CursiveCharacterPrediction

Cursive character (CC) recognition faces challenges owing to multifarious CC writing styles and complex character structures. Therefore, a universal CC identification system is essential. This paper proposes a novel deep learning-based CC prediction model and describes the development and implementation of a corresponding mobile application. The proposed model comprises three primary steps: data collection (in which images are collected and cleaned); image preprocessing (where images are normalized, resized, and augmented); and prediction (involving hyperparameter optimization, regularization, and prediction by deep learning models). Transfer learning from convolutional neural networks such as ResNet, DenseNet, VGG, and EfficientNet is also employed. The model was comparatively evaluated on a dataset comprising Wang Xizhi's Grass Jue Song and Cursive Thousand Characters images. Using DenseNet-201, it exhibited superior performance, with accuracy, precision, recall, and F1-score of 0.92, 0.92, 0.93, and 0.92, respectively. Furthermore, the CC prediction application allows users to input CCs and obtain accurate prediction results.

Getting started

Setup Environment

• Clone the repository: git clone https://github.com/LeeKwonWoo/CursiveCharacterPrediction.git
• Installed the requierement:
  • Python version 3.8.0
  • TensorFlow 2.5.0 library
  • CUDA-Toolkit 11.0
  • Virtual device with Android version 9
  • Model’s name Nexus 5X, API 28
  • Application's SDK was 30

Android Implementation

To explore the Android implementation, open the Cursive_Character/ directory in Android Studio. This directory contains all the code, resources, and build scripts needed to run the mobile application. Make sure your Android Virtual Device (AVD) matches the specified configuration to properly test the app.

Notes on Reproducibility

• Environment Consistency: Ensure you use Python 3.8.0 and TensorFlow 2.5.0 to match the original training conditions.
• Dataset Version: Download the dataset from the provided DOI link.
• Virtual Device Configuration: Use the specified virtual device settings (Nexus 5X, Android 9, API 28) and application SDK version 30 to replicate the environment where the model was deployed and tested.

Dataset

The dataset could be downloaded from Zenodo at the DOI: https://doi.org/10.5281/zenodo.14745650

Model

After downloading the complete folder, run the following:

python PredictServer.py

Reference:

(Not Yet availabel)