Research On Image Retrieval Algorithm Based On High-Level Semantic Features

Image retrieval is a very popular research direction in the field of image processing.So far,it has undergone several decades of research,has produced numerous research results and has been widely used in today's society.Image retrieval is mainly divided into two parts: image feature extraction and similarity matching.To obtain accurate retrieval results,the best image features should be extracted firstly,and the second is to select the appropriate matching method,both of which are indispensable,therefore,this paper is devoted to the research of these two aspects,and has made improvements and sublimation on the basis of the existing algorithms.The main research content of this paper is image retrieval technology based on high-level semantic features.It is well-known that the "semantic gap" between the low-level features and the high-level semantic features of image has always been a difficult problem in the field of image retrieval,influencing the development of image retrieval technology,so how to overcome this problem and improve the accuracy of image retrieval has become the focus of this paper.This paper firstly introduces the background,significance and development status of image retrieval.Secondly,it elaborates the principle of related algorithms used in this paper.Then it proposes two image retrieval algorithms based on high-level semantic features: image retrieval algorithm based on the fusion information and the label voting strategy,and image retrieval model based on multiple objective region.For the phenomenon of "semantic gap",this paper adopts the popular convolutional neural network algorithm and combines the category probability of images,proposes an image retrieval algorithm based on the fusion information and the label voting strategy.CNN is used to extract features of the image features and category probability respectively,and then combine both of them.In the process of similarity matching,the tag voting strategy is combined to sort the retrieval results according to the voting score.Experiments on three well-known image datasets caltech101,caltech256 and corel10 k show that compared with the traditional algorithms,the retrieval accuracy of our algorithm in all three datasets has been greatly improved,the increase of mAP value is 21.8%-59.65%,which shows that our algorithm is effective improve the accuracy of image retrieval.At the same time,for the problem of the image retrieval based on global feature algorithm can not process the multi-objective retrieval problem well,an image retrieval model based on multiple objective region is proposed,and an efficient retrieval algorithm is coded.Firstly,the object detection algorithm is utilized to locate the multiple objective regions of the image,and then the Convolutional Neural Network is utilized to extract their features.Finally,the newly proposed similarity measure based on multiple objective regions is utilized to measure the similarity between it and every database image and the retrieval result is sorted accordingly.Experiments on two datasets PASCAL VOC2007 and PASCAL VOC2012 show that,compared with other image retrieval based on the objective region algorithm,the proposed algorithm has better performance in the retrieval task of multi-object image.