Research On Instance Image Retrieval Method Based On Deep Convolution Feature Aggregation

In recent years,the rapid development of computer technology has led to the explosive growth of image data,so the direction of computer vision image processing has become the focus of many researchers,and image retrieval technology has become one of the most popular research directions as the basic field of image processing.With the emergence of deep convolutional neural networks,content-based instance image retrieval has developed rapidly,and the problem of how to quickly and accurately retrieve images that people need from a complex image database needs to be solved urgently.Those primary thought from the instance image retrieval dependent upon convolutional neural system is that to extricate those features of the image,after that measure the similarity about these features,and at last get those retrieval effect.For the purpose of moving forward the retrieval accuracy,query expansion,reordering and other methods are usually used as supplements.Nowadays,deep convolutional neural networks bring been broadly utilized within the field from image retrieval.Furthermore,image retrieval utilizing convolutional neural networks get to be a prevalent technology.However,since the three-dimensional convolutional features obtained by the convolutional neural network are troublesome to utilize at ascertaining similarity.It is usually necessary to aggregate the three-dimensional convolutional features into one-dimensional features during similarity retrieval,and the information of convolution features cannot be avoided losing in this process.How to design a better convolutional feature aggregation method so that the obtained onedimensional descriptors can better delegate the message involved in the image has become the focus of research.This paper uses not only convolutional neural networks but also feature aggregation algorithms to carry on an in-depth research on instance image retrieval methods.The main work is as takes after:1.This paper uses the pre-trained VGG16,which is an unnecessary training process for drawing out the feature from image,and achieves good results.2.For the 3D convolutional feature maps obtained by deep convolutional neural networks,this paper first considers 3D feature aggregation from the perspective of global features.This paper proposes an improved feature aggregation algorithm Generalized R-MAC(GR-MAC)based on Lp norm,aiming at the traditional R-MAC aggregation method only considering the local information that contributes more in each scale.The article uses the fusion of the largest and the average eigenvalue in each scale,which greatly reduces the loss of information in the scale region,and is more conducive to obtaining the global information of convolutional features.3.On the other hand,the article considers the aggregation of 3D features in terms of salient features that are more conducive to retrieval.For traditional feature aggregation algorithms,the correlation,importance and saliency between feature channels are not considered.A convolutional feature aggregation method based on saliency weighting is proposed,which includes variance channel selection and nonparametric saliency algorithm weighting.It pays more attention to the convolutional features of the target region.After fusing it with the proposed Generalized R-MAC,it is more beneficial to obtain better global features to form the final image descriptor.4.This paper will be approved for various classic instance image retrieval datasets.The test outcomes indicate that the two feature aggregation modules proposed in this paper can promote instance retrieval,and the fusion of the two algorithms can achieve a better retrieval effect.Moreover,compared with the previous classical feature aggregation algorithms,our method significantly improves the retrieval accuracy.The retrieval results reach 79.87% on the Oxford5 k dataset,89.55%on the Paris6 k dataset,and 91.51% on the Holidays dataset.