基於生成對抗式網路的人臉影像身分重建

Abstract

基於卷積神經網路的人臉辨識技術已經達成極高的準確率並且廣泛應用於各種場域，然而在特定的應用場合人臉辨識技術還是有很大的挑戰，特別是影像品質不佳的監控設備環境下，人會與鏡頭有較大的距離，導致人臉影像解析度過低造成人臉身分難以辨識，為此我們提出一個新穎的基於生成對抗式網路的人臉影像重建網路，藉由學習低解析度的人臉影像與高解析度的人臉影像之映射關係，我們可以從低解析度人臉影像重建出高解析度人臉影像，此外我們使用Feature Embedding的方法從網路的輸出層得到人臉身分特徵，並且利用三元組損失計算人臉身分特徵用以訓練神經網路，使神經網路可以藉由人臉身分特徵表達做相應的高解析度人臉影像重建，實驗階段我們使用了公開的人臉資料集CASIA-WebFaces進行訓練，並與近年來基於深度學習所提出的底解析度人臉重建方法作為比較並稱為傳統方法。實驗結果表明我們所提出的極低解析度人臉重建網路在大倍率放大的影像品質與辨識率優於傳統方法。

Recently, face recognition technology based on convolutional neural networks has achieved extremely high accuracy and has widely used in various fields, but face recognition technology still has great challenges in specific applications, especially in the environment of surveillance equipment with poor image quality, people usually have a large distance from the lens, resulting in low face image resolution and difficult recognition of the face identity. For this reason, we propose a novel generative adversarial networks-based face hallucination framework for producing high-resolution face images from very low-resolution ones, by training the network to find the mapping relationship between low-resolution face images and high-resolution ones, we can hallucinate corresponding high-resolution faces, and Embedding's feature adopted to extract face identity features from the output layer of the network, and a triplet-based identity preserving loss for extracting identity-aware facial representations. to train the face identity information to update the parameters of the network so that the face reconstruction network can well super-resolve very LR face images.