One-shot deformed face recognition via Siamese neural network

Abstract

CNN network classes require multiple images per class to train. This makes facial recognition using CNN imprac- tical, as it is often hard to obtain a sufficient number of images of one person. Siamese Networks, on the other hand, uses oneshot learning, meaning that only one input image will be needed to train the network for each person. We build a facial recognition system using Siamese Network. In Siamese Networks, a single image of one person is input, and the network will learn to recognize the person by learning the embedding of the image. The embedding is used to compute a similarity score – similar images will have higher similarity scores. Another image will then be input to the same network, and the system will compare two embeddings to determine whether they contain the same person, giving a true or false output. Using the ORL and LFW dataset, we performed several experiments on multiple aspects of the Siamese Network. We experimented on the Random Erasing function for our augmented data to test the reliability of the network in facial recognition. Results show significant improvement on model accuracy for model trained on random erasing masking. This kind of facial recognition systems is versatile and can be applied to numerous use cases. For example, this kind of system can be used to provide facial recognition for persons with disability that manifests in the deformation of facial features.