虹膜取像系統及切割方法比較

虹膜辨識是受人矚目的生物特徵辨識之一，其辨識步驟依序為取像、虹膜內外緣定位、特徵抽取、特徵比對，其中取像及虹膜區域切割是關鍵的兩個步驟。取得清晰的虹膜紋理並正確切割虹膜區域，為達成高精度虹膜辨識的要件。目前的虹膜取像裝置以及切割虹膜區域的演算法都只適用於人類，對不同物種的虹膜取像及切割仍有待發展。本論文開發出適用於鳥類的虹膜取像裝置，以及比較切割三種虹膜區域的演算法。我們在取像裝置的鏡頭前加裝三角形光圈，用來簡化估算聚焦好壞的計算。切割虹膜區域方面，由於鳥類眼睛構造與人相異，因此我們定義兩種錯誤指標分別為評估將虹膜誤判為背景的第一型誤差 (Type-1 Error) 以及將背景誤判為虹膜的第二型誤差 (Type-2 Error)，以量化切割方法的誤判率。基於兩種錯誤指標，我們評估兩種不同的虹膜切割方法，分別為區域偵測技術 (U-Net 與 SD-UNet) 以及邊界偵測技術 (HED) ，用於切割虹膜區域的錯誤率。經由實驗結果發現三種方法都可以達到不錯的切割效果，而比較三種不同的切割方法，HED 和 SD-UNet 分別實現最差和最佳的分割效果。

關鍵字

虹膜取像；虹膜切割；深度學習

並列摘要

Iris recognition is one of the most popular biometric identification techniques. The recognition procedure includes image acquisition, iris region segmentation, feature extraction, and feature matching. Among the four computation steps, image acquisition and iris region segmentation are two critical steps. Obtaining a clearly focused iris image and correctly segmenting the iris area are essential for achieving high-precision iris recognition. Existing iris image acquisition devices and algorithms are designed for human iris recognition and, thus, both the imaging system and the segmentation method for different species are yet to be developed. In this work, we develop an iris acquisition device suitable for birds and evaluate three approaches for segmenting the iris area. To simplify focus quality evaluation, we install a triangular aperture in front of the lens of the imaging device. As regarding to the iris area segmentation, we first define two segmentation error measures, namely, Type-1 error and Type-2 error. Type-1 error is defined to be the percentage of pixels within the iris region which are incorrectly classified as background. Whereas Type-2 error is the ratio of the number of background pixels that are miss-classified as iris region to the number of pixels of the ground-true iris area. Based on the two segmentaiton error measures, we evaluate two different approaches for iris segmentation. The first approach is region-based image segmentation (including U-Net and SD-UNet) and the second one is boundary-based segmentation (Holistically-Nested Edge Detection, HED) method. Experimental results show that all the three methods can achieve satisfactory segmentation results and, among the three segmentation methods, the HED method and the SD-UNet method achieve the worst and the best segmentation results, respectively.