siamese-transfer-learning

Deep learning frameworks such as Convolutional Neural Networks (CNNs) take a lot of samples to train, even if the lower layers are transferred from a pretrained CNN. In this work, we propose that the ability to compare a pair of images for similarity should also be considered for such transfer learning. We are developing a Siamese network based CNN architecture which compares pairs of images for similarity; the output of the network can be transformed into a Mercer kernel to allow utilization of wide margin classification properties of a SVM that is useful when the amount of training data is scarce.