Enhancing Convolutional Neural Networks for Face Recognition with\n Occlusion Maps and Batch Triplet Loss

Abstract

Despite the recent success of convolutional neural networks for computer\nvision applications, unconstrained face recognition remains a challenge. In\nthis work, we make two contributions to the field. Firstly, we consider the\nproblem of face recognition with partial occlusions and show how current\napproaches might suffer significant performance degradation when dealing with\nthis kind of face images. We propose a simple method to find out which parts of\nthe human face are more important to achieve a high recognition rate, and use\nthat information during training to force a convolutional neural network to\nlearn discriminative features from all the face regions more equally, including\nthose that typical approaches tend to pay less attention to. We test the\naccuracy of the proposed method when dealing with real-life occlusions using\nthe AR face database. Secondly, we propose a novel loss function called batch\ntriplet loss that improves the performance of the triplet loss by adding an\nextra term to the loss function to cause minimisation of the standard deviation\nof both positive and negative scores. We show consistent improvement in the\nLabeled Faces in the Wild (LFW) benchmark by applying both proposed adjustments\nto the convolutional neural network training.\n