基于弱语义嵌入的图像相似性度量

提出了一种旨在缩小"语义鸿沟"的弱语义嵌入方法及其相似性度量模型.在该方法中,用户反馈的相关图像被当成查询图像的弱语义直接嵌入到视觉特征向量中,从而使固定的视觉特征向量成为可伸缩的包含语义和视觉特征的集合.根据心理学中集合论相似性理论,图像之间的相似性被表示成语义特征与视觉相似性之比.基于Corel图像库的实验结果表明,该方法在基于内容的图像检索中是非常有效的.

Image similarity measures based on weak semantic embedding

To bridge the semantic gap, a novel approach that weak semantic features are embedded into visual features is proposed. Moreover, a novel similarity model is employed to measure the similarity between two images with both semantic and visual features. In interactive image retrieval, similar images are defined as weak semantic features and embedded into the feature vector of the querying image. Therefore, the fixed visual feature vector becomes an expandable feature set, which integrates semantic and visual features. Based on the set-theoretic similarity, the similarity between two im ages is expressed as the ratio of the measures of semantic features to those of visual features. Experimental results, over Corel image collections, show that the approach is effective in content-based image retrieval.