A Modified Siamese Network for Facial Assimilation

Abstract

We have proposed a face recognition model that uses modified Siamese Networks to give us a distance value that indicates whether 2 images are the same or different. We have used a pre-trained Xception CNN model based on InceptionV3 for Encoder. The Siamese networks take 3 input images (anchor, positive and negative), and Encoder encodes images to their feature vectors. The main objective of this research is to propose a model for face recognition with high accuracy and low classification process time, that is why we have implemented the model using a custom training and testing loop and loss function to be able to compute the triplet loss using three embeddings produced by Siamese Network. The model is trained using batches of triplets, and testing is performed using test triplets. The performance of the proposed model shows high accuracy. Also, the custom loop lowers the computational time during training and testing.