Key Technology Research Of Instance Search

With the rapid development of the Internet&mobile Internet,a large amount of images and videos have been generated,and they need to be analyzed and understood for quick information exchange and distribution.The instance image search has appeared a very practical value in this respect.In addition,deep learning speeds up the progresses of instance retrieval.This paper systematically explores the key techniques of instance search and proposes improved algorithms from following three aspects:An unsupervised image representation algorithm based on Indicator and Gram matrix for spatial-channel weighting is proposed.The Indicator matrix is first used to highlight the objects of interest,and then the Gram matrix is applied to intensify the distinctive channels.As a result,the correlation between the database images and different channels is measured and a more efficient global feature representation is obtained.To address the problem of the limited labelled samples,an end-to-end weakly supervised learning network is designed to generate a compact global feature representation.In addition,new loss functions,which can be trained in both unsupervised and supervised ways,are proposed to significantly improve the generalization ability of the model.An improved query expansion algorithm is presented.For the problem that the neighbor parameter is not robust,the similarity weighting algorithm is adopted to replace the average summation process.Afterwards,data diffusion is introduced addressed the problem that the internal structure of the dataset cannot be fully utilized.Moreover,an efficient image processing and computation is proposed to speed up the model training.The experimental results on multiple standard datasets such as Oxford,Paris and Holidays etc.indicate that the proposed algorithms significantly improve the accuracy of instance search.