Large-scale Image Retrieval Based On Deep Hash Learning

Image retrieval technology plays an important role in search engines,e-commerce,medical fields,industries and so on.In recent years,with the rapid development of Internet technology,especially the popularity of social networks such as Weibo and Douban,heterogeneous data such as images,videos,audios,and texts have been growing at a rapid rate every day.For pictures containing rich visual information,how to quickly and accurately retrieve the images required by users in these massive images database becomes a hot topic in the field of computer vision and information retrieval.The image retrieval methods based on hash learning is become a powerful tool for image retrieval and will become an effective solution for massive image retrieval,it liberates people from a lot of labor,material and financial resources.Image will increase continuously in the future Internet,and image retrieval technology will play an important role in corresponding areas.Image retrieval can be divided into Text Based Image Retrieval(TBIR)and Content Based Image Retrieval(CBIR).TBIR needs to mark the images with corresponding texts manually,so it is only suitable for small-scale image data,and it consumes a lot of manpower and financial resources to mark image data,so these methods are suitable only for small-scale image data.However,these manual works are not necessary for CBIR,so most scholars tend to focus on CBIR recently.The image retrieval method based on hash learning are CBIR methods,These methods map the image features from high dimensional space to low dimension space as Hamming space,and generates a low dimensional binary hash codes sequence to represent a picture.Hash learning avoids the dimension disaster of high dimension and reduces the requirement of computing memory in the image retrieval system,also can quickly respond to users' retrieval and becomes an effective solution to large-scale image retrieval.To address the above problems,we employed the deep hash learning method to learn binary hash codes of images.Deep learning,which is able to extract the potential relationship between the image data from the underlying data and extract the high level semantic information of images,is a powerful feature learning method.In information retrieval,hash learning,which maps the high dimensional features to a low dimensional Hamming space by the hash function and generates corresponding compact binary code sequences,is an efficient method.The representation of features in hash learning,which is not only simplified,but also able to be quickly compared with the Hamming algorithm when measuring the similarity,greatly improves the rate of retrieval and achieves the real-time requirement of retrieval.The main work of this paper consists of 2 aspects:(1)An end-to-end deep hash model based on deep residual network.Since convolutional neural networks can extract image features very well and represent them hierarchically,CNNs have been successfully applied to face recognition,image classification,object detection applications and so on.It is feasible to extract features of images by using deep residuals network to capture features,and then embed the semantic information of the tag,learn to hash,get the hash code of the image.Reduces the dimensionality of the image,reduces storage space and speeds up retrieval.In order to verify the performance of the proposed model.We applied the proposed method on a variety of image data sets and compares it with a variety of mainstream hash learning algorithms.The results shown that this end-to-end deep hash learning method achieves high retrieval accuracy,and at the same time,it achieves better results for finely classified image data retrieval.(2)Large-scale image retrieval model based on deep residual network and Iterative Quantization Hashing(ITQ).Firstly we employed deep residual network to extract deep features,then use ITQ hash learning algorithm to get the compact binary hash codes.The proposed method is experimentation on a variety of image datasets,and the result compared with several state of the art hash learning methods.The experimental results shown that compared with other methods,it is able to extract the feature by deep learning,and improve the accuracy of image retrieval.