應用於虹膜辨識之成對特徵學習與遮罩注意力機制神經網路

Abstract

虹膜圖案複雜、隨機的特性使得虹膜辨識成為一種具有高度準確性以及可靠性的身分辨識方法。跟其他生物特徵相比，虹膜的穩定特性不像指紋一樣會隨著時間磨損毀壞，而它平面幾何組成又比臉部特徵容易計算預測，因此虹膜辨識已經被廣泛應用在多種領域之中，像是資料庫存取、金融認證服務和個人穿戴式裝置上更是有很大的潛力。 自從 1993 第一個開創性的虹膜辨識膜方法問世以後，許多與之相關的研究也隨之展開。經過幾十年來的發展，對在控制條件下拍攝眼睛照片的虹膜辨識已經被廣泛地解決，因此大部分研究開始專注在增強系統的穩定性或是克服非理想情況下拍攝的照片。近幾年很多機器學習和類神經網路的方法被證明對影像處理等多媒體領域有很大的幫助，許多虹膜辨識的研究也開始使用深度學習的架構來進行研究。 有鑑於虹膜辨識在未來極大的潛力以及深度學習現今廣泛應用的背景之下，我們架設了一個基於成對特徵學習以及遮罩注意力機制的神經網路來進行虹膜辨識的任務。除此之外，為了檢驗這個系統架構在真實世界的應用價值，我們更是利用商業用穿戴式眼球追蹤眼鏡所拍攝的照片來進行驗證。總結來說，我們提出了一個可靠又準確且有實際應用價值的虹膜辨識模型。

Highly accurate, invariable, and robust biometric authentication on the basis of iris patterns has been widely applied to many aspects of life in recent years. Despite numerous commercialized iris recognition products on the market, more advanced techniques are developed as a result of multiple types of research in deep learning methods. However, experiments in most of the works are only conducted on established open datasets. Hence, it is worthwhile to verify the practical value of an iris recognition system when it comes to real images taken by apparatus in a realistic world. In this thesis, we present a novel neural network architecture based on a pairwise feature learning method and mask attention mechanism. Specifically, the proposed algorithm learns the similarity between different iris image pairs by their spatially according features. Moreover, mask attention mechanisms are added to achieve more reliable results and to reinforce the effectiveness of the pairwise feature extraction network. In the experiments, our framework achieves state-of-the-art performance in three well-known open datasets. Additionally, we used our own camera on a wearable device to capture a variety of iris image data under certain conditions and thus evaluate the utility of the system.