https://github.com/meetps/tf-devanagri
Convolutional Neural Networks to identify Devanagri characters implemented in Tensorflow
https://github.com/meetps/tf-devanagri
character-recognition convolutional-neural-networks tensorflow
Last synced: about 2 months ago
JSON representation
Convolutional Neural Networks to identify Devanagri characters implemented in Tensorflow
- Host: GitHub
- URL: https://github.com/meetps/tf-devanagri
- Owner: meetps
- License: mit
- Created: 2016-10-05T20:14:04.000Z (almost 10 years ago)
- Default Branch: master
- Last Pushed: 2017-02-20T14:53:56.000Z (over 9 years ago)
- Last Synced: 2025-02-23T01:34:51.668Z (over 1 year ago)
- Topics: character-recognition, convolutional-neural-networks, tensorflow
- Language: Python
- Size: 17.3 MB
- Stars: 1
- Watchers: 2
- Forks: 1
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# tf-devanagri
Convolutional Neural Network to identify Devanagri characters implemented in Tensorflow.
### Data
* n_classes = 104
* Size = 320x320
* Format = Binary, PNG
### Results and Preprocessing
* 94.6 +/- 0.07 % accuracy on test data.
* 30 epochs in 30 mins of training time on a 4 GB NVIDIA GT920M GPU and i7 CPU.
* 10x downsampling to get 32x32 images.
* 3-iterations of binary dilation to make characters thicker.
### Network
#### Architecture
* Conv2d 7x7 1
* Conv2d 5x5 32
* Conv2d 5x5 64
* Conv2d 3x3 128
* Conv2d 3x3 256
* Dense 256x2x2
* Dense 1024
* Softmax 104 = `n_classes`
#### HyperParameters
* Learning Rate = 0.003
* Momentum = 0.9