CAJ | 학술논문

二代身份证人脸核实问题是指判断二代身份证人像和身份证使用者当前头像是否为同一人。具体来说，即将二代身份证模糊人像和实际在光照、背景等因素不可控环境下拍摄的若干张二代证使用者的视频人像作匹配，判断是否为同一个人。由于低分辨率模糊图像和清晰视频图像属于2种不同的图像模态，因此该问题属于异构人脸识别问题。考虑到跨模态人脸图像的差异，传统的特征抽取方法很难抽取判别性特征来描述不同模态图像，使得传统方法难以达到精准辨别。针对这个问题，提出了一种新的基于深度学习的解决方法，其基本思想是通过深度信念网络（ DBN）的非监督贪心逐层训练来提取人脸图像的高层特征，结合传统的图像预处理和相似性度量技术，达到人脸核实的目的。通过在256人的真实二代证数据集上和传统特征降维方法PCA、LDA进行比较，证实了所提出方法在准确率上相比PCA有约12％的提升，相比LDA有约8％的提升。实验同时表明，针对数据量增大的情况，基于深度学习的解决方法要优于传统的人脸识别方法。
이대신빈증인검핵실문제시지판단이대신빈증인상화신빈증사용자당전두상시부위동일인。구체래설，즉장이대신빈증모호인상화실제재광조、배경등인소불가공배경하박섭적약간장이대증사용자적시빈인상작필배，판단시부위동일개인。유우저분변솔모호도상화청석시빈도상속우2충불동적도상모태，인차해문제속우이구인검식별문제。고필도과모태인검도상적차이，전통적특정추취방법흔난추취판별성특정래묘술불동모태도상，사득전통방법난이체도정준변별。침대저개문제，제출료일충신적기우심도학습적해결방법，기기본사상시통과심도신념망락（ DBN）적비감독탐심축층훈련래제취인검도상적고층특정，결합전통적도상예처리화상사성도량기술，체도인검핵실적목적。통과재256인적진실이대증수거집상화전통특정강유방법PCA、LDA진행비교，증실료소제출방법재준학솔상상비PCA유약12％적제승，상비LDA유약8％적제승。실험동시표명，침대수거량증대적정황，기우심도학습적해결방법요우우전통적인검식별방법。
The objective of the face verification method for the second?generation identity card is to determine whether the original head?photo stored in the corresponding identity card image and the currently captured head photo of the card?holder by using a video camera image actually belongs to the same person or not. To obtain a good verification result for the heterogeneous face verification method is a very challenging task because the two different types of ima?ges belong to two different modalities (e.g., different image resolutions, different illumination conditions). Consider?ing the difference of trans?modal face images, it is hard to use traditional feature extraction methods to extract dis?criminative feature for description of images with different modes. Traditional feature extraction methods cannot distin?guish images exactly. In this paper, a deep learning?based face verification method is proposed. The proposed deep learning?based face verification method integrates the deep belief network ( DBN ) , which employs unsupervised greedy layer?by?layer training for high?level feature extraction of face photo and combines the popularly used image preprocessing and similarity measurement technologies to realize the purpose of face verification. The results were e?valuated on a real dataset with two different modalities of 256 different people. This method outperforms the traditional principal component analysis ( PCA) and linear discriminant analysis ( LDA) methods with 12% and 8% improve?ments in terms of the verification accuracy, respectively. The results validated the advantage of the proposed method, especially when the amount of entries increases.