduler: ReduceLROnPlateau (reduces LR if validation loss stalls)
  • Early stopping: patience = 10

Evaluation

• Best model checkpoint auto-saved (e.g., bestcnn_model_scheduler.pth).
• Evaluation outputs:
  • Test accuracy
  • Classification report
  • Confusion matrix
  • Top misclassifications
  • Prediction confidence scores

---

Sample Predictions

CNN Model (96.94%)

Sample	True Label	Predicted Label	Confidence
0	94	94	1.0000
1	81	81	1.0000
2	91	91	1.0000
3	44	44	1.0000
4	120	120	0.9989
5	134	134	0.9309
6	14	15	0.6577
7	148	148	1.0000
8	30	30	1.0000
9	31	31	1.0000

DenseNet Model (96.87%)

Sample	True Label	Predicted Label	Confidence
0	94	94	1.0000
1	81	81	1.0000
2	91	91	1.0000
3	44	44	1.0000
4	120	120	0.9983
5	134	134	0.6897
6	14	14	1.0000
7	148	148	1.0000
8	30	30	1.0000
9	31	31	1.0000

HCCR Model (96.14%)

Sample	True Label	Predicted Label	Confidence
0	94	94	0.9999
1	81	81	0.9999
2	91	91	1.0000
3	44	44	1.0000
4	120	120	0.9972
5	134	134	0.9105
6	14	14	0.7637
7	148	148	1.0000
8	30	30	0.9999
9	31	31	1.0000

CapsNet CNN Model (93.10%)

Sample	True Label	Predicted Label	Confidence
0	94	94	0.9997
1	81	81	1.0000
2	91	91	1.0000
3	44	44	1.0000
4	120	120	0.8329
5	134	134	0.9634
6	14	15	0.8758
7	148	148	1.0000
8	30	30	1.0000
9	31	31	1.0000

Dataset Source and Reference Paper

• Dataset: https://www.kaggle.com/datasets/faizalhajamohideen/uthcdtamil-handwritten-database
• N. Shaffi and F. Hajamohideen, "uTHCD: A New Benchmarking for Tamil Handwritten OCR," in IEEE Access, vol. 9, pp. 101469-101493, 2021, doi: 10.1109/ACCESS.2021.3096823.
• For more details about the dataset, Please refer the above paper.