Handwritten Vietnamese Recognition

Failure is only the opportunity to begin again. Only this time, more wisely. Good times become good memories, but bad times become good lessons. - Uncle Iroh

Motivation

Not many people know that my name is not Thu, but it's actually Thư, and the ways these two words are pronounced are also different. Vietnamese written alphabet is a Latin-based script but it has various dialectical marks that make Vietnamese a very nuanced language. My motivation for pursuing this project is: 1) I love my mother tongue language because the nuances make it so special and beautiful, and 2) I believe in the meaningful applications of handwriting recognition in many settings, such as education, clinical settings, banking, etc.

The repo has 3 folders, each consisting of different Vietnamese datasets (one produced by me, which is a huge limitation because the test set is too similar to the train set that the models' accuracy is really high, yet not too useful). Below is the description of the models I used for the self-produced dataset.

Pipeline 1: Classify variations of the letter A - A, Ă, Â

Data Acquisition

The dataset consists of 300 samples of handwritten Vietnamese letters which belong to 3 classes: a, ă, â. Each class consists of 100 samples. I wrote all these letters on regular A4 papers and used my phone to scan them through CamScanner. The original pictures were cropped into squared frames.

Data Preprocessing

1. Resizing
2. Binarization
3. Normalization

Feature Extraction

1. Pixel Features

Model Selection

1. K-Nearest Neighbor Classifier
2. Naive Bayes
3. Multinomial Logistic Regression
4. Neural Networks

Pipeline 2: Classify variations of 12 letters - A, Ă, Â, D, Đ, E, Ê, O, Ô, Ơ, U, Ư

The second pipeline implemented Convolutional Neural Networks (CNNs) and Transfer Learning (MobileNet, ResNet18, VGG-16)

Pipeline 3: Generate 12 letters - A, Ă, Â, D, Đ, E, Ê, O, Ô, Ơ, U, Ư

The final pipeline implemented different types of autoencoders to create the letters