Research On Local Feature Aggregation In Image Retrieval

Nowadays,computer science and technology are developing rapidly and image data is growing vigorously,image processing is inseparable from the lives of the people.According to the needs of people,how to search accurately and efficiently in hundreds of millions of images has always been a research topic in the field of image retrieval.In the 1970 s,researchers proposed text-based image retrieval(TBIR).In the feature extraction stage,TBIR requires manual labeling of images,which consumes a lot of manpower and time.Content-based image retrieval(CBIR)has evolved and matured since the 1990 s.The CBIR extracts image features according to the properties of the image itself,so that the features obtained can describe the image more accurately and get rid of the dependence on manual annotation.In CBIR,the retrieval algorithm based on SIFT feature aggregation is the most commonly used.SIFT features are well distinguishable and robust.At present,the mainstream SIFT feature aggregation algorithms include VLAD,Bo W,Fisher Vector,and T-embedding.Generally speaking,the process of generating image global descriptors from local feature aggregation can be divided into two steps: embedding and aggregation.The embedding process maps low-dimensional local features to high-dimensional space,and calculates residual between local features and clustering centers.The final image descriptor is generated by summing residuals or taking the mean of residuals in the aggregation process.For example,VLAD distributes local features,the principle of distribution is to assign the cluster centers closest to the local features,and then use additive aggregation after calculating the residuals.T-embedding calculates the residual of local features and all clustering centers in the process of embedding,and retains the direction information of the residual,and then generates image descriptors by democratic aggregation.T-embedding balances the contribution of local features to image similarity matching.However,this kind of approximation-completely fair aggregation method is insufficient,the importance of local features to images is different.The local feature retrieval algorithm provides saliency information of features when extracting SIFT feature.The saliency information describes the importance of features to the image.Based on this idea,this paper proposes the Saliency T-embedding method,which is improved by the saliency of local features.The Saliency T-embedding method proposed in this paper improves the aggregation process of local features by using the saliency of local features.Saliency T-embedding method distinguishes local features of images by the saliency information,which represents the importance of features.The local features with high saliency play a major role in the aggregation process,and the local features with low significance play a supporting role.Saliency T-embedding realizes the control of local features by adjusting the distribution of local features,which improves the aggregation of local features.Nowadays,the deep convolution image retrieval algorithm has made remarkable achievements in the field of image retrieval.In this paper,the Saliency T-embedding method is applied to SIFT-based and CNN-based retrieval algorithm respectively.After many experiments on INRIA Holidays and Oxford datasets,the proposed aggregation algorithm based on local feature saliency has better retrieval effect.