Research On Instance Retrieval Algorithm Based On Regional Representation

Thanks to the growth of the Internet and multimedia technology,the number of images on the network has increased rapidly.How to retrieve image efficiently and accurately has become one of the research hotspots of multimedia information retrieval.In the past ten years,image retrieval has been fully explored.From text-based image retrieval to content-based image retrieval technology,the performance of retrieval has made a qualitative leap.However,these traditional retrieval methods often use only shallow visual features and cannot cross the semantic gap.During these years,algorithms based on deep networks have emerged in many tasks of computer vision.Among them,the convolutional neural network has the characteristics of weight sharing,sparse link and multi-feature map,which is one of the most representative deep networks in deep learning.In image retrieval,pooling methods are required to generate compact feature representation for methods based on convolutional neural networks.The pooling method directly affects the final retrieval effect.Therefore,this paper has focused on the pooling method in convolutional neural networks.The main work is as follows:1.The advantages and disadvantages of text-based and content-based image retrieval are analyzed,and the development process based on content image retrieval is organized.2.The feature extraction process based on convolutional neural network is summarized,including convolutional neural network,pooling,dimension reduction and normalization.Then,the common similarity measurement methods are summarized and introduced.Finally,the retrieval datasets and the retrieval performance metrics are studied.3.The pooling method of image retrieval based on deep learning is studied,especially R-MAC pooling algorithm is studied.Although the algorithm utilizes the region information of the feature map to some extent,it only considers the local information whose contribution is large while not to fully consider the global information.Therefore,this paper proposes an entropy-based R-MAC improved algorithm,which adds entropy as a supplement to the feature representation of R-MAC,which further improves the distinguishability of features.The experimental results show that the retrieval performance has been significantly improved after the introduction of entropy.4.The effectiveness of the improved R-MAC algorithm based on the entropy is verified.And the re-ranking algorithm in which key area coarse positioning is carried out,further improves the retrieval performance.