基于底层特征融合的图像检索算法研究

发布时间：2018-02-14 05:46

本文关键词： 图像检索底层特征融合特征降维位平面熵多矩结构描述符　出处：《山东师范大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着Flickr、Facebook等社交网站的流行,图像资源正在以惊人的速度不断增长,如何从海量的图像中快速有效地提取用户所需要的资源已成为人们工作和生活中必须解决的关键问题。因此,不少学者进行了卓有成效的研究,尤其是在对图像中的单一底层特征或者综合特征的提取算法方面取得了丰硕成果。遗憾的是,单一底层特征算法只提取一种图像特征,如颜色、纹理、形状等,这属于部分特性,忽略了全局空间信息,将导致检索结果不准确。综合特征算法大多是从查询图像中直接提取所需的特征,未考虑特征之间的关联性,具有维数高、计算复杂的缺点。为解决上述问题,同时结合人类视觉机制对图像空间特性的感知特点,提出将图像中的底层特征进行有效地融合,使融合后的特征尽可能多地表达图像信息,从而提高检索精度和效率。具体创新性研究成果如下:(1)在图像底层特征提取过程中,往往存在大量的高维特征,这些高维特征不仅增加了计算复杂度,而且在一定程度上影响底层特征的准确性。为此,本文在不影响图像表达效果的前提下,在底层特征提取过程中采用对图像特征量化的方法进行降维,从而提高了图像的检索效率。(2)人类视觉感知机制对图像的空间结构信息较为敏感,尤其是对图像中分层信息和角度变化信息。针对这一特性,本文提出一种对颜色特征进行分层、对纹理基元特征进行角度变化,然后将两种特征变化相融合的算法。首先,颜色特征微观部分利用颜色直方图,刻画每种颜色的像素和占整个图像总像素和的比例;宏观部分应用位平面熵对颜色特征进行分层,取特征较明显的前4层,并对每层的位平面熵加权;然后,根据5个方向不同的基本结构基元中各像素点的颜色值和角度值的统计信息,结合颜色特征,实现图像检索。实验结果表明,融合后的算法能有效的表达图像空间结构信息,提高了准确率和召回率。(3)由于局部特征更能表现图像的细节,同时根据颜色、纹理各自的特点和对图像表达所起的作用,提出了一种新的图像特征描述子,命名为多矩结构描述符(Multi-Rectangle Structure Descriptor,MRSD)。MRSD基于纹理基元的空间结构定义了3种结构描述符,用一种7-4-2-1加权方法突出每种结构描述符在特征表达中所起作用的不同,这样更能深刻地表达局部特征的重要性。同时融合了更契合人类视觉感知机制的HSV颜色模型,能有效地表达底层颜色信息。实验结果表明,将HSV颜色模型和MRSD融合能够更好地表达图像特征信息,解决了单一特征表达信息量不足的问题。
[Abstract]:With the popularity of social networking sites such as Flickr-Facebook and other social networking sites, image resources are growing at an alarming rate. How to quickly and effectively extract the resources needed by users from massive images has become a key problem in people's work and life. Many scholars have carried out fruitful research, especially in the extraction algorithm of single bottom feature or synthetic feature in the image. Unfortunately, the single bottom feature algorithm only extracts one image feature. Such as color, texture, shape and so on, which are part of the characteristics, ignoring the global spatial information, will lead to inaccurate retrieval results. In order to solve the above problems and combine the perception of human visual mechanism to the spatial characteristics of images, we propose to fuse the underlying features of images effectively. The fusion features can express image information as much as possible, so as to improve the retrieval accuracy and efficiency. The specific innovative research results are as follows: 1) in the process of image bottom feature extraction, there are often a large number of high-dimensional features. These high-dimensional features not only increase the computational complexity, but also affect the accuracy of the underlying features to some extent. In the process of feature extraction, the method of image feature quantization is used to reduce the dimension, which improves the retrieval efficiency of the image. (2) the human visual perception mechanism is sensitive to the spatial structure information of the image. Especially for the layered information and angle change information in the image, this paper proposes an algorithm to layer the color feature, change the angle of the texture primitive feature, and merge the two features. The microscopic part of the color feature uses the color histogram to depict the proportion of the pixel sum of each color to the total pixel sum of the whole image. Then, according to the statistical information of color value and angle value of each pixel in five basic structure primitives with different directions, the image retrieval is realized by combining the color features. The experimental results show that, The fusion algorithm can effectively express the spatial structure information of the image, improve the accuracy and recall rate. The local features can represent the details of the image more effectively. At the same time, according to the characteristics of the color, texture and the function of the image expression, the fusion algorithm can express the spatial structure information of the image effectively, improve the accuracy and recall rate. In this paper, a new image feature descriptor named Multi-Rectangle Structure descriptor MRSDG. MRSD is proposed. Three kinds of structure descriptors are defined based on texture primitives. A 7-4-2-1 weighting method is used to highlight the different functions of each structure descriptor in feature representation, which can express the importance of local features more profoundly. Meanwhile, the HSV color model, which is more compatible with human visual perception mechanism, is fused. Experimental results show that the fusion of HSV color model and MRSD can better express image feature information and solve the problem of lack of information in single feature expression.
【学位授予单位】：山东师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】