Research On Image Retrieval Based On Core Semantics

Image retrieval is an important research direction of image understanding.With the rapid increase of data,the demands of image retrieval system are getting higher and higher.The method of image retrieval based on semantic gradually enter the sight of people,and the automatic image semantic annotation becoming one of the important parts of it.This paper devoted itself to the research of image automatic semantic annotation and image semantic retrieval.First of all,a quick glance at an image is sufficient for human to point out and describe an immense amount of details about the visual scene,based on that,this paper try to simulate human visual system,proposed a method of image automatic semantic annotation based on deep residual network.The approach is mainly divided into two parts as semantic alignment model and description generation model.In the training of semantic alignment model,it use the deep residual network to extract the image features,aiming at perfect the semantic alignment model to improve the result of automatic descriptions.Second,the sentence annotation is detailed which may well describe the image contains and what they are doing,but sometimes,people just want to know the core semantic of images.With regard of this,this paper proposed a measure based on multimodal recurrent neural networks to generate core semantic of images.The main work is in the dependence of frame work of image automatic semantic annotation model,according the characteristics of the neural network structure,pick out the word which can describe the image semantic.The generated annotations express the core semantic of images well.At last,this paper proposed an image retrieval algorithm based on image core semantics.The purpose of semantic description is to better realize the process of semantic retrieval,therefore,this paper automatically annotated the sample query image and the candidate image library with the algorithm proposed of the core semantic generated model.After that,according compare the text results of the auto-annotated image core semantics to find out images that are similar to the core semantic of the query sample image.In the end,it demonstrate the effectiveness of the method.