摘要
视觉特征的描述和相似度的测量是图像检索的两大关键。本文提出在较为均匀的L*a*b*颜色空间,利用平均值位移聚类算法提取主色,相似度的测量采用EMD方法,利用粗略检索减少查询时间。然后利用相关反馈与多次检索有效地提高检索准确率。最后通过在原型系统上的两个实验,从查准率、平均序号和检索时间验证了此方法的有效性。
The description of the visual character and the similarity measure are two keys in an Image Retrieval System. A clustering algorithm based on mean shift is used to extract the dominant color and the color feature in CIE L^＊a^＊b^＊ color model. Earth Mover＇s Distance is used to calculate dissimilarity between images. In order to reduce the computation time, cursory search is used. The technique of relevance feedback is used to enhance the effectiveness. A simple prototype system is developed to compare the retrieval accuracies, the average rank and the execution time by two experiments. The results show that the proposed approach is effective.
出处
《电路与系统学报》
CSCD
北大核心
2007年第1期62-67,共6页
Journal of Circuits and Systems
基金
国家自然科学基金资助项目(60472099)
国家教育部留学回国人员基金
计算机软件新技术国家重点实验室资助
关键词
图像检索
主色
平均值位移
EMD测量
相关反馈
image retrieval
dominant color
mean shift
EMD measure
relevance feedback
