摘要
将鉴别信息引入到距离测度中,利用这个新的局部距离测度代替欧氏距离构建k-近邻,提出一种新的局部线性近邻扩展算法。将此用于图像检索的相关反馈机制,产生基于局部自适应逼近的半监督反馈算法FLANNP(feedback locally adaptive nearest neighbor propagation)。该方法首先由支持向量机构建的判别函数来确定最优判别方向,基于此方向产生一个局部自适应距离算法,进而确定数据点间的权重。最后,标签信息由全局一致性假设,通过局部最近邻,从有标签数据点开始进行全局扩散标注。该方法使用有鉴别信息的距离测度,提高了图像检索的准确度。
In this paper, identification information was put into the distance measure, using this new distance measure instead of the Euclidean distance to construct k-neighbor, we proposed a new local linear nearest neighborhood propagation method. This provides a semi-supervised feedback algorithm based on the local adaptive approximation for image retrieval relevance feedback mechanism FLANNP (feedback locally adaptive nearest neighbor propagation). The decision function constructed by SVMs was used to determine the most discriminant direction in a neighborhood around the query. Such a direction provides a local adaptive distance algorithm. By this the reconstruction weights were computed. After all the labels were propagated from the labeled points to the whole dataset using the local linear neighborhoods with sufficient smoothness. The approach makes use of identification information in distance measure and improves the accuracy of image retrieval.
出处
《计算机科学》
CSCD
北大核心
2010年第7期280-284,共5页
Computer Science
基金
国家863高技术研究发展计划(No.2006AA01Z119)
国家自然科学基金(No.60473039)资助
关键词
相关反馈
半监督学习
局部自适应逼近
线性近邻扩展
Relevance feedback, Semi-supervised learning, Locally adapt approximating, Linear neighborhood propagation