基於pix2pix深度學習模型之條件式虹膜影像生成架構;A Deep Learning Framework for Conditional Iris Image Generation Based on Pix2Pix Model

NCU Institutional Repository > 資訊電機學院 > 資訊工程研究所 > 博碩士論文 > Item 987654321/84113

jsp.display-item.identifier=請使用永久網址來引用或連結此文件: http://ir.lib.ncu.edu.tw/handle/987654321/84113

题名:	基於pix2pix深度學習模型之條件式虹膜影像生成架構;A Deep Learning Framework for Conditional Iris Image Generation Based on Pix2Pix Model
作者:	彭宇喧;Peng, Yu-Syuan
贡献者:	資訊工程學系
关键词:	深度學習;生成對抗式網路;虹膜辨識;Deep Learning;Generative Adversarial Network;Iris Recognition
日期:	2020-08-17
上传时间:	2020-09-02 18:06:16 (UTC+8)
出版者:	國立中央大學
摘要:	在市場上主流的生物辨識技術有指紋辨識、人臉辨識、虹膜辨識，而在安全性需求高的門禁系統，虹膜辨識往往佔有舉足輕重的角色。近年來興起的深度學習技術也漸漸應用到虹膜辨識技術上。眾所皆知，應用深度學習技術需要大量且有良好人工標籤的資料集，越大量的資料越能提升演算法效能。現今社會大眾重視隱私權及法律的限制下，收集個人虹膜影像資料變得非常困難，更遑論能收集到有人工標籤且品質良好的虹膜影像資料，於是我們提出了一個基於pix2pix的深度學習網路架構，並人工標記了兩個虹膜資料集，CASIA-Iris-Thousoud與ICE，使每張虹膜影像資料有對應的虹膜遮罩與眼周遮罩，只要提供相對應的虹膜遮罩與眼周遮罩，透過生成對抗式模型生成擬真的虹膜影像資料，以增加虹膜影像資料庫。並提出方法產生合理的虹膜遮罩與眼周遮罩，為後續研究基於深度學習的虹膜分割演算法，提供大量的人造資料集，以提升演算法的準確度。 ;The mainstream biometrics technology in the market include fingerprint recognition, face recognition, and iris recognition. Iris recognition often plays a pivotal role on access control system with high security requirements. The deep learning techonology that rise in recent years has gradually been applied to iris recognition technology. As we all know, applying deep learning techniques requires a large amount of data sets with high quality manual labels. The larger the amount of data, the better the algorithm performs. Nowadays, the general public pays more attention to privacy and legal restrictions, so it is very difficult to collect personal iris image data. Not to mention the high quality iris image data with decent manual labels. In this work, we proposed a deep learning network architecture based on pix2pix. We have manually produce iris masks for two iris datasets: CASIA-Iris-Thousoud dataset and ICE dataset. Based on the original iris image in Cartesian domain, we create contour information and binary mask for each iris image. The ultimate goal for this study is a conditional iris image and mask generator, which takes inputs of iris and eye mask are provided, and outputs a photo-realistic iris image by applying the proposal conditional Pix2Pix generative models. We also proposed a method to produce a reasonable iris mask and periocular mask for follow-up research based on deep learning iris segmentation algorithm, which enables researchers to generate unlimited number of artificial iris data sets. Such large-scale iris dataset is very hard to be collected in practical situations. The work described in this thesis will enable researchers to have enough training data to train iris segmenter and mask producer based on deep learning technique.
显示于类别:	[資訊工程研究所] 博碩士論文

文件中的档案:

档案	描述	大小	格式	浏览次数
index.html		0Kb	HTML	175	检视/开启

在NCUIR中所有的数据项都受到原著作权保护.

社群 sharing

数据加载中.....