Classical Computer Vision + OCR · Public GitHub + Kaggle Project

Handwriting OCR

HOG Features · KNN Classification · EMNIST Letters

A classical OCR pipeline for handwritten English letter recognition using HOG feature extraction and a distance-weighted KNN classifier, extended with character segmentation and custom handwritten word reconstruction.

RoleDeveloper

Year2025

StatusPublic GitHub + Kaggle Project

Handwriting OCR HOG and KNN project cover

OVERVIEW

A classical OCR pipeline built around shape features

This project recognizes handwritten English letters using Histogram of Oriented Gradients (HOG) features and a K-Nearest Neighbors classifier. EMNIST Letters provides the official evaluation set, while an additional isolated-character dataset enriches training.

The notebook extends isolated-character recognition into a practical OCR workflow by segmenting custom handwritten word images into letters, classifying each character, and reconstructing the predicted word.

DATASET & FEATURES

103,504Training letter images after enrichment

14,800Official EMNIST test images

1,296HOG features per image

7KNN neighbors, distance weighted

RECOGNITION PIPELINE

Word image → character segmentation → HOG → KNN → text

01Character Segmentation

Custom word images are separated into individual handwritten character regions.

02HOG Feature Extraction

Each normalized 28×28 letter is converted into a 1,296-dimensional gradient-orientation descriptor.

03KNN Classification

A distance-weighted KNN model with 7 neighbors predicts the corresponding English letter class.

04Word Reconstruction

Predicted characters are placed back in reading order to form the recognized word.

OFFICIAL EVALUATION

The official score is measured on the EMNIST Letters test split. The additional handwritten-character dataset is used to enrich training rather than replace the official test set.

PROJECT GALLERY

From benchmark letters to real handwritten words

The gallery combines outputs from the completed Kaggle notebook with a real handwriting demonstration sheet, showing both character-level evaluation and end-to-end word reconstruction.

Multiple handwritten words with character boxes and OCR annotations — Real handwriting examples · segmented words and reconstructed text

Sample handwritten letters from the EMNIST Letters dataset — EMNIST Letters · sample training characters

Confusion matrix for the HOG and KNN handwriting recognition model — Official evaluation · EMNIST Letters confusion matrix

OCR result for the handwritten word Dream — Custom word test · Dream

OCR result for the handwritten word Freelance — Custom word test · Freelance

OCR result for the handwritten word Puzzle — Custom word test · Puzzle

WHAT THIS PROJECT DEMONSTRATES

Feature engineering and classical ML remain useful for OCR

The project emphasizes a complete classical computer-vision workflow: image preparation, handcrafted HOG features, KNN classification, benchmark evaluation, segmentation logic, and visual inspection of errors on custom handwriting.

TECH STACK

PythonOpenCVHOGKNNscikit-learnEMNIST

PUBLIC PROJECT

Code, trained model release, and full notebook.

The GitHub repository contains the lightweight demo structure and links to the pre-trained KNN + HOG model release. Kaggle contains the complete training and evaluation notebook.

View GitHub Repository View Kaggle Notebook Back to top