請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/65883
標題: | 用於雙魚眼影像人臉辨識之臉部校正 Deep Face Rectification for 360 degrees Dual-Fisheye Cameras |
作者: | Yi-Hsin Li 李奕欣 |
指導教授: | 陳宏銘 |
關鍵字: | 魚眼鏡頭相機,魚眼影像,人臉辨識,透視轉換,深度學習, Fisheye camera,fisheye distortion,face recognition,image rectification,deep learning, |
出版年 : | 2020 |
學位: | 碩士 |
摘要: | 當套用線性影像的人臉辨識模型於背對背雙魚眼鏡頭組成的360度相機拍攝到的魚眼影像時會產生嚴重的辨識準確率下降。我們提出一個創新的端到端深度人臉辨識架構來對抗魚眼扭曲對人臉辨識產生的影響。此架構由一個分類網絡、修復網絡、特徵取得網絡,以及特徵比對網絡組成。前兩個網絡是為了處理魚眼影像的非準質特性所特別設計的,並整合進線性影像的人臉辨識系統中的特徵取得以及特徵比對模組。分類網絡根據魚眼影像的扭曲程度進行分類。修復網絡將扭曲的魚眼影像作為輸入並將之修復為線性人臉結構。特徵取得網絡將修復後的影像投影到高維空間中的特徵向量,而特徵比對網絡利用此特徵向量進行人臉驗證以及人臉識別。提出方法的人臉驗證成效在模擬的野生標示資料集(Labeled Faces in the Wild, LFW)上達到99.18%的準確率並在真實的資料集上達到95.50%的準確率;平均較最佳的技術高出6.57%的準確率。而人臉識別方面則較最佳的技術高出4.51%的準確率。 Rectilinear face recognition models suffer from severe performance degradation when applied to fisheye images captured by 360 degrees back-to-back dual fisheye cameras. We propose a novel end-to-end deep face framework to combat the effect of fisheye image distortion on face recognition. The framework consists of a classification network, a restoration network, a feature extraction network, and a feature matching network. The first two networks are specifically designed to handle the non-linear property of fisheye images and to integrate with the feature extraction and feature matching modules of rectilinear face recognition systems. The classification network classifies an input fisheye image according to its distortion level. The restoration network takes a distorted image as input and restores the rectilinear geometric structure of the face. The feature extraction network maps the restored image to a feature vector, which is then used by the feature matching network for face verification and identification, in the latent space. The face verification performance of the proposed approach achieves 99.18% accuracy when tested on images in the synthetic Labeled Faces in the Wild (LFW) dataset and 95.50% for images in a real image dataset; the average improvement over the state-of-the-art technique is 6.57%. For face identification, the average improvement over the state-of-the-art technique is 4.51%. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/65883 |
DOI: | 10.6342/NTU202000453 |
全文授權: | 有償授權 |
顯示於系所單位: | 電信工程學研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-109-1.pdf 目前未授權公開取用 | 2.32 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。