System of comparison of images of faces, in multiple resolutions, based on Siamese Neural Networks

Abstract

In this work a system of face image comparison is presented, capable of being used with variousresolutions, using a method for calculating similarity, which allows to be used, for example, inbiometrics applications. The method works for data sets with many categories and few examples percategory, as well as categories not visualized during the training phase. The main idea is to learn afunction that maps input patterns to a destination space so that the L1 norm in the destination spaceapproximates the "semantic" distance in the input space. The learning process minimizes adiscriminative loss function that converges to a similarity metric, which should be small when the pairsof faces are of the same person and large for pairs of faces of different people. A Siamese architectureof convolutional neural networks, robust to geometric distortions, was used to map the input patternsin the target space. The system has been tested on the CyberExtruder Ultimate Face Matching DataSet, which has 10205 face images.