Deep Discriminant Neural Network for Digit Classification

This project implements a deep neural network for classifying handwritten digits. Built with Keras, the model learns to recognise hand-written digits. The network combines convolutional layers with batch normalization and dropout to achieve robust classification performance. Through data augmentation and a carefully structured CNN architecture, it aims to generalise effectively to new handwriting styles.

Implementation

Network Architecture

The network follows a progressive deepening structure, with each block increasing in complexity to learn more sophisticated features.

Block 1

• Two Conv2D layers (32 filters, 3×3 kernel) - Captures basic edges and shapes
• BatchNormalisation after each Conv2D - Stabilises training
• MaxPooling (2×2) - Reduces spatial dimensions and computational load
• SpatialDropout2D (20%) - Prevents feature map co-adaptation

Block 2

• Two Conv2D layers (96 filters, 3×3 kernel) - Learns intermediate-level patterns
• BatchNormalisation after each Conv2D - Maintains consistent feature scaling
• MaxPooling (2×2) - Further dimension reduction
• SpatialDropout2D (20%) - Continues regularisation

Block 3

• Two Conv2D layers (128 filters, 3×3 kernel, 'same' padding) - Identifies complex digit features
• BatchNormalisation after each Conv2D - Normalises deeper features
• MaxPooling (2×2) - Final spatial reduction
• SpatialDropout2D (20%) - Ensures robust feature learning

Dense Layers

• Flatten layer - Converts 2D features to 1D
• Dense layer (1050 units) with L2 regularisation - Rich feature combination
• BatchNormalisation - Stabilises deep network training
• Dropout (50%) - Prevents overfitting
• Output layer (10 units, softmax) - Produces digit probabilities

Data Augmentation

Static Augmentation (Albumentations)

• ElasticTransform - Simulates natural handwriting deformations
• GaussNoise - Adds resilience to image noise
• CoarseDropout - Improves robustness to missing parts
• RandomBrightnessContrast - Handles varying image qualities
• Image Inversion - Adapts to different digit colours

Real-time Augmentation (ImageDataGenerator)

• Rotation - Handles tilted handwriting
• Width/Height shifts - Accounts for different digit positions
• Zoom range - Manages varying digit sizes

Training Strategy

Optimisation

• Adam optimiser with gradient clipping - Prevents explosive gradients
• Initial learning rate: 1e-3 - Balanced between speed and stability
• Batch size: 384 - Provides stable gradient estimates

Training Callbacks

• Early Stopping - Prevents overfitting by monitoring validation loss
• Model Checkpoint - Preserves best model during training
• ReduceLROnPlateau - Adapts learning rate when progress plateaus

Regularisation Techniques

• BatchNormalisation - Stabilises training throughout the network
• SpatialDropout2D - Specifically designed for convolutional features
• Standard Dropout - Prevents dense layer overfitting
• L2 regularisation - Controls weight growth

Dataset Management

• Multiple dataset combination - Increases training diversity
• Image standardisation (28×28) - Ensures consistent input size
• Value normalisation [0,1] - Stabilises network training
• 90-10 split - Provides sufficient validation data
• Fixed random seed - Ensures reproducible results

Tech Stack

• Python - Core programming language
• TensorFlow - Deep learning framework
• Keras - Neural network API
• NumPy - Numerical computing library
• Albumentations - Image augmentation library
• OpenCV - Computer vision library
• PIL (Pillow) - Image processing library
• scikit-learn - Machine learning utilities
• scikit-image - Image processing algorithms
• Matplotlib - Data visualization

Project Setup

You can set up this project using either Conda or Poetry.

Using Conda

Create a new conda environment:

conda create -n digit-classifier python=3.9

Activate the environment

conda activate digit-classifier

Install dependencies

conda install tensorflow-gpu
conda install numpy
conda install -c conda-forge albumentations
conda install -c conda-forge opencv
conda install pillow
conda install scikit-learn
conda install scikit-image
conda install matplotlib

Using Poetry

Install all the project dependencies using:

poetry install

Datasets

Below are the datasets that have been using for this project. Datasets marked with * need to be downloaded manually. Kaggle not implemented.

• MNIST
• *EMNIST
• *USPS
• *Handwritten Digits Dataset (not in MNIST)

Download these and extract them into a folder called data within the project directory. Name the folders with the dataset as shown below:

• emnist <- EMNIST
• handwritten-digits-not-mnist
• usps <- USPS