Handwritten Digit Recognition (C++)

This project explores handwritten digit recognition using basic machine learning algorithms written from scratch in C++.

It is not a finished system. It is a learning-focused implementation.

Overview

The system reads raw dataset files (MNIST format), processes them, and applies a classification algorithm.

Main components:

• Data: Represents a single data sample (image + label)
• DataHub: Handles dataset loading and association
• KNN: Implements k-nearest neighbors classification

Dataset

• MNIST handwritten digit dataset
• Binary IDX file format
• Image and label files are manually parsed

Implemented Algorithm

K-Nearest Neighbors (KNN)

• Distance-based classification
• Uses dataset loaded through DataHub
• Example usage:
  • K = 3
  • Finds nearest samples and predicts label

Project Structure

/Data

• Data representation and parsing

/Data Hub

• Dataset loading and management

/KNN Algorithm

• KNN implementation

/archive

• Dataset files (MNIST)

/main.cpp

• Entry point

How It Works

1. Load dataset paths (train + test)
2. Parse IDX files into memory
3. Associate images with labels
4. Run KNN on test data
5. Predict labels based on nearest neighbors

Notes

• Written for learning purposes
• Hardcoded dataset paths (needs refactor)
• No optimization for large datasets
• Error handling is minimal

Known Issues

• Some unstable behavior during dataset parsing
• No validation for corrupted files
• Performance drops with large input size

TODO

• Remove hardcoded paths (make configurable)
• Improve parsing robustness
• Add dataset normalization
• Optimize KNN (distance calculation, memory usage)
• Add accuracy evaluation (confusion matrix, metrics)

Build

mkdir build
cd build
cmake ..
make
Run

Make sure dataset paths are correctly set inside main.cpp.

Why This Project

The goal is to understand:

• How raw data is handled in ML systems
• How simple algorithms like KNN actually work internally
• Memory and performance constraints in low-level implementations