Shape similarity identification with siamese network

Learning from a few samples is a difficult task in machine learning. The siamese network architecture provides a computationally reasonable way to check the similarity of sample pairs.

This implementation provides an example to test this concept, by generating a sample of pairs of similar and non-similar input images.

Steps:

1. Create a toy database, generate different shapes pairs label logic: 1 - similar shapes ; 0 - different shapes
2. Visualize data
3. Create the siamese model
4. Test similarity with the reference image

Model:

Outputs:

With high similarity:

With low similarity, below the threshold:

References

1. Siamese Neural Networks for One-shot Image Recognition, G Koch et al
2. Few-Shot Learning with Siamese Networks and Label Tuning T Muller arXiv:2203.14655 [cs.CL]

/Enjoy.