Siamese neural networks for biometric hashing

Official URL: http://risc01.sabanciuniv.edu/record=b1589820 (Table of Contents)

In this project we propose a new method for biometric hashing. In our method we use a deep neural network structure to learn a similarity preserving mapping. For this purpose we train a neural network architecture that consists of two identical neural nets called Siamese neural nets where each one performs the mapping for hashing. The weights are tuned in training such that two di erent biometric data of a person yield a similar code but codes corresponding to di erent subject's images are far away. The neural network outputs a pre-hash vector which is then converted to a biometric hash vector through random projection and thresholding. We use angular distance measure to train pre-hash vectors which is more related with the Hamming distance that is used between hashes in veri cation time. Our experiments show that the introduced method can outperform a PCA-based baseline. We also show that a biometric hashing system which was trained using the angular distance can achieve better verification rates than another one trained with the Euclidean distance.