CAJ | 학술논문

目前在图像检索领域，由于视觉字典其性能突出，已成为图像检索领域构建视觉词典的主流方法。但传统的视觉字典方法存在运行时间效率低、内存消耗大等缺点。因此本文采用ROOTSift算法提取图像的特征点并利用高效的K-means聚类算法建立支持动态扩充的随机视觉字典。该方法基于视觉字典构建视觉词汇直方图和倒排序索引文件，并对视觉词重新分配权重以提高检索命中率。最后利用欧氏距离法查询完成相似性匹配。试验结果表明该方法能提高图像检索的准确率，对大规模的图像检索能够达到很好的检索质量。
목전재도상검색영역，유우시각자전기성능돌출，이성위도상검색영역구건시각사전적주류방법。단전통적시각자전방법존재운행시간효솔저、내존소모대등결점。인차본문채용ROOTSift산법제취도상적특정점병이용고효적K-means취류산법건립지지동태확충적수궤시각자전。해방법기우시각자전구건시각사회직방도화도배서색인문건，병대시각사중신분배권중이제고검색명중솔。최후이용구씨거리법사순완성상사성필배。시험결과표명해방법능제고도상검색적준학솔，대대규모적도상검색능구체도흔호적검색질량。
In the current research field of image searching, due to the good performance, visual vocabulary has currently become the mainstream method for creating visual dictionaries. The traditional method based on visual vocabulary has the disadvantages of low efficiency, high memory consumption and so on. Thus, the system in this paper utilizes the ROOTSift algorithm to extract feature points in images and adopts the efficient k-means cluster algorithm to create random visual vocabulary which can be expanded dynamically. While there are some open problems in current image searching system, such as the problems of ambiguity and synonymity in visual vocabulary, object positioning and detection performance with big data, this system creates the visual vocabulary histogram and index files with reverse order based on visual vocabulary and reallocates the weights to visual vocabulary to improve the hit ratio of detection. Finally, the similarity matching is achieved through Euclidean distance searching.The experiments show that this method can increase the success rate of image searching and improve the efficiency of the searching ,and it can achieve a nice retrieval quality of the image searching.