- Tensorflow 1.14
- OpenCV 4.5.2
All images in the Dataset are extraxted from Sinhala Font
- BhashitaComplex
- Iskoola Potha
- LK-LUG
- Nirmala UI
- Noto Sans Sinhala
- Noto Sans Sinhala Bold
classes = ['ක', 'ඛ', 'ග', 'ඝ', 'ඟ', 'ච', 'ඡ', 'ජ', 'ට', 'ඩ', 'න', 'ණ', 'ත', 'ථ', 'ද',
'ධ', 'ප', 'ඵ', 'බ', 'භ', 'ම', 'ඹ', 'ය', 'ර', 'ල', 'ව', 'ශ', 'ෂ', 'ස', 'හ', 'ළ', 'ෆ']There are 32 classes (letters). For the Training dataset, each class has 6 examples. So Total number of examples in the training dataset is 6x32 = 192. For the Testing dataset, each class has 1 examples. So Total number of examples in the testing dataset is 1x32 = 32.
A data example has 70 rows and 70 columns, then it is resized into 28x28.
Tensorflow default ANN architecture.
<Tensorflow v1> : https://github.com/tensorflow/docs/tree/master/site/en/r1/tutorials
Model: "sequential"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
flatten (Flatten) (None, 784) 0
_________________________________________________________________
dense (Dense) (None, 512) 401920
_________________________________________________________________
dropout (Dropout) (None, 512) 0
_________________________________________________________________
dense_1 (Dense) (None, 32) 16416
=================================================================
Total params: 418,336
Trainable params: 418,336
Non-trainable params: 0
_________________________________________________________________
| Accuracy : 96.88% |
| Loss : 0.9942 |
Limitation: Each letter should have enough margin. In the Contour detection, character height and width should between 32 - 320 pixels. Further: Increase the dataset with more characters