Hiragana-Recognition

The project is aimed to the Neural Network course project, which compares the state-of-the-art image classification networks for hand-written hiragana characters.

The experiment will be divided into two parts: single character recognition and multi-characaters recognition. The second task will localize the characters in the image and identify them.

Usage

1. Install all libraries required in "requirements.txt"
2. Change the training configuration in "config.py"
3. Use the following command to try the training process.

python inference.py --train --vgg

• To continue training process, run the following command:

python inference.py --train --vgg --load /path/to/checkpoint/file.tar

• Available models:

  • --vgg: Use VGG 19 with Batch Normalization. Minimum input size 224x224x3
  • --inception: Use Inception V3 Net. Minimum input size 299x299x3, batch size must > 1.
  • --simple: Use customized CNN.

• Available modes:

  • --train: Start training process.
  • --test: Start testing process.

• Available inputs:

  • --dataset: Path to the dataset.
  • --load: path to the saved 'tar' checkpoint file.

Dependencies

Please refer to the "requirements.txt" file in the project folder.

1. torch~=1.4.0
2. numpy~=1.18.2
3. tqdm~=4.45.0
4. torchvision~=0.5.0
5. scikit-learn~=0.22.2.post1
6. matplotlib~=3.2.1
7. opencv-python~=4.2.0

Results

\	VGG 19 with BN	Inception v3	Customized CNN
Input Size	224x224x3	299x299x3	50x50x1
Normalized	True	True	True
Batch Size	8	8	8
Learning Rate	1e-3	1e-3	1e-3
Train/Test Split	85%/15%	85%/15%	85%/15%
Epochs	30	30	30
Parameters	126,001,031	24,543,342	1,712,103
Validation Loss	0.7290	0.2385	0.6475
Validation Accuracy	98.74%	99.5%	98.47%
Inference Time	0.0294s/it	0.0189s/it	0.0020s/it

VGG 19 with BN	Inception v3	Customized CNN

Training Accuracy	Training Loss

Validation Accuracy	Validation Loss