Research On Image Retrieval Based On Deep Learning

In recent years,due to the development of mobile Internet,especially the popularity of social networking websites such as Flickr and Facebook,heterogeneous data such as images and videos are becoming more and more important on the Internet.Now,there are more than 1 billion users on Facebook producing more than 1 billion images per month,then users on Flickr uploaded as many as 728 million images in 2015,averaging about200 per day by per user.In China,Taobao,the country's largest e-commerce system,holds more than 28.6 billion image data on the back end.Therefore how to quickly and accurately retrieve and return images to users in a large number of image libraries containing rich visual information is a research hotspot in the field of multimedia information retrieval.In image retrieval,there is a lack of consistency between the information extracted from visual data and the interpretation of these data to users in specific situations.This phenomenon is called the semantic gap,and the semantic gap is a problem that cannot be ignored in image retrieval.This paper attempts to analyze the causes of the semantic gap from the perspective of image understanding,and attempts to solve it through image semantic based retrieval.The work of this paper focuses on the following sections:(1)This paper analyses the current research status of image retrieval technology in recent years,and then discusses the causes of the semantic gap.This paper turns out that the non-independent relationship between images tags leads to the problem of semantic gap.To solve this problem,we need to learn the semantic embedding of images,and the semantic embedding of images means the semantic information of images.(2)We proposed an image retrieval algorithm based on image multi-objective semantics.In this paper,a deep convolutional neural network is constructed to detect the classification and location information of multiple targets in the image,and after filtering and merging the information,we classify the image by the sequence generation model based on recurrent neural network.Finaly images are ranked by comparing the similarity of the image classification to get the retrieval task.(3)Considering the influence of different levels of image semantics on retrieval,this paper attempts to replace the multi-objective semantics with image concept semantics in the algorithm.The method first extract the relational short sentences from the image as the conceptual semantic content through the image caption and syntactic analysis,and then generate a model to classify which the image belongs by using the sequence generation model based on the attention mechanism,then wo rank the images by comparing the similarity of the categories.(4)In this paper,two image retrieval methods are trained and tested on multiple datasets by both single-label and multi-label retrieval,which proved that methods implemented in this paper have certain advantages.