Convolution Neural Network for Face Similarity Estimation

Abstract

We present a convolution neural network used to determine face similarity given two images as input, thus face identification task. The main focus is on the shape of the input data. We propose schemes where two pictures are connected in four different ways. The input sample is concatenated horizontally and vertically, giving the first two schemes. The other two input shapes include the intertwining by column and by row. Analysis of precision versus recall has been provided for each input schema. Some of traditional approaches focus on deriving feature vectors of an individual, and then the obtained vectors are compared with each other. Our paper offers a new approach to face identification problems where two images of an individual are directly feed to the neural network. Then it is directly the task of the neural network to determine the similarity score.