Arabic-handwritten-one-shot-learning

Implementation of Siamese network for handwritten Arabic characters using Keras.

Dataset

• The dataset is available on Kaggle here.
• I have used 60% of the characters (16 characters) for training and used the other 40% for testing (12 characters).
• I chose 20 images for each class (character) randomly using different seeds and the performance was similar for each seed.

Performance

• The validation was done using N-way one shot learning over 500 trials with N in range (2-12).
• I have also tested it against Nearest neighbour.
• The training and validation were done using GPU and Google Colab.
• Model weights are available here with 92.8% accuracy on test data.

References

• One shot learning with Siamese Netwirk using Keras blog
• Siamese explanation line by line blog
• Siamese Neural Networks for One-shot Image Recognition