Research On Image Retrieval Methods Based On Hashing And Deep Learning

With the rapid development of the internet,users produce enormous images data at PB rate on a daily basis.Facing so much data,users continuously enlarge their demand for searching images,especially in some e-Commerce industries.For instance,when users occasionally find their interesting items in the street or mall and they do not know their name,it is great hard for them to search on the internet or shopping-online store.Therefore,it is crucial for the e-Commerce company to develop an efficient and fast retrieval algorithm.Nowadays,hashing algorithm has been widely applied into our life.In fact,the essence of hashing algorithm is to convert high-dimension image presentation into low-dimension hash codes and use bit operation to calculate the similarity between any two images,which can ensure high efficiency when calculating the similarity.In the field of video,Instance Searching is an extremely significant application.For instance,in some criminal investigation industry,when polices or detectives intend to find the fames where suspects probably appear.If there is existing an algorithm to fast find out them,maybe it can reduce the searching time a lot.In early 1960 s,many image retrieval algorithm based on text content has been proposed by many researchers,for example,the early searching engines.At that time,many searching engine companies needed many people to label enormous images and provide a long description for the images,which is so time-costing and money-consuming.Fortunately,those algorithms obtained high performance.They still,However,exist many shortages: first,most images in the internet have no labels or tags and it is challenging to label them.Second,to store those images' content and index is highly space-consuming.To solve above problems,many researchers start to focus on nearest neighbors searching.At the earliest,researchers proposed plenty of algorithms most of which were based on nearest neighbors searching in high European space,such as HOG,SIFT and GIST.Later,people gradually found searching in high dimension space inefficient,especially calculating the image similarity.In the recent decades,people began to research how to use a low dimension feature to present a image.Naturally,hashing retrieval algorithms become a hot topic in the retrieval field.To address the above three crucial problems,a enormous number of scholars start pay a attention to the hashing algorithm.Therefore,we propose three efficient hashing frameworks to address the above three problems:1)Deep region hashing,which is aimed at solving the instance search problems.We are inspired by the up-to-date object detection algorithm(.e.g RCNN,Fast-RCNN and Faster-RCNN)and deep hashing methods,and proposed an end-to-end hashing work which be able to efficiently tackle the instance retrieval issue.From the multiple compared experiments,our performance has been seen a significant rise either on image retrieval or query speed.2)Deep discrete hashing with self-supervised labels,which focuses on the discrete hashing learning problems without relaxation operation by involving intermediary variable.In addition,we combine the complete training process into an end-to-end framework.In the aspect of experiments,we test our method on two tasks,image retrieval and image recognition,both of which show our results outperform the state-of-art.3)Binary generative adversarial network,which is utilized to retrieve images and reconstruct images.We combine the up-to-data generative adversarial network and autoencoding system and explore a novel deep network.We conduct many constructive experiments on three public datasets.The results show that our algorithm achieves the state-of-art results.Overall,hashing algorithm has great researching value and prospect.