一种基于语义子空谱聚类的自动图像标注的方法.docVIP

一种基于类的自动图像标注graph spectral clustering Algorithn Guo Yu-tang1 ， Han Chang-gang2 1．Department of Computer Science and Technology in Hefei Normal College, Hefei 230061,China 2．School of Computer Science and Technology in Anhui University, Hefei 230039,China aieyt@ Abstract: Due to existing the semantic gap, images with the same or similar low level features are possibly totally different on semantic level. How to find the underlying relationship between the high-level semantic and low level features is one of the difficult problems for image annotation. In this paper, a new image annotation method based on graph spectral clustering with the consistency of semantics is proposed with detailed analysis on the advantages and disadvantages of the existed image annotation methods. This method firstly cluster image into several semantic classes by semantic similarity measurement in the semantic subspace. Within each semantic class, images are re-clustered with visual features of.region Then, the joint probability distribution of blobs and words was modeled by using Multiple-Bernoulli Relevance Model. We can annotate a unannotated image by using the joint distribution. Experimental results show the the effectiveness of the proposed approach in terms of quality of the image annotation, the consistency of high-level semantics and low level features is efficiently achieved. Key Words: image annotation, K-Harmonic Means, graph spectral clustering, semantic gap ? 1 引言 自动图像标注指的是指借助计算机视觉、机器学习与模式识别等多学科技术，从一组已标注好的图像集合中学习图像特征与文本标注两种模态间的相关性，为未标注的图像推理出最为可能的标注或关键字。一旦图像被成功标注，图像检索问题就可以转化就可以用这样一组关键词来进行，当前技术已相当成熟的文本检索方法就可以用于图像检索[1]中。因此，图像语义标注得到了国内外相关领域的研究人员的广泛重视。目前，国内外已经有许多学者提出了各种不同的方法，并取得了相当多的成果。如共现模型[2]、翻译模型 [3-4]、交叉媒体相关模型 [5]和多伯努利相关模型[6]等。 聚类分析是机器学习领域中的一个主要分支，是人们认识和探索事物之间内在联系的有效手段。近几年来，基于聚类分析的自动图像标标注技术运应而生，并取得了可喜的效果。在参考文献[7-8]中，聚类用于一组训练图像的低层特征，利用统计学方法获得低层特征与标注词间的一组映射。用获得的映射规则标注图像。 这些方法都基于一种假设：视觉特征相同或相似的图像，其语义也应相同相似。由于语义鸿沟的存在，实际情况并非如此，那些有着相似的视觉特征的图像它们的语义并不相同，甚至相差很远。如“天空”与“大海”是不同的语义，但它