Hanzi Vision OCR

Project started 06/26/25
README last updated on 12/30/25
Project last updated on 12/30/25

Authors/Contributors: Keagan Kautzer
Note: This repo will be sparsely updated throughout the academic year, as I focus primarily on my coursework

Overview

This repository is home to my Hanzi OCR Project. It contains five individual subdirectories (model, analysis, frontend, server, and tests) each outlined briefly below.

model

This is the main directory for all model-related data and information. This is where training and evaluation scripts, raw training / evaluation / test data, logs, and previous model weights are stored. The model utilizes a tweaked Inception-v1 (GoogLeNet) architecture for grayscale input and a dynamic number of output classes, as this is the best available trade-off for time and accuracy given the current project constraints. Models have been trained on sets consisting of the most popular 500, 750, 1000, 1500, 2000, 2500, ..., 5000 characters, trained for 100 epochs each. In the future, additional models may be trained using other underlying architectures.

analysis

This is the main directory for all data analysis operations. The main component is a script to generate various plots and tables based on model metadata and training data. Currently, figures are generated for:

• Metadata Table (Name, Architecture, Number of Characters, Highest Accuracy, Highest Accuracy Epoch, and Number of Epochs)
• Number of Characters vs. Highest Accuracy
• Epoch vs. Accuracy (Individual)
• Epoch vs. Accuracy (Combined / Aggregate)
• Epoch vs. Average Accuracy

As deemed necessary, the script may be updated to generate additional figures. Furthermore, a future update will implement necessary p-value calculations to complement the existing figures.

frontend

This is the main directory for all frontend web-interface relevant code. The frontend is a straightforward and intuitive React app - it allows users to either upload and evaluate an image from their computer, or to evaluate a drawing from an on-screen canvas. This directory contains all of the pages, components, routing, and styling seen on the frontend webpage.

server

This is the main directory for behind-the-scenes connections between the frontend interface and the trained models. It is a Python Flask API, allowing for the frontend to quickly and easily communicate with and evaluate images based on trained character models.

tests

This directory contains comprehensive test files, which should be run (and passed) before committing / pushing / merging any changes.

This directory is still a work in progress - the full test set is coming soon!.

Current Goals

The current core objectives for this project are as follows:

1. Create a series of tests to run before commits are made to this repository, ensuring no breaking changes are incorrectly pushed.
2. Build a mobile interface / app (very similar to web interface from frontend)

Future Directions

Some brief future directions for this project are as follows:

1. Image Segmentation
   • This will significantly expand the project's functionality and real-world application, allowing for additional future directions to open up (transcription, translation, etc.)
   • Small updates to the frontend and server may be required (separate single-character and multi-character processing). Otherwise, the interactive process is largely the same.
2. Examine the accuracy impact of incorporating novice-level handwriting into the training data
3. Implement additional architectures to train and evaluate new models on