【作者】 谢峰；

【导师】 何晓飞；

【作者基本信息】 浙江大学 ， 计算机科学与技术， 2012， 硕士

【摘要】 目前大多数流行的图像搜索引擎,都利用从图像的上下文中抽取出来的文本信息来建立索引。这种机制下的检索结果往往含有大量的垃圾信息,因为文本上下文与检索关键词匹配,不一定代表图像本身与检索关键词匹配；最关键的是这种机制忽略了图像视觉内容信息本身。针对Google Images图像搜索引擎的检索返回结果,我们提出了一种新的、基于聚类方法的垃圾图片过滤方法。首先,我们对Google Images的检索结果进行图像特征抽取。为了准确得描述图像视觉内容信息,我们采用了三种不同性质的图像特征。同时设计了线性混合内核函数将这些图像特征混合在一起,构建出相似性矩阵。然后,利用K-way Min-MaxCut算法对返回图像进行聚类,与聚类过程相迭代的是线性混合内核的权重系数调整过程,从而各内核系数的设定得以自动完成。聚类完成后,对各个类进行排序。排序的策略是先选出质量最优秀的类,剩余的类再按照其与质量最优类之间的相似性大小进行排序。最后,根据所需的返回图像数目,过滤掉排在末尾的图像类。我们的垃圾图片过滤算法拥有很具竞争力的实验表现。在低召回率的实验中,我们的算法在多数情况下比当前最优秀的过滤算法表现优秀,说明了聚类算法和最优类选择策略的鲁棒性。在多召回率的试验中,我们的算法比Google Images有很好的精度提高表现,说明了类排序算法的有效性。整体来说,我们的算法能切实得对Google Images的检索返回结果过滤垃圾图像。更多还原

【Abstract】 Currently most existing image search engines such as Google Images index web im-ages majorly using text keywords extracted from the context, which may return large amount of junk information. The matching between query and context doesn’t necessarily mean the matching between query and image itself, ignoring the visual content of image is the es-sential reason. Based on the search result of Google Images, we propose a novel clustering based filtering method to filter those junk images.Firstly we extract three different image features from images returned by Google, meanwhile we design appropriate kernel functions to capture the visual similarity property. Linear mixture method is used to mix all three feature kernels. Based on the kernel matrix, we apply K-way min-max cut to cluster those images into multiple clusters, and kernel weights in the mixture-of-kernel can be determined automatically. Secondly we select the best cluster in a robust way, and rank all the rest clusters according to their similarity with the best one. Finally those low-rank clusters can be filtered out as junk clusters.We obtain very comparative filtering performance throughout the experiments. In the part of low-recall level comparision, our method outperforms the current state-of-the-art in most cases, which states the robustness of clustering algorithm and best cluster selection algorithm. In the part of multi-recall level comparison, our method improves Google Images consistantly and significantly, which states the effectiveness of ranking algorithm.更多还原