《基于协作式标注图像的垃圾标签检测和垃圾图像过滤方法研究》-毕业论文.doc

浙江大学硕士学位论文 STYLEREF 标题,章标题(无序号) \* MERGEFORMAT Abstract PAGE i 基于协作式标注图像的垃圾标签检测 和垃圾图像过滤方法研究 摘要 随着web的普及，越来越多的人上传图片和标注图片，这种由用户上传并添加标注的图片称为协作式标注图像，这些标注提供了大量语义信息，但由于用户标注的随意性造成一些标注与图像内容的关联性并不强，直接影响到协作式标注图像检索的准确性，因此，本文设计了一种垃圾标签检测和垃圾图像过滤方法，提高标注和图像之间的关联性，从而提高协作式标注图像检索的准确性。 本文垃圾标签检测的方法是通过计算同一图像的所有标注词之间的语义相似性进行聚类，得到的聚类标签较少的那类标签即为垃圾标签。其中标注词之间的语义相似性采用本文设计的多重信息源的标注语义相似性，它整合了标签的WordNet语义词典相似性以及标签的同现频率的统计特性，能够更加准确的描述标签之间的语义相关性。垃圾图像过滤方面，通过对同一标注下所有图像的视觉特征进行基于混合核的最大最小聚类从而过滤垃圾图像。混合核的方法更加准确的描述了图像的视觉相关性。垃圾标签检测和垃圾图像过滤互为补充，共同提高图像和标签之间的关联性。最后，本文设计了双曲可视化的图像显示界面。实验表明，本文方法能够较为准确的检测协作式标注图像的垃圾标签以及过滤垃圾图像，从而提高检索准确性。 关键词：　协作式标注，垃圾标签，垃圾图像，WordNet语义词典，混合核，SIFT 浙江大学硕士学位论文 Abstract 浙江大学硕士学位论文 Abstract PAGE ii PAGE ii Abstract With the popularity of the web, more and more people begin to upload images and label images, this kind of images which uploaded and labeled by users is called collaborative tagging images, these annotations can provide plenty of semantic information, but the users subjective caused some correlation between annotations and image content is not strong, it will directly affect the accuracy of image retrieval for collaborative tagging images, and therefore, we design a junk tag detecting and junk image filtering method for the collaborative tagging images and annotations, it improve the correlation between images and annotation, so as to improve the accuracy of collaborative tagging image retrieval. In this paper, junk tags detecting method is got by clustering by semantic similarity for all of the annotations of one image, the kind of tags which the number of tag in class is less is junk tags. The semantic similarity between tags is based on multiple information sources, which integrates semantic similarity by the wordnet semantic dictionary and co-occurrence frequency belongs to statistical pr