Hanacaraka Recognition HD-CNN

Image recognition for Javanese script using YOLOv4 Darknet and HD-CNN. This project uses YOLOv4 as the object detector, and each detected object will be classified by HD-CNN.

Environment

• CentOS Stream release 9
• CUDA Toolkit 11.8
• cuDNN 8.6
• python 3.8.16
• tensorflow 2.12.0
• opencv 4.6.0

Table of Contents

• Hanacaraka Recognition HD-CNN
• Environment
• Table of Contents
• Usage
  • Installation
  • Training and Testing YOLOv4 Darknet
  • Training and Testing HD-CNN
  • Run YOLOv4 + HD-CNN
• Acknowledgements

Usage

Installation

git clone https://github.com/jansen062001/Hanacaraka-Recognition-HD-CNN.git

Training and Testing YOLOv4 Darknet

1. Preparing The Dataset

   • Download and unzip the augmented dataset with YOLO Darknet format from roboflow: https://universe.roboflow.com/thesis-dicgg/hanacaraka-recognition

     Hanacaraka YOLOv4 Darknet.v14i.darknet
     │   README.dataset.txt
     │   README.roboflow.txt
     │
     └───train
           ...
           9_png.rf.f7a6d330b72103e36cc779f7a2c5d075.jpg
           9_png.rf.f7a6d330b72103e36cc779f7a2c5d075.txt
           _darknet.labels

   • Because our YOLOv4 model will use 416x416 (WxH), so each image in the dataset needs to be sliced into 416x416. Use this Github repo to do this work.
   • Copy all sliced images (train and test) and classes.names into ./yolov4_darknet/dataset/raw/
   • Rename classes.names to classes.txt
   • Run these commands to re-label and move the dataset into ./yolov4_darknet/dataset/processed/

     python -m yolov4_darknet.generate_yolo_dataset --train_size=70 --valid_size=20 --test_size=10

2. Training

   python -m yolov4_darknet.train

   .weights file from the training process will be placed on ./yolov4_darknet/weights/

3. Testing

   • Put the image file in the directory ./yolov4_darknet/data/

   • Run these commands

     python -m yolov4_darknet.test --width=416 --height=416 --filename=example.jpg

     There is a file called output_example.jpg in the directory ./yolov4_darknet/data/ as the output

Training and Testing HD-CNN

1. Preparing The Dataset

   • Download and unzip the augmented dataset with YOLO Darknet format from roboflow: https://universe.roboflow.com/thesis-dicgg/hanacaraka-recognition

     Hanacaraka YOLOv4 Darknet.v14i.darknet
     │   README.dataset.txt
     │   README.roboflow.txt
     │
     └───train
           ...
           9_png.rf.f7a6d330b72103e36cc779f7a2c5d075.jpg
           9_png.rf.f7a6d330b72103e36cc779f7a2c5d075.txt
           _darknet.labels

   • Copy all files inside the train folder into ./hd_cnn/dataset/raw/
   • Rename _darknet.labels to classes.txt
   • Run these commands to re-label and move the dataset into ./hd_cnn/dataset/processed/

     python -m hd_cnn.generate_hdcnn_dataset --train_size=70 --valid_size=20 --test_size=10

2. Training

   python -m hd_cnn.train --model=hd_cnn

   .weights file from the training process will be placed on ./hd_cnn/weights/

3. Testing

   • Put the image file in the directory ./hd_cnn/data/
   • Run these commands

     python -m hd_cnn.test --filename=example.jpg

Run YOLOv4 + HD-CNN

• Put the image file in the directory ./
• Run these commands

  python main.py --filename=example.jpg

• After the process is complete, there is a file called result.jpg in the directory ./ as the final output

Acknowledgements

Expand

• https://github.com/AlexeyAB/darknet
• https://github.com/slanj/yolo-tiling
• https://github.com/justinessert/hierarchical-deep-cnn
• https://github.com/satyatumati/Hierarchical-Deep-CNN