基于核典型相关分析与神经网络的跨模态哈希算法研究

发布时间：2018-01-10 02:12

本文关键词：基于核典型相关分析与神经网络的跨模态哈希算法研究　出处：《安徽大学》2017年硕士论文　论文类型：学位论文

【摘要】：近年来,随着各种模态数据的快速增长,如何实现跨模态数据之间的相互检索已经成为信息检索领域中的一个研究热点。哈希算法是一种非常有效的信息检索方法,具有占用储存空间小和检索速度快的优点。哈希算法可以分成单模态哈希算法与跨模态哈希算法,单模态哈希算法用于单一模态数据之间的信息检索,跨模态哈希算法用于不同模态数据之间的信息检索。哈希算法的核心思想是先通过哈希函数将数据转换成对应的哈希编码串,其中哈希编码串之间的汉明距离与原始数据之间的语义相似度相对应,然后通过计算哈希编码串之间的汉明距离返回检索结果。本文提出了两种不同的跨模态哈希算法:基于核典型相关分析的跨模态哈希算法(KCMH)和基于神经网络的跨模态哈希算法(NNCH)。1.提出了基于核典型相关分析的跨模态哈希算法(KCMH)。首先,通过锚点核典型相关分析算法(AKCCA)进行公共核空间的学习,将不同模态的数据映射到公共核空间中,该部分对同一模态内数据之间的相关性和不同模态间数据之间的相关性进行考虑。不同于现有的公共空间的学习算法,AKCCA算法结合了典型相关分析算法、K-means算法以及核技巧,可以非常有效地学习到数据之间的非线性关系。在学习到公共核空间之后,利用内积与汉明距离的对应关系提出相应的目标函数。最后,使用迭代优化算法进行最优化求解。2.提出了基于神经网络的跨模态哈希算法(NNCH),该算法成功地将神经网络技术应用到跨模态哈希算法领域中。神经网络是一种模仿生物大脑活动的简化生物模型,它在人工智能、计算机视觉以及语音识别等领域有着广泛应用。在NNCH算法中,提出了独特且非常有效的网络结构、损失函数以及神经网络的训练方法。NNCH算法的神经网络是由图片神经网络和文字神经网络组成,其中图片神经网络主要是对图片进行二进制编码,文字神经网络主要是对文字进行二进制编码,图片神经网络和文字神经网络通过共享分类器权重的方式连接起来。NNCH算法提出了将Softmax分类损失与哈希特性损失相结合的损失函数,并且利用该损失函数和caffe深度学习框架完成神经网络的训练。
[Abstract]:In recent years, with the rapid growth of various modal data. How to achieve cross-modal data retrieval has become a research hotspot in the field of information retrieval. Hash algorithm is a very effective information retrieval method. The hashing algorithm can be divided into single mode hash algorithm and cross mode hash algorithm, and single mode hash algorithm is used for information retrieval between single mode data. Cross-modal hash algorithm is used to retrieve information between different modes of data. The key idea of the hashing algorithm is to convert the data into the corresponding hash encoding string through the hash function. The hamming distance between the hash coding string and the semantic similarity between the original data corresponds to each other. Then the hamming distance between the hash coding strings is calculated to return the retrieval results. In this paper, two different cross-modal hash algorithms are proposed: the cross-modal hash algorithm based on kernel canonical correlation analysis (KCMH). And neural network-based cross-modal hashing algorithm / NNCH.1.A cross-modal hash algorithm based on kernel canonical correlation analysis is proposed. First of all. The common kernel space is studied by the anchor kernel canonical correlation analysis algorithm (AKCCA), and the data of different modes are mapped to the common kernel space. This part considers the correlation between the data in the same mode and the data between different modes. It is different from the existing learning algorithm in common space. The AKCCA algorithm combines the typical correlation analysis algorithm with K-means algorithm and kernel techniques, which can effectively learn the nonlinear relationship between the data, after learning the common kernel space. Using the corresponding relation between inner product and hamming distance, the corresponding objective function is proposed. Finally, the iterative optimization algorithm is used to solve the optimization problem. 2. A cross-modal hash algorithm based on neural network is proposed. The algorithm successfully applies neural network technology to the field of cross-modal hashing algorithm. Neural network is a simplified biological model that mimics biological brain activity and is applied in artificial intelligence. Computer vision and speech recognition are widely used. In the NNCH algorithm, a unique and very effective network structure is proposed. Loss function and neural network training method. The neural network of NNCH algorithm is composed of picture neural network and text neural network, in which picture neural network is mainly binary coding of pictures. Text neural network is mainly used for binary coding of text. Image neural network and text neural network are connected by sharing classifier weights. NNCH algorithm proposes a loss function which combines Softmax classification loss with hash characteristic loss. The loss function and the caffe depth learning framework are used to complete the training of neural network.
【学位授予单位】：安徽大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP183;TP391.3

【相似文献】