大规模旅游景点图像检索

发布时间：2018-01-14 19:09

本文关键词：大规模旅游景点图像检索　出处：《安徽大学》2016年硕士论文　论文类型：学位论文

【摘要】：随着旅游业的飞速发展,网络上各个景点图片泛滥,图像数据越来越庞大,很多时候对于浏览的某张美丽的风景图片但是并不知道属于哪个景点。因此,如何从如此大规模的图像数据中快速准确的搜索到所需的景点信息是一个很实用并亟待解决的问题。近年来,大规模图像检索方法已日渐成熟,为能够从海量景点图像中准确地找到目标图像提供了可能性。论文结合近几年大规模图像检索和图像处理等方法来实现对大规模旅游景点图像的检索。论文主要工作如下：(1)研究了图像检索的基本理论知识及常用方法,主要包括：基于图像全局特征的描述符GIST和基于图像局部特征构建的词袋模型(bag-of-feature,简称BOF)。GIST特征是利用多组Gabor滤波器与图像卷积,并对图像进行网格划分,将不同网格的卷积结果级联得到图像的全局特征。但是这种特征提取方法往往依赖于对网格的划分而且只有在不同图像整体相似度较高的情况下检索效果好。BOF模型是借鉴于对文本检索的方法,一般是通过SIFT(scale invariant feature transform)算法提取图像局部特征并利用k-means进行聚类得到低维的视觉词典,用基于视觉词典的直方图向量表征图像特征。但是在图像种类较多时,视觉词典的维数会很大,不便于构建BOF模型。(2)为改进这两种方法在大规模图像检索中准确度等缺点,论文利用Alex等人提出的8层卷积神经网络作为特征提取的框架,取最后一个全连接层作为图像特征,并通过主成分分析方法对其进行降维,然后利用基于局部敏感哈希(locality sensitive hashing,简称LSH)的近似最近邻算法构建低维的特征索引结构。利用卷积神经网络提取特征的优势和哈希索引结构在检索中的高效率,解决了传统方法在图像检索时准确度等方面的不足。(3)论文利用该算法对北京市1740个旅游景点进行图像检索实验,结果表明,与上述两种算法相比,对于多数的景点该方法在检索的准确度上都有比较明显的优势,当景点图像库中的某个景点相似图像较少时,查询结果的准确率与前两种方法相差不大。
[Abstract]:With the rapid development of tourism industry, the network spread all the attractions of images, the image data is more and more large, most of the time to browse a beautiful scenery picture but does not know which belong to the attractions. Therefore, how to from such a large-scale image data quickly and accurately search the information required is a very practical and scenic spots the problem to be solved. In recent years, large-scale image retrieval method has become more mature, in order to massive attractions image accurately find the target image provides the possibility. To achieve retrieval of large-scale tourist attractions image based on recent years large-scale image retrieval and image processing method. The main work is as follows: (1) study the basic theory of knowledge and methods, including: image retrieval based on global image features descriptor GIST and build the image based on the local characteristic The bag of words model (bag-of-feature, referred to as BOF).GIST feature is the use of multiple sets of Gabor filter and image convolution, and meshes for the image, the convolution result of different grid cascade get the global features of the image. But this feature extraction method often depends on the division of the grid and only in different images under the condition of higher overall similarity good retrieval effect.BOF model is reference to the method of text retrieval, usually by SIFT (scale invariant feature transform) algorithm to extract local image features and clustering to get the low dimensional visual dictionary using k-means, characterized by image feature vector histogram based on visual dictionary. But in many kinds of image, the dimension of the visual dictionary very not easy to construct the BOF model. (2) in order to improve the shortcomings of these two methods in large-scale image retrieval accuracy, the use of Alex Proposed 8 layer convolution neural network as feature extraction framework, finally get a fully connected layer as the image characteristics, and method for its dimension reduction through principal component analysis, and then use the local sensitivity based on Hashi (locality sensitive hashing, referred to as LSH) the approximate nearest neighbor algorithm to construct feature index structure of low dimensional. The advantage of feature extraction and Hashi index structure in retrieval using convolutional neural network with high efficiency, solves the shortcomings of traditional methods such as the accuracy of the image retrieval field. (3) the use of the algorithm for image retrieval experiments on 1740 tourist attractions in Beijing city. The results showed that compared with the above two algorithms, for most the attractions of the method has obvious advantages in retrieval accuracy, when a tourist attractions image database of similar images is small, the accuracy of the query results from the first two There is little difference in the method.

【学位授予单位】：安徽大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】