Feature Representation And Retrieval Of Instance

Research on content-based instance retrieval began since the 1990s.Through the development of social networks,the amount of multimedia information has grown rapidly.Relevant images describing the given instance in the specified pictures should be quickly and accurately searched from massive picture data.This is not only required in the social business field,but also a strong technical support for automatically datasets generated for other tasks of computer vision.The key technology of instance retrieval is to use as little as possible memory to store features that have strong descriptive capabilities,which can express the visual and high-level semantics of image.Studying the feature encoding of image makes full use of the advantages of artificial intelligence,improves the retrieval efficiency,and will obtain more efficient and accurate retrieval results.The main research contents of this article are as follows:First of all,this article proposes an unsupervised feature representation based on clustering and space weighting.Through clustering the image dataset,and set spatial weighting weights for different categories to adjust the distribution of similar samples in high-dimensional space,the proposed method fuses local features and resolution features to enhance the distinguishability of image feature representation and improve the accuracy and recall of search results.Secondly,this article proposes image feature representation based on LSTM context-aware network.Due to the uneven distribution of image features,the spatial relative relationship of the local information of the image is simulated.The long-term of the LSTM model can be used to sequence the top-to-bottom and left-to-right sequence features of the image.The Encoding method mines spatial relationships of features,and obtains feature representations with stronger description capabilities.Finally,this article also studies the optimization of reranking.According to the query expansion,the results obtained extra information after the first sort.By adaptively selecting the similarity threshold,each query image screens for accurate samples to supply additional information.The query expansion combined with data diffusion provides a more stable reranking effect and achieves the purpose of correcting the results.The innovative algorithms proposed are experimented and performed well on standard image retrieval datasets such as Oxford5k,Paris6k etc.