CAJ | 학술논문

传统的视觉单词仅通过无监督聚类方法生成,标注的精度和效率较低。加权概念格是一种有效的层次数据分析工具,本文采用加权概念格对视觉单词进行分析与约简,提出了一种新的视觉单词生成方法。首先生成训练图像视觉词包的形式背景,并通过信息熵获取视觉单词的权值;其次针对各语义类别,根据用户所设定的内涵重要性阈值,构造出视觉词包模型频繁加权概念格;然后依据外延数阈值,提取对分类贡献大的描述图像语义的约简视觉单词,进一步提高了标注的精度和效率;最后通过实验验证了该方法是有效的和可行的。
전통적시각단사부통과무감독취류방법생성,표주적정도화효솔교저。가권개념격시일충유효적층차수거분석공구,본문채용가권개념격대시각단사진행분석여약간,제출료일충신적시각단사생성방법。수선생성훈련도상시각사포적형식배경,병통과신식적획취시각단사적권치;기차침대각어의유별,근거용호소설정적내함중요성역치,구조출시각사포모형빈번가권개념격;연후의거외연수역치,제취대분류공헌대적묘술도상어의적약간시각단사,진일보제고료표주적정도화효솔;최후통과실험험증료해방법시유효적화가행적。
Traditional visual word is generated through unsupervised clustering method, and the image semantic classification performance is not high. Weighted concept lattice is an effective tool for data analysis. Through analy- zing and reducing visual words by making use of weighted concept lattice, the paper presents a novel method of building visual words. First, the formal context about the BOV of training image is generated, and the weight value of visual word is acquired through information entropy. Second, the frequent weighted concept lattice of Bag-of- vis- ual words model is constructed for each semantic category according to the intent threshold given by the user. Then, the reduced visual words which contribute to scene category were selected to construct visual words dictiona- ry according to the extent threshold, and further the precision and efficiency of label was improved. Final experi- ments show that it is effective and feasible.