图像搜索重排序关键技术研究
发布时间：2018-10-24 17:27
【摘要】：基于内容的图像分析技术在图像检索中的应用已经引起越来越广泛的关注，图像搜索重排序技术是其中一种利用图像的视觉信息对初始文本搜索结果进行再次分析与排序的新技术。有效的视觉表征是其中的关键技术之一，然而由于视觉特征具有高维及存在“语义鸿沟”等问题，直接应用现有视觉特征难以获得较好的排序性能。维数约简方法可以在一定程度上克服这些缺点，但是传统的维数约简维数约简算法往往是针对分类任务提出的，，并不适合于排序问题。排序学习与分类任务并不等同，因此设计适用于图像搜索重排序学习的维数约简算法显得尤为重要。为此，本文有针对性地进行了若干研究，主要工作及创新为： （1）基于PCA（Principal Component Analysis）降维后每个维度具有的不同比重的贡献率，提出了一种基于主成分分析的相似度计算方法SM-PCA，并在此基础上提出了一种利用少量标注样本即可得到较好的排序性能的直推式半监督重排序方法。在该方法中采用迭代的方式计算扩展训练样本集合，并利用训练样本集合训练排序模型，最后对待排序的样本进行重排序，在网络搜索引擎下载的图像数据库验证了算法性能的有效性。 （2）提出了一种基于典型相关性分析的排序维数约简算法。在排序学习中广泛存在的是样本的相关性等级信息，其与样本的类别标签信息有很大的不同。基于此，在典型相关性分析算法的基础上，把排序问题中样本的相关性等级信息引入到维数约简技术中，设计适用于多模态数据的维数约简算法。将其应用到图像搜索重排序中，大量实验表明所提算法可以显著地改善图像检索性能。
[Abstract]:The application of content-based image analysis technology in image retrieval has attracted more and more attention. Image search reordering is one of the new techniques to reanalyze and sort the original text search results using the visual information of the image. Effective visual representation is one of the key technologies. However, due to the high dimension of visual features and the existence of "semantic gap", it is difficult to obtain better ranking performance by direct application of existing visual features. Dimension reduction can overcome these shortcomings to some extent, but the traditional dimension reduction algorithm is often proposed for classification tasks and is not suitable for sorting problems. Sorting learning is not the same as classification task, so it is very important to design dimension reduction algorithm which is suitable for image search and resort learning. Therefore, this paper has carried out a number of targeted studies. The main work and innovations are as follows: (1) based on the contribution rate of different proportion of each dimension after dimension reduction by PCA (Principal Component Analysis), In this paper, a similarity calculation method based on principal component analysis (SM-PCA,) is proposed, and a direct-push semi-supervised reordering method based on a small number of labeled samples is proposed. In this method, an iterative method is used to calculate the set of extended training samples, and the training set is used to train the sorting model. Finally, the sorted samples are reordered. The performance of the algorithm is validated in the image database downloaded by the network search engine. (2) A sort dimension reduction algorithm based on canonical correlation analysis is proposed. The correlation level information of samples exists widely in sorting learning, which is very different from the category label information of samples. Based on this, based on the typical correlation analysis algorithm, the correlation level information of the sample in the sorting problem is introduced into the dimension reduction technology, and a dimension reduction algorithm suitable for multi-modal data is designed. A large number of experiments show that the proposed algorithm can significantly improve the performance of image retrieval.
【学位授予单位】：天津大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP391.41
本文编号：2292057
[Abstract]:The application of content-based image analysis technology in image retrieval has attracted more and more attention. Image search reordering is one of the new techniques to reanalyze and sort the original text search results using the visual information of the image. Effective visual representation is one of the key technologies. However, due to the high dimension of visual features and the existence of "semantic gap", it is difficult to obtain better ranking performance by direct application of existing visual features. Dimension reduction can overcome these shortcomings to some extent, but the traditional dimension reduction algorithm is often proposed for classification tasks and is not suitable for sorting problems. Sorting learning is not the same as classification task, so it is very important to design dimension reduction algorithm which is suitable for image search and resort learning. Therefore, this paper has carried out a number of targeted studies. The main work and innovations are as follows: (1) based on the contribution rate of different proportion of each dimension after dimension reduction by PCA (Principal Component Analysis), In this paper, a similarity calculation method based on principal component analysis (SM-PCA,) is proposed, and a direct-push semi-supervised reordering method based on a small number of labeled samples is proposed. In this method, an iterative method is used to calculate the set of extended training samples, and the training set is used to train the sorting model. Finally, the sorted samples are reordered. The performance of the algorithm is validated in the image database downloaded by the network search engine. (2) A sort dimension reduction algorithm based on canonical correlation analysis is proposed. The correlation level information of samples exists widely in sorting learning, which is very different from the category label information of samples. Based on this, based on the typical correlation analysis algorithm, the correlation level information of the sample in the sorting problem is introduced into the dimension reduction technology, and a dimension reduction algorithm suitable for multi-modal data is designed. A large number of experiments show that the proposed algorithm can significantly improve the performance of image retrieval.
【学位授予单位】：天津大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP391.41
【参考文献】
相关期刊论文 前1条
1 王黎;帅建梅;;图像重排序中与查询相关的图像相似性度量[J];计算机系统应用;2010年11期
本文编号：2292057
本文链接：https://www.wllwen.com/kejilunwen/sousuoyinqinglunwen/2292057.html
