Research On Large Scale Image Hashing Retrieval Based On Deep Learning

Content-based image retrieval aims at searching for images which are similar to query images on image database and it has been widely used in various applications such as search engine,E-commerce and medical systems.With the development of the Internet and the population of digital device,the dramatic growth of image database leads to low time efficiency of retreival.Thus approximate nearest neighbor(ANN)search that applies to large scale image retrieval is proposed.Among existing ANN search methods,hashing is one of the most popular techniques due to its fast search speed,low storage cost and good retrieval performance.Traditional hashing methods map hand-crafted features into short hash codes which should preserve semantic similarities in original space.However,hand-crafted features can only represent low-level visual information and thus there are semantic gap.Recently,deep features that can represent high-level semantic information of images are used in image retrieval and good performances are achieved.Consequently,deep convolutional neural network and hashing are combined to form the methods called deep supervsied hashing which reduce the semantic gap and improve retrieval precision.Existing deep supervised hashing methods can be roughly grouped into metric-learning-based and classificationbased methods.The former methods which learn complex semantic relationship by a large number of pair or triplet has a slower convergence speed and a higher retrieval precision.The latter methods which embed a hash layer in classification networks has a faster convergence speed and a lower precision.In view of the shortcomings of metric-learning-based and classification-based methods,this thesis propose improved methods respectively to speed up training process and improve retrieval precision.Our contributions are summarized as follows:(1)Deep multiple-negatives supervised hashing based on metric-learning is proposed.In existing metric-learning-based methods,only the semantic similarities in pair or triplet are preserved and rich supervised information is not made full use of,which result in a slow convergence speed and unsatisfactory retrieval precision.To address this issue,this thesis propose multiple-negatives learning method which preserve the semantic relationship among a query,its positive and multiple negatives simultaneously.Firstly,we propose an efficient strategy to construct multi-negative-tuplet which can obtain adequate multi-negative-tuplet with a small number of images.Then,we propose the cost function based on multi-negative-tuplet which enables a query to interact with multiple negatives concurrently during the training process.Because supervised information is made full use of in this method,retrieval precision and training efficiency are improved.In addition,margin,a parameter is incorporated into the cost function to enlarge the hamming distance among dissimilar images and retrieval precision is further improved.(2)Deep large-margin supervised hashing based on classification is proposed.Existing classification-based deep hashing methods do not well in controlling the hamming distance among dissimilar images and thus the retrieval performance is lower.Based on the framwork used in image classification and the large-margin classification method,this thesis propose large-margin-classification function which are suitable for single label and multi-label data respectively.This function can enlarge the distance among classification-planes and can make the hamming distance among images in difference classes as large as possible.Besides,center loss is added to make the hamming distance among images in same class as small as possible.Hence this method can better control inter-class and intra-class hamming distance of images.Besides,this method has a faster convergence speed as a result of that the framework of classification is utilized.In this thesis,extensive experiments are conducted on several benchmark datasets.The experimental results validate that the proposed methods outperform existing state-of-the-art hashing methods in terms of retrieval performance and convergence speed.In addition,we analysis the impacts of deep convolutional neural networks and hyper-parameter in proposed methods and determine the optimal hyper-parameters.