The Aurora Image Retrieval Based On Convolutional Neural Network

Aurora is made between the sun and earth’s magnetosphere high-energy charged particles in the polar ionosphere ionization effect and a natural glow is generated,the aurora image research is helpful to explore the internal relations of interaction between the sun and the earth’s magnetic field,to study the effect of the sun the earth between law and solar-terrestrial physics research has very important significance.Deep learning models especially CNN have made amazing achievements in natural image retrieval.However,the aurora image retrieval is still carried out through artificial selection or traditional SIFT based methods.How to use the advanced CNN model to retrieve the aurora image is a new topic of great significance.Therefore,combined with the background of aurora images and the excellent performance of CNN,this thesis mainly does the following work:(1)Aurora image retrieval method based on convolutional neural network and VLAD coding.In order to solve the problem that the performance of aurora image retrieval is still poor by manual selection or traditional SIFT based method,this thesis combined the advantages of CNN and VLAD coding to carry out the aurora image retrieval.In this method,the traditional Vector of Locally Aggregated Descriptors(VLAD)is embedded into CNN,and the back propagation is used for optimization to achieve end-to-end auroral image retrieval.The local features that better represent the aurora images extracted by CNN are used to replace the traditional manual features based on SIFT.Then,the local descriptors are aggregated and reorganized by VLAD,and the triple loss function based on distance constraint is proposed to solve the problems of triple loss,which makes the inner class more compact.On the other hand,due to the different convolutional layer of the information contained in the extracted features are different,in order to obtain more abundant image characteristics,different characteristics of convolution layer fusion,the features of the different levels are complementary,and contains the aurora of high-level semantic information of the image and the underlying information,can be more comprehensive characterization of aurora image content.(2)Aurora image retrieval method based on deep hashing algorithm.An end-to-end deep hashing algorithm for aurora image retrieval is proposed based on the good performance of CNN in image feature extraction and the fact that hash coding can meet the retrieval time requirment of large-scale image retrieval.Firstly,spatial pyramidal pooling(SPP)and power mean transformtion(PMT)are embedded in CNN to extract multi-scale region information in the image.Secondly,a hash layer is added between the fully connected layer to mean average precision(m AP)the high-dimensional semantic information that can best represent the image into a compact binary hash code,and the hamming distance is used to measure the similarity between the image pairs in the low-dimensional space.Finally,combined with the idea of multi-task learning mechanism,the joint loss function is designed,combined with the auroral image label information and the auroral image pair similarity information to jointly train the network model,combined with the loss of the classification layer and the hash layer as the optimization objective,so that the hash codes can maintain better semantic similarity.The feasibility of the two methods mentioned above has been verified on the aurora dataset,and the experimental results show that the proposed method effectively improves the performance of aurora image retrieval.