Research On Visual Saliency Oriented To Image Retrieval

A key problem of content-based image search is how to correctly understand the intention of the user,and establish correct correlation between query images and images in the database.Since there is a strong correlation between the visual saliency area of an image and the user's visual attention,it is possible to improve the accuracy of image search by introducing visual saliency into image search.However,most of the existing image search models directly use the existing saliency model to extract the salient region of the image,and then weighted the correlation between the images according to the intensity of saliency.However,because of the uneven prediction performance of the current saliency models,it is difficult to accurately evaluate the role of real saliency in image search.In addition,the current saliency models are not optimized for image search tasks,and the performance promotion space is limited.Therefore,this paper aims at these problems,and carries out targeted research.The main research contents are as follows:(1)We build a real saliency annotated dataset for image search.In order to verify the function of true saliency in image search,a saliency evaluation database for image search is first annotated by using an eye movement instrument.In addition,based on this database,we design four saliency embedding methods to verify the significance of saliency in the image search task.(2)The research of adding the visual saliency information to bag of word model.This paper studies four methods to add the human visual saliency information to bag of word model,including:saliency filter;saliency intensity embedding;saliency representation embedding;saliency representation re-ranking and combination of four methods.Selecting the optimal threshold by threshold selection method.The visual saliency experiments of human eyes prove the important:role of visual saliency information on image retrieval task.The retrieval results improve 10 percentage on BOW model and improve 3 percentage on BOW+HE model.(3)Image search enhancement algorithm based on model integration.The main idea of the algorithm is to achieve the advantages of all models by integrating a variety of saliency models to improve the accuracy of the image search.In this algorithm,we integrate 9 classical saliency models,and analyze the essence of saliency models in image search.The RCM+CAF method is proposed to verify the importance of the stability of saliency models and the consistency of saliency models for similar images prediction to image search.The experimental results show that the proposed algorithm realizes the complementary advantages of the saliency models and significantly enhance the accuracy of the image search.And the retrieval accuracy of this algorithm is superior to that of the single saliency model embedded image search algorithm.