Research On Case-level Image Retrieval Based On Deep Convolutional Neural Network

This paper focuses on the problem of instance-level image retrieval.Compared with the traditional object retrieval,instance-level image retrieval has a series of difficulties,such as:the difference between the same categories(for example,lighting,rotation,occlusion,cropping,etc.),the difference between categories and categories is not great(Coca Cola bottle and Sprite bottle),the image contains a large amount of interference information(such as background images)and a large number of unlabeled interference images.Recent developments have shown that Convolutional Neural Networks(CNN)can provide an image feature representation that is superior to traditional methods.However,the features extracted from the entire image by the convolutional neural network contain a large amount of interference information,which may cause the retrieval performance to be less than expected.In order to solve this problem,this paper proposes two solutions.The first is a method based on Faster R-CNN detection for instance-level image retrieval.It has two stages,namely Faster R-CNN offline training and online instance retrieval.First,the Faster R-CNN model is trained to locate the area where the object is located.Then,CNN features of the region where the object is located are extracted and integrated into the overall features of the image.Finally,the Euclidean distance between the overall features is calculated to obtain the search result.This paper carries out experiments on Instre and Oxford datasets respectively.The experimental results verify the effectiveness of the proposed method.The second is a new instance-level image retrieval framework.The framework consists of two phases.First,this paper uses the Regional Proposal Network(RPN)to detect the image,and its detection result is input into the Dual Regularized Triple Network(DLRTN).Second,by calculating the loss function of the ranking subnetwork and the classification subnetwork,and using the calculation results to optimize the network.Then,this paper introduces the regional generalized mean pooling(RGMP)layer,pools the feature maps from the output of the dual regularized triplets network and generates the regional generalized convolution activation(R-GAC)as the global image representation.Finally,experiments on image retrieval datasets demonstrate the effectiveness of the proposed image retrieval framework.