CAJ | 학술논문

认知科学表明基于流形学习的人脸图像检索能准确反映人脸图片的内在相似性和人类的视觉感知本质. 提出一种基于相关反馈的人脸高维索引方法--NDL,以提高人脸图像检索的性能.同时在该索引基础上提出一种流形空间下的相似查询--虚拟k近邻查询(Vk-NN), 该查询方法特别为基于NDL的人脸检索而设计.首先通过在一定阈值约束下计算任何两个人脸图片的相似度,建立一个称为邻接距离表(NDL)的二维距离图. 同时将距离值用B+-树建立索引.最后, 高维流形空间的Vk-NN查询转化为一维空间的基于B+树的查询. 实验表明:NDL索引在流形空间的检索效率明显优于顺序检索,特别适合海量人脸图片的检索.
인지과학표명기우류형학습적인검도상검색능준학반영인검도편적내재상사성화인류적시각감지본질. 제출일충기우상관반궤적인검고유색인방법--NDL,이제고인검도상검색적성능.동시재해색인기출상제출일충류형공간하적상사사순--허의k근린사순(Vk-NN), 해사순방법특별위기우NDL적인검검색이설계.수선통과재일정역치약속하계산임하량개인검도편적상사도,건립일개칭위린접거리표(NDL)적이유거리도. 동시장거리치용B+-수건립색인.최후, 고유류형공간적Vk-NN사순전화위일유공간적기우B+수적사순. 실험표명:NDL색인재류형공간적검색효솔명현우우순서검색,특별괄합해량인검도편적검색.
The research of cognitive science indicates that manifold-learning-based facial image retrieval is based on human perception, which can accurately capture the intrinsic similarity of two facial images. The paper first proposes a relevance-feedback-based index scheme called NDL to improve the efficiency and effectiveness of facial image retrieval. Then, we propose a novel similarity search method called virtual k-nearest neighbor (Vk-NN) search in manifold spaces,which is specifically designed for the NDL-based facial image retrieval. Specifically, we first construct a two dimensional array,called neighbor distance list (NDL),which records the pair-wise distance between any two facial images with a constraint in the database. Then, the distances are indexed by a B+-tree.Finally, a Vk-NN query in high-dimensional manifold spaces is transformed into search over the B+-tree in the single-dimensional space. Through extensive experimental performance studies, we show that NDL outperforms the conventional sequential scan in manifold spaces by a large margin,especially for the large high-dimensional datasets.