A Research On Image Retrieval Based On Semantic Hashing

As an important research branch in the field of computer vision,large-scale im-age retrieval tasks are aimed at finding effective image representation.Binary hashing has become the leading method to learn image representation of image retrieval system,for its computational and storage efficiencies of binary hash codes.However,how to overcome the semantic gap between the underlying visual features and the high-level semantic information,in order to make the image representation closer to the user's un-derstanding,remains a key problem in the current image retrieval field.To handle with the research challenges of image retrieval,this paper attempts to propose the following three solutions:Firstly,a Triple Hashing Semantics Ranking(THSR)based method for image re-trieval.A good hash function helps to map high-dimensional visual features into low-dimensional binary hash codes as effective image representation,which will signifi-cantly reduce the computational cost when similarities between samples are computed.THSR utilizes Word Embedding method in Natural Language Processing to convert labels into semantic vectors to maintain as much semantic information of labels as pos-sible.Then relative distances of semantic vectors are exploited to define a soft-margin triple loss.The motivation of soft-margin triplet loss is to adjust the boundary condi-tions for each triplet to stop training dynamically,and enhance the semantic expression ability of hash codes.The extensive experiment performed on two benchmark data sets commonly used in image retrieval also verifies the effectiveness of the THSR.Secondly,a Deep Semantic Hashing Based(DSHB)model for image retrieval.De-signing deep end-to-end networks suitable for image retrieval has become vital research interest in the image retrieval community.It's^shown that the deep convolutional archi-tecture implementin g classical manual image feature learning methods,such as SIFT,VLAD,BoW and FV,could achieve desirable performance.Inspired by the idea,DSHB combines three deep modules,ROI Align,Fully Connected layer and Normalization layer with deep convolutional blocks to learn binary hash codes.Besides,a soft-margin triplet loss similar to mentioned THSR's is constructed with semantic vectors,aim-ing at learning hashing layer by end-to-end manner.Since commonly-used image sets of instance-level image retrieval tasks usually contain few labels or similar labels,in which word vector of label is not available,DSHB takes a hard-margin strategy to deal with such tasks.The result of the comparative experiment proves the rationality of the structure of the DSHB,and the benefit of soft-margin strategy to improve the retrieval performance.At last,a Semantic Clustering Based Hashing model(SCBH)for multi-label image retrieval.The content and theme of multi-label images are more complex than single-label images,which are closer to the actual application scenarios of image retrieval tasks.Effective similarity measurement between images tends to be a key challenge in multi-label retrieval community,expecting the learned hash codes to keep multi-label semantic information as much as possible.SCBH improves the common label similarity strategy,by clustering the word vectors,with help of Word Embedding independent from the label distribution into K groups,aiming at transforming multi-label retrieval task to single-label retrieval task.The goal of SCBH is to avoid information loss during transforming process as possible.The extensive experiment verifies the effectiveness of the SCBH.