Research Of Deep Hashing Algorithm For Web Image Retrieval

Hashing has been widely used for web image retrieval applications due to its low storage cost and fast query speed.Most existing content-based hashing methods only consider visual features but ignore the valuable semantics involved in the associated texts.Furthermore,the hand-crafted features might not be optimally compatible with hashing learning procedure.Thus,these hashing methods with hand-crafted features may not achieve efficient satisfactory performance.In order to improve the retrieval performance of image hashing algorithm,the research of this paper is carried out as follows.First at all,a novel unsupervised visual hashing algorithm called Semantic Transfer Deep Visual Hashing(STDVH)is proposed.In this hashing algorithm,the semantic information of the training text is extracted by spectral clustering.Later,a deep convolutional neural network is constructed to transfer the text semantic information into the learning of the image hash codes.Finally,the image hash codes and hash functions are trained in a unified framework,and the effective retrieval of large-scale image data is completed in low-dimensional Hamming space.STDVH integrates the objective function of hash learning and the eigenfunction of the convolutional neural network in a joint objective function.This integration transforms hash learning into an optimization process of this objective function.After the optimization of the integrated objective function through the back-propagation algorithm,the high-level features and hash codes of images obtained can improve each other.Secondly,STDVH algorithm is expanded to semi-supervised condition.A semisupervised image deep hashing algorithm called Semantic Transfer Deep Visual Hashing for Semi-Supervised(STDVH-SS)is proposed for the condition that training data have limited semantic information.STDVH-SS uses a small amount of supervised information and a large amount of unsupervised text information to learn abundant semi-supervised semantic information,which can be transferred into the learning procedure of image deep hashing.Futhermore,deep hash learning algorithm based on unified framework can improve the semantic expression ability of image hash codes.Finally,in order to compare STDVH and STDVH-SS comprehensively,a web image retrieval system is constructed based on semantic transfer deep hashing algorithm in this paper.Only a part of image and text data of the database is needed in the training phase of the system.Hash codes and hash functions are learned through either STDVH or STDVH-SS algorithm according to the quantity of labels in training set.The rest of the database image is converted into hash codes through the learned hash function for storage.In the retrieval stage,only the query web image is needed to be provided.Then it is mapped into binary codes with hash functions.At last,the similarities between query image and database images are calcaulated in Hamming space,and the retrieval system returns database images in ascending order of distance.Comprehensive experiments are conducted on two publicly available image datasets Wiki and MIR Flickr.Results highlight various advantages of STDVH and STDVH-SS and demonstrate that these two algorithms significantly outperform several state-of-the-art content-based and cross-modal hashing methods from various perspectives.It also shows that,with semantic transfer,the valuable extracted semantics can be effectively encoded in the binary hash codes.At the same time,the image feature expression and hash learning can promote with each other in the unified deep neural network framework,which can improve the performance of web image retrieval.