基於深度學習方法之高精確度瞳孔放大片偵測演算法;Ultra-Accurate Detection of the Existence of Cosmetic Contact Lens for Iris Images based on Deep Learning

NCUIR > College of Electrical Engineering & Computer Science > Executive Master of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/83940

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/83940

Title:	基於深度學習方法之高精確度瞳孔放大片偵測演算法;Ultra-Accurate Detection of the Existence of Cosmetic Contact Lens for Iris Images based on Deep Learning
Authors:	劉榮勝;Liu, Rong-Sheng
Contributors:	資訊工程學系在職專班
Keywords:	瞳孔放大片;深度學習;虹膜分割;虹膜識別;Cosmetic contact lens;Deep learning;Iris segmentation;Iris recognition
Date:	2020-08-11
Issue Date:	2020-09-02 17:43:40 (UTC+8)
Publisher:	國立中央大學
Abstract:	近年來，瞳孔放大片已經成為許多民眾的生活用品，更是不少愛美時尚男女的生活必需品。為了符合更多需求，廠商也針對色澤、風格和紋理提供更多的選擇，豐富產品的多變性。這些瞳孔放大片也因改變虹膜紋理在虹膜辨識上受到考驗。若要透過深度學習的方法，則需要收集大量的資料進行網路模型的訓練，從資料中萃取出複雜的規則；另外在深度學習進行訓練之前，需要對影像做預處理，如：影像的分割、影像形式的轉換及利用影像處理的方法增加影像的數量等，來達到高準確以及穩健的結果。本篇論文收集台灣市售的9家廠牌、18種款式及101位配戴虹膜放大片與未配戴時的樣本，實驗中使用的圖像總數為30390，透過深度學習的方式訓練模型，使得測試準確度可達到99%以上的水準。 ;In recent years, Cosmetic Contact Lens (CCL) has become a daily necessity for many people, and it is also a necessity for many people who love beauty and fashion. In order to meet more needs, manufacturers also provide more choices for color, style and texture to enrich the variability of products. These Cosmetic Contact Lens (CCL) also becomes a challenge for iris recognition because it changes the appearance of the texture of the iris. However, in deep learning method, one needs to collect a lot of data for the training of the network model, and extract rules from the data. In addition, before training a deep learning, it is better to preprocess the image for the sake of data augmentation, such as : image cropping, scaling, rotating to achieve higher accuracy and robustness. This paper collects CCL samples from 9 brands and 18 styles from Taiwan. We invite 101 participants and collect eye images with and without wearing CCL. The total number of images used in the experiment is 30390. At the end, we can achieve an accuracy higher than 99% using deep learning based models.
Appears in Collections:	[Executive Master of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	188	View/Open

社群 sharing

Loading...