Handwritten Digit Classifier (Custom SVM)

01 / OVERVIEW

The idea behind it.

End-to-end machine learning project where the training data is entirely self-collected handwriting samples of digits 0–9. Captured samples using a drawing canvas, preprocessed the images (grayscale, thresholding, normalization, HOG feature extraction), then trained a multi-class Support Vector Machine with a tuned RBF kernel. The result is a classifier that recognizes my specific handwriting style. Includes a live inference interface where you draw a digit and the model predicts it in real time.

02 / FEATURES

What it does.

• 01Self-collected training dataset of handwritten digits 0–9
• 02Image preprocessing pipeline: grayscale, thresholding, resize, normalization
• 03HOG (Histogram of Oriented Gradients) feature extraction
• 04Multi-class SVM with RBF kernel and hyperparameter tuning via GridSearchCV
• 05Live drawing canvas for real-time digit prediction
• 06Confusion matrix and per-class accuracy visualization

03 / CHALLENGES

The problems worth solving.

• 01Collecting enough varied handwriting samples per class for generalization
• 02Choosing and tuning the right feature descriptor (HOG vs raw pixels)
• 03Handling class imbalance across digits in the self-built dataset
• 04Tuning C and gamma hyperparameters without overfitting to personal writing style

04 / LEARNINGS

What I took away.

• 01Full ML pipeline: data collection → preprocessing → feature engineering → training → evaluation
• 02How SVMs use kernel functions to separate non-linearly separable classes
• 03HOG feature extraction and why spatial gradient histograms outperform raw pixels for digit recognition
• 04Hyperparameter search with GridSearchCV and cross-validation
• 05Bias introduced by training on a single person's handwriting
• 06OpenCV image preprocessing techniques for ML input preparation