- Overview
- Documentation
- System Requirements
- Installation Guide
- DCNN DropBlock Algorithm flow
- Dataset
- Results
- Visualization
- Credits
- Citation
- License
Kuzushiji-Dropblock is a Project based on Recognizition of Japanese Historical Image classification.
📝 View our research paper titled "Japanese Historical Character Recognition using Deep Convolutional Neural Network (DCNN) with DropBlock Regularization" availble at (http://dx.doi.org/10.35940/ijrte.b2923.078219)
🚀 Kuzushiji-DropBlock package requires only a standard computer with enough RAM to support the in-memory operations.
This code is supported for Windows, macOS and Linux. The code has been tested on the following systems:
- Windows: Professional (10)
- macOS: Mojave (10.14.1)
- Linux: Ubuntu 16.04
Kuzushiji-DropBlock mainly depends on the Python scientific stack.
Keras
numpy
scipy
pandas
matplotlib
scikit-learn
PyTorch
Tensorflow
DropBlock
In Bash,
pip install DropBlock2D
In Google Colab,
!pip install DropBlock2D
GitClone
git clone https://github.com/sujatasaini/Kuzushiji-DropBlock
cd Kuzushiji-DropBlock
📁 Kuzushiji-MNIST is a drop-in replacement for the MNIST dataset (28x28 grayscale, 70,000 images), provided in the original MNIST format as well as a NumPy format.
Kuzushiji-MNIST contains 70,000 28x28 grayscale images spanning 10 classes (one from each column of hiragana), and is perfectly balanced like the original MNIST dataset (6k/1k train/test for each class).
| File | Examples | Download (MNIST format) | Download (NumPy format) |
|---|---|---|---|
| Training images | 60,000 | train-images-idx3-ubyte.gz (18MB) | kmnist-train-imgs.npz (18MB) |
| Training labels | 60,000 | train-labels-idx1-ubyte.gz (30KB) | kmnist-train-labels.npz (30KB) |
| Testing images | 10,000 | t10k-images-idx3-ubyte.gz (3MB) | kmnist-test-imgs.npz (3MB) |
| Testing labels | 10,000 | t10k-labels-idx1-ubyte.gz (5KB) | kmnist-test-labels.npz (5KB) |
Mapping from class indices to characters: kmnist_classmap.csv (1KB)
it's recommended to download in NumPy format, which can be loaded into an array as easy as:
arr = np.load(filename)['arr_0'].
Kuzushiji-49, has 49 classes (28x28 grayscale, 270,912 images), is a much larger, but imbalanced dataset containing 48 Hiragana characters and one Hiragana iteration mark.
Kuzushiji-49 contains 270,912 images spanning 49 classes, and is an extension of the Kuzushiji-MNIST dataset.
| File | Examples | Download (NumPy format) |
|---|---|---|
| Training images | 232,365 | k49-train-imgs.npz (63MB) |
| Training labels | 232,365 | k49-train-labels.npz (200KB) |
| Testing images | 38,547 | k49-test-imgs.npz (11MB) |
| Testing labels | 38,547 | k49-test-labels.npz (50KB) |
Mapping from class indices to characters: k49_classmap.csv (1KB)
Kuzushiji-Kanji is an imbalanced dataset of total 3832 Kanji characters (64x64 grayscale, 140,426 images of both common and rare characters), ranging from 1,766 examples to only a single example per class.
The full dataset is available for download here (310MB).
💾 You can run python download_data.py to interactively select and download any of these datasets!
You can also download the data from Kaggle and participate in Kaggle Kuzushiji Recognition Competition.
📈 Results of our DCNN Model with different regularization methods (DropBlock, Dropout, SpatialDropout) on MNIST, Fashion-MNIST Kuzushiji-MNIST and Kuzushiji-49, trained on Google Colab average over 3 runs.
| Models | MNIST | Fashion-MNIST | Kuzushiji-MNIST | Kuzushiji-49 |
|---|---|---|---|---|
| DCNN-DropBlock | 99.47% | 93.40% | 97.66% | 95.67% |
| DCNN-Dropout | 97.99% | 85.47% | 86.43% | 95.34% |
| DCNN-Spatial-Dropout | 97.17% | 84.44% | 81.08% | 58.18 |
Have more results to add to the table? Feel free to submit an issue or pull request!
| Models | MNIST | Fashion-MNIST | Kuzushiji-MNIST | Kuzushiji-49 |
|---|---|---|---|---|
| 4-Nearest Neighbor Baseline | 97.14% | 85.97% | 91.59% | 86.00% |
| Naive-Bayes | 98.06% | 86.60% | 92.17% | 88.44% |
| AlexNet | 98.19% | 87.47% | 91.82% | 81.01% |
| Simple CNN | 99.08% | 92.54% | 95.02% | 90.42% |
| Transfer Learning with CNN | 99.34% | 97.46% | 97.06% | 83.96% |
| LeNet-5 | 99.13% | 91.33% | 94.66% | 89.64% |
| MobileNet | 99.20% | 93.04% | 95.09% | 91.06% |
| DCNN-DropBlock | 99.47% | 93.40% | 97.66% | 95.67% |
📊 *Used dropblock with block_size=7 and keep_prob=0.8 over 100 iterations
Example available here
Please cite Kuzushiji-DropBlock in your publications if it helps your research::100:
@article{2019,
doi = {10.35940/ijrte.b2923.078219},
url = {https://doi.org/10.35940%2Fijrte.b2923.078219},
year = 2019,
month = {jul},
publisher = {Blue Eyes Intelligence Engineering and Sciences Engineering and Sciences Publication - {BEIESP}},
volume = {8},
number = {2},
pages = {3510--3515},
title = {Japanese Historical Character Recognition using Deep Convolutional Neural Network ({DCNN}) with Drop Block Regularization},
journal = {International Journal of Recent Technology and Engineering}
}
