基于深度卷积神经网络的语义图像检索研究

发布时间：2018-12-23 13:36

【摘要】：图像检索问题作为计算机视觉的一个基本问题,已经有很多学者对它进行过分析和研究。本世纪初,人们逐渐认识到低级特征和高级语义特征之间固有的“语义鸿沟”,开始涉及基于语义的图像检索,但收效甚微。近年来,深度学习技术为饱受到传统人工特征困扰的基于语义的图像检索问题打开了新的大门。相比较传统方法,深度学习模型在挖掘深层次特征上有惊人的能力,更可能满足图像语义特征学习。但是,深度学习本身也处在一个快速发展的阶段,并不是完全成熟的工具,在使用过程中需要克服一些深度学习本身带来的问题。例如,深度模型在实际训练中的参数优化问题。同时我们该选择什么样的深度模型来应对基于语义的图像检索问题,是我们首先需要考虑的。本文梳理了图像检索技术和深度学习技术的发展历程。在此基础上,提出了一种基于语义的图像检索框架。该框架中主要包含三部分:(1)特征提取网络,在本文中称为多层级的图像语义特征提取网络,框架通过该网络提取语义特征;(2)特征存储结构,它将提取出来的特征降维并转化成图像语义距离度量公式可以计算的格式;(3)图像语义距离度量公式,框架最终通过它实现图像语义距离的计算。多层级的图像语义特征提取网络是该框架的核心,它的性能直接影响着整个框架的最终精度,该网络能通过其层级结构分次对图像中的语义特征进行提取,独特的结构设计使得它可以挖掘图像中更深层的语义。特征存储结构将提取出来的图像语义特征转化成一种更低维更易计算的融合特征表,而图像语义距离度量公式则将通过融合特征表计算图像之间的语义距离,从而生成检索结果序列。在基于语义的图像检索框架中,多层级的图像语义特征提取网络可以被替换成其他分类网络,以灵活地应对不同的待检索图像。在多个数据集上的测试结果表明,与现有方法相比,本文框架能够更加准确地完成基于语义的图像检索工作。
[Abstract]:As a basic problem of computer vision, image retrieval has been analyzed and studied by many scholars. At the beginning of this century, people gradually realized the inherent "semantic gap" between low-level features and high-level semantic features, and began to involve semantic-based image retrieval with little effect. In recent years, depth learning technology has opened a new door for semantic image retrieval, which is beset by traditional artificial features. Compared with the traditional methods, the depth learning model has an amazing ability in mining deep-level features, and it is more likely to satisfy the semantic feature learning of images. However, deep learning itself is also in a rapid development stage, is not a fully mature tool, in the use of the process need to overcome some of the problems caused by the depth of learning itself. For example, the parameter optimization problem of depth model in practical training. At the same time, we should consider what depth model we should choose to deal with the semantic image retrieval problem. This paper reviews the development of image retrieval technology and depth learning technology. On this basis, a semantic-based image retrieval framework is proposed. The framework consists of three parts: (1) feature extraction network, which is called multi-level image semantic feature extraction network in this paper, is used to extract semantic feature; (2) feature storage structure, which reduces the dimension of the extracted feature and converts it into a format that can be calculated by the image semantic distance metric formula; (3) the image semantic distance measurement formula, by which the frame finally realizes the calculation of the image semantic distance. Multi-level image semantic feature extraction network is the core of the framework, and its performance directly affects the final accuracy of the whole framework. Unique structural design allows it to dig deeper semantics in images. The feature storage structure transforms the extracted image semantic features into a fusion feature table with lower dimension and easier to calculate, and the image semantic distance measurement formula calculates the semantic distance between images through the fusion feature table. Thus, the retrieval result sequence is generated. In the semantic-based image retrieval framework, the multi-level semantic feature extraction network can be replaced with other classification networks to flexibly deal with different images to be retrieved. The test results on multiple datasets show that the proposed framework is more accurate than the existing methods in image retrieval based on semantics.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【相似文献】