用于图像语义检索的深度哈希算法

发布时间：2018-05-22 10:24

本文选题：图像检索 + 哈希算法　；参考：《浙江大学》2017年硕士论文

【摘要】：随着互联网的飞速发展,网上的图像资源也在以爆炸的速度增长,为了在如此大规模的数据中根据不同用户各自的需求来高效地检索相关的图像,哈希相关的检索算法应运而生。然而在现存的大多数传统哈希算法中,图像特征提取和后续的哈希函数学习是割裂的,这使得特征很难与哈希函数学习过程适配,影响了检索效果。基于深度学习的哈希算法的出现缓解了这一问题,但仍面临优化复杂难于训练,生成哈希码冗余性强等一系列问题。在本文中,我们提出了基于深度学习的语义检索哈希算法(Deep Hashing for Semantic Retrieval,简称DHSR),该方法使用深度卷积神经网络同时进行特征提取和哈希函数学习,实现了完全端到端的训练学习,将图像特征的提取和哈希函数的学习融入到同一个过程中,使这两部分互相适配促进,还提出结合了样本点标签和样本对标签两种监督信息以及量化误差的优化目标,使得神经网络模型可以在学习较好的个体语义特征的基础上,生成保持样本对原始相对位置关系的哈希码。在此基础上,本文进一步提出了一种基于神经网络的复合哈希码层次检索策略,通过单次训练神经网络生成两组哈希码,结合哈希表查找和哈希码排序两种方法的优点。另外,本文还采用分块全连接模块代替全连接部分,生成各位更加不相关的哈希码,并通过模型融合以及微调训练的方式以及一系列优化技巧提升算法的稳定性和准确率。大量实验结果证明,本文提出的DHSR方法在MNIST,CIFAR-10和NUS-WIDE三个公开图片数据集上都取得了比现有的哈希算法更好的效果。
[Abstract]:With the rapid development of the Internet, the image resources on the Internet are also increasing at the rate of explosion. In order to efficiently retrieve the related images according to the respective needs of different users in such large-scale data, Hash-related retrieval algorithm emerges as the times require. However, in most of the existing traditional hash algorithms, image feature extraction and subsequent hash function learning are fragmented, which makes the feature difficult to match with the learning process of the hash function, and affects the retrieval effect. The emergence of Hash algorithm based on deep learning alleviates this problem, but it still faces a series of problems, such as complex optimization and difficult training, strong redundancy of generating hash code, etc. In this paper, we propose a deep learning based semantic retrieval hashing algorithm called Deep Hashing for Semantic Retrieval,. This method uses deep convolution neural network to extract features and learn hash functions simultaneously, and realizes the complete end-to-end training learning. The extraction of image features and the learning of hash function are integrated into the same process, which makes the two parts fit and promote each other, and proposes the optimization goal of combining the two kinds of monitoring information of sample point label and sample to label, as well as the quantization error. The neural network model can generate a hash code to preserve the relationship between the sample and the original relative position on the basis of learning better individual semantic features. On this basis, a hybrid hash code retrieval strategy based on neural network is proposed in this paper. Two groups of hash codes are generated by a single training neural network, which combines the advantages of the two methods of hash table lookup and hash code sorting. In addition, the block full connection module is used instead of the full connection module to generate more irrelevant hash codes, and the stability and accuracy of the algorithm are improved by model fusion, fine tuning training and a series of optimization techniques. A large number of experimental results show that the proposed DHSR method is more effective than the existing hash algorithm in the three open picture datasets of MNIST CIFAR-10 and NUS-WIDE.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】