面向检索的多视觉特征融合
发布时间：2018-11-05 09:32
【摘要】：近几十年来,伴随着互联网的快速发展以及智能终端的普及,互联网上数字图像的数量呈现爆炸式增长。面对海量的图像信息,如何快速高效地检索这些图像一直是学术界和工业界研究的热点课题。图像的特征表达是基于内容的图像检索的最基本问题之一。为了提升检索的准确度,研究人员从颜色、纹理等不同的角度提出不同的视觉特征表达来表征图像。选择不同的视觉特征对于图像检索的准确度有很大的影响。一般来说,采用具有一定互补性的多种特征进行融合是提升图像检索准确度的一种方法。为了把基于不同特征得到的图像检索结果融合在一起,我们有两个关键的问题需要解决。第一个关键问题是如何使基于不同特征空间的距离度量是可比拟的。因为通常使用不同的特征,如SIFT,HSV,CNN特征,算得的距离是不在一个尺度空间的。直接把不在一个尺度空间的"距离"进行相加是不合适的。第二个需要关注的关键的问题是,如何自适应的度量不同的特征的有效性。因为对于某些查询图像来说,局部特征就能取得较好的检索结果。然而对于另外某些查询图像,用全局特征比如CNN特征才能够得到比较好的检索结果。对于同一个查询图像,我们需要比较并量化不同特征的有效性。基于上面的两个关键点,我们的工作主要归纳如下:(1)基于图模型的自适应加权特征融合方法。在此方法中,图模型把本来在不同尺度空间的距离度量,都统一到一个Graph里面,并用统一的度量方法Jaccard系数来度量各个图片之间的相似度。同时为了衡量不同特征的有效性,我们使用PageRank算法对不同特征构建的图进行分析,并根据最后得到的PageRank值的分布来衡量不同特征的有效性。最后根据特征对特定检索图像的有效性,完成不同特征构建的图的自适应加权融合。根据最后融合得到的图,我们解出最后的图片检索排序。(2)基于邻域相似度分布的自适应多特征融合方法。该方法是根据图像在给定的视觉特征下的近邻空间的分布情况,来进行特征融合。不同特征对于一个具体的查询图像得到的k近邻的距离空间分布是不一样的。我们通过探索k近邻的空间分布特性,来进行衡量不同特征的有效性。我们提出了有效性系数的概念-REC(Rank Effectiveness Coefficient)。REC 反映 了一个特征对一个具体图像的有效性。通过有效性系数对原来特征的相似度进行加权融合,最后得到融合后的相似度得分。根据融合后的相似度得分,可以给出最后的图像检索排序结果。
[Abstract]:In recent decades, with the rapid development of the Internet and the popularity of intelligent terminals, the number of digital images on the Internet has increased explosively. In the face of massive image information, how to retrieve these images quickly and efficiently has been a hot topic in academia and industry. Image feature representation is one of the most basic problems in content-based image retrieval. In order to improve the retrieval accuracy, the researchers proposed different visual features to represent the image from different angles such as color and texture. The selection of different visual features has great influence on the accuracy of image retrieval. Generally speaking, it is a method to improve the accuracy of image retrieval by using a variety of complementary features. In order to fuse the image retrieval results based on different features, we have two key problems to be solved. The first key problem is how to make distance metrics based on different feature spaces comparable. Because different features, such as SIFT,HSV,CNN features, are usually used, the distance calculated is not in the same scale space. It is not appropriate to add "distances" that are not in a scale space directly. The second key concern is how to measure the effectiveness of different features adaptively. Because for some query images, local features can obtain better retrieval results. However, for some other queried images, global features such as CNN features are used to obtain better retrieval results. For the same query image, we need to compare and quantify the validity of different features. Based on the above two key points, our work is summarized as follows: (1) Adaptive weighted feature fusion method based on graph model. In this method, the graph model unifies the distance measure in different scale space into one Graph, and uses the unified measure method Jaccard coefficient to measure the similarity of each picture. At the same time, in order to evaluate the validity of different features, we use PageRank algorithm to analyze the graph constructed by different features, and evaluate the validity of different features according to the distribution of PageRank values. Finally, according to the validity of the feature to the specific retrieval image, the adaptive weighted fusion of the graph constructed by different features is completed. According to the final fusion graph, we solve the final image retrieval ranking. (2) Adaptive multi-feature fusion method based on neighborhood similarity distribution. This method is based on the distribution of the image in the nearest neighbor space under the given visual features. The distance distribution of k-nearest neighbor is different for a specific query image with different features. We evaluate the effectiveness of different features by exploring the spatial distribution of k-nearest neighbors. We propose the concept of validity coefficient-REC (Rank Effectiveness Coefficient). REC reflects the validity of a feature to a specific image. The similarity of the original feature is weighted by the validity coefficient, and the score of the similarity after the fusion is obtained. According to the similarity score of fusion, the final result of image retrieval and sorting can be given.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
本文编号：2311654
[Abstract]:In recent decades, with the rapid development of the Internet and the popularity of intelligent terminals, the number of digital images on the Internet has increased explosively. In the face of massive image information, how to retrieve these images quickly and efficiently has been a hot topic in academia and industry. Image feature representation is one of the most basic problems in content-based image retrieval. In order to improve the retrieval accuracy, the researchers proposed different visual features to represent the image from different angles such as color and texture. The selection of different visual features has great influence on the accuracy of image retrieval. Generally speaking, it is a method to improve the accuracy of image retrieval by using a variety of complementary features. In order to fuse the image retrieval results based on different features, we have two key problems to be solved. The first key problem is how to make distance metrics based on different feature spaces comparable. Because different features, such as SIFT,HSV,CNN features, are usually used, the distance calculated is not in the same scale space. It is not appropriate to add "distances" that are not in a scale space directly. The second key concern is how to measure the effectiveness of different features adaptively. Because for some query images, local features can obtain better retrieval results. However, for some other queried images, global features such as CNN features are used to obtain better retrieval results. For the same query image, we need to compare and quantify the validity of different features. Based on the above two key points, our work is summarized as follows: (1) Adaptive weighted feature fusion method based on graph model. In this method, the graph model unifies the distance measure in different scale space into one Graph, and uses the unified measure method Jaccard coefficient to measure the similarity of each picture. At the same time, in order to evaluate the validity of different features, we use PageRank algorithm to analyze the graph constructed by different features, and evaluate the validity of different features according to the distribution of PageRank values. Finally, according to the validity of the feature to the specific retrieval image, the adaptive weighted fusion of the graph constructed by different features is completed. According to the final fusion graph, we solve the final image retrieval ranking. (2) Adaptive multi-feature fusion method based on neighborhood similarity distribution. This method is based on the distribution of the image in the nearest neighbor space under the given visual features. The distance distribution of k-nearest neighbor is different for a specific query image with different features. We evaluate the effectiveness of different features by exploring the spatial distribution of k-nearest neighbors. We propose the concept of validity coefficient-REC (Rank Effectiveness Coefficient). REC reflects the validity of a feature to a specific image. The similarity of the original feature is weighted by the validity coefficient, and the score of the similarity after the fusion is obtained. According to the similarity score of fusion, the final result of image retrieval and sorting can be given.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
【参考文献】
相关期刊论文 前1条
1 周文罡;李厚强;卢亦娟;田奇;;Encoding Spatial Context for Large-Scale Partial-Duplicate Web Image Retrieval[J];Journal of Computer Science & Technology;2014年05期
，本文编号：2311654
本文链接：https://www.wllwen.com/shoufeilunwen/xixikjs/2311654.html
