Deep Hash Image Retrieval Method Based On Hierarchical Semantics

With the rapid development of Internet and the popularization of smart devices,image data has made an explosive growth on the Internet,how to retrieval the required image from the large image database quickly has become a burning question,and image retrieval technology has provide technological support for this purpose.Now,on account of the progress of content-based Image Retrieval,especially the deep learning algorithm has provided powerful image features for image retrieval,which mitigates the ”semantic gap” problem and enhance the accuracy of image retrieval.On these bases,this paper presents the problem of mitigating the semantic gap based on the image retrieval of hierarchical semantics.The model find the correlations between the class by creating hierarchical semantics tree,and represent it as semantic tag vector which integrated it into deep learning model to extract high quality image features.Compared with traditional image features,this model takes into account the correlation between hierarchical semantics and improves the retrieval accuracy.Although we can get high quality image features from deep learning Internet model based on hierarchical semantics,but it also caused the high dimension features.In order to ensure the efficiency of retrieving large data sets,this paper introduced Hash method and presented image retrieval model based on deep Hash,added the Hash layers in RestNet network ends and designed new loss function.Added regular terms in loss function to replace add-threshold to do binary constraint which made the Hash layer learn Hash functions effectively and output binary hash code.The machine can finish the retrieval task according to compare the low-dimensional binary codes of the image.On these bases,this paper integrated two model ways,designed image retrieval system based on hierarchical semantics and deep Hash.The system adopts the hierarchical retrieval strategy from ”rough” to ”fine”to do the image retrieval.In the end,experiments are carried out on Cifar100 and ImageNet data sets to verify the effectiveness of the proposed method,the accuracy of the CNN model is 6%-12% higher than the traditional image retrieval method based on the hierarchical semantics.While compared with the no-Hash method,the CNN model based on Hash improves the retrieval speed bu nearly 10 times.The retrieval system presented in this paper integrates the advantages of both,which not only improves the speed of retrieval,but also guarantees the accuracy of retrieval.