【作者】 黄鑫； 张世佳； 汪国平； 王衡；

【Author】 Huang Xin+, Zhang Shi-jia, Wang Guo-ping, Wang Heng 1(Human-Computer Interaction & Multimedia Lab, Department of Computer Science and Technology, Peking University, 100681, China)

【机构】 北京大学信息科学技术学院人机交互与多媒体实验室；

【摘要】 颜色特征是基于内容的图像检索方法中使用最广泛的底层特征之一。本文提出了一种基于主色匹配的图像检索方法,通过使用聚类算法提取出图像的主色特征,作为整幅图像内容的表示和描述。同时本文根据主色匹配的思想,定义了基于主色特征的相似度计算方法。实验结果表明本文的方法具有优良的检索性能,提取出的特征文件空间复杂度低,有利于图像库的索引和管理。更多还原

【Abstract】 Color is one of the most important low-level features used in image retrieval because of its characteristic of invariance with respect to image scaling and orientation. By exploiting the fact that the human eye cannot perceive a large number of colors at the same time, nor is it to distinguish close colors well, we aim to create small color descriptor, which is suitable for image representation, matching and retrieval. In this paper we introduce a compact color feature extraction scheme in the CIELAB color space, which helps us for developing perceptually based image similarity metric. According to the idea of similarity measure by the dominant colors matching, we make use of the improved LBG algorithm, which is commonly used in vector quantization area, to extract the dominant colors vector which is most matching human perception of image. The paper also develops the corresponding method of similarity measure. The basic principle is that similarity between two images in terms of color composition should be measured by a combination of color and area differences. The dominant color and the corresponding area percentage allow the definition of a color similarity measure closely match the human perception. We have developed a content-based image retrieval system to validate the effectiveness of proposed algorithms and techniques in our paper. From the retrieving results we can see that our scheme is well defined, and achieves much better retrieving results than other three previously proposed methods. We use average retrieval rate (ARR) and average normalized modified retrieval rank (ANMRR) to evaluate the performance of our proposed technique in the image database. We also compare feature file sizes extracting by our method and other three previously proposed methods, which indicates that our method largely reduces the feature file sizes needed for storage and indexing. The contribution of this paper is that we propose a clustering technique to extract perceptually dominant colors, and design a compact dominant color codebook and corresponding similarity metric. The proposed scheme in our paper has other important advantages for image retrieval system too, such as great reduction of the feature file sizes needed for storage and indexing. Because the perceptually dominant colors are well clustered, the proposed techniques can be easily extended to include the shape features or spatial information to measure similarity, which is our interesting research direction in the near future.更多还原