Research And Application Of Image Retrieval Technology Based On Deep Learning

"Fourteenth Five-Year Plan" outline points out: accelerate digital development,create new advantages of digital economy,build a digital China.Digitization is bound to lead social change profoundly,and the realization of digitization is a major mission for contemporary scientific research workers to combine theory with reality.The realization of comprehensive digitization needs a powerful computing center and a good algorithm system.The research of image retrieval is an important part of digital society.How to quickly and accurately in a huge image to retrieve the required images is a challenging task,in this paper,on the basis of the traditional image retrieval is improved,the classic algorithms for further improvement,combined with the depth of the deep learning technology in extracting semantic features advantages,developed a deep learning framework based image retrieval algorithm can deal with huge amounts of data quickly,main content and innovation are as follows:1?Given the deep learning in terms of depth of extract semantic features powerful ability,based on the bilinear attention Inception-v3 module as the core to join mechanism depth of neural network structure,the size of the data at the same time in order to deal with a defect data and further extracting image features,this paper generated against network architecture,the building against enhancement module,data feature extraction ability to strengthen the backbone network.The comparative tests on CUB-200-2011,Stanford CARS and FGVC AIRCRAFT three fine-grained image data sets prove that the method presented in this paper has a better extraction performance for the distinguishable row features of fine-grained images,thus laying a good foundation for the accuracy of retrieval.2?In terms of index construction,this paper adopts the hashing algorithm based on deep neural network,and it has a cascade relationship with feature extraction network,which is helpful for the training of the overall model.In this paper,the concept of hashing center is introduced and a new central similarity measurement method is proposed to quickly generate high quality binary hash codes,and the central quantization loss function is proposed to train the convergence of the model.With Res Net50 and feature extraction in this paper the network as backbone,with CNNH,DNNH similarity measures,DHN,Hash Net and center hash layer for the hash code learning,with Image Net,MS COCO,PASCAL VOC,Veg Fru,Food101 for experimental data sets more groups of cross experiment,proved in this paper,the feature extraction of backbone and center similarity measures hash retrieval framework on three sets of index obviously superior to other combinations,with high precision and the retrieval speed is quick and easy training,triple traits.3? On the basis of the retrieval algorithm,this paper uses Java and C++ to test the physical and chemical experimental image key frame retrieval system in the development of Windows and Cent OS platform,and verifies the feasibility of the algorithm.