| Image retrieval is a technique that allows users to search for specific images accurately and quickly,and it is a cross-cutting issue in the field of applied mathematics and machine learning.This paper proposes new image retrieval algorithms based on convolutional network.First of all,two kinds of feature fusion methods are proposed to solve the problem of poor generalization ability of single-feature retrieval system,one is the fusion of deep features and traditional features,the other is the fusion of multi-network deep features.Experiments show that the fused features can express more details and have strong discriminating ability.In order to improve the retrieval efficiency,this paper proposes an image hierarchical retrieval method based on multi-feature fusion.Firstly,the image is pre-retrieved based on the binarized CNN features to obtain the sub-library.Then,we carry out accurate retrieval based on the fused features.Hierarchical retrieval strategy reduces the scope of the query,which effectively reduces the retrieval time.Due to the difference of descriptors and matching methods,some related images are often difficult to be retrieved.In this paper,a retrieval method based on query expansion is proposed.The results of initial retrieval are regarded as new queries,and the features are re-matched according to the proposed weighted distance.At the same time,this paper uses the kd-tree indexing method to store the database images,which avoids the search for most of the data.Experiments show that the proposed method combines accuracy and timeliness.Finally,the proposed algorithms are validated on four benchmark datasets:Oxford5k,Ukbench,Holidays and Cifar-10.An N-S score of 3.76 is obtained on Ukbench,with a mean average precision of 87.3%,88.42%and 73.60%on Oxford5k,Holidays and Cifar-10. |
