Research On Cassini ISS Image Classification For Astrometry Based On Extracting Feature By Deep Convolution Network

Imaging Science Subsystem is carried by the Cassini Space Detector,which is a dedicated optical telescope system.It has acquired hundreds of thousands of images until now,but only a few of them can be used for astrometry.The traditional approach is to select images that can be used for astrometry from a large number of ISS images by professionals based on past astrometry experience,which is a heavy and time-consuming work.It will be important for the field of astrometry to find an automated ISS images filtering methods.The purpose of this study is to attempt to automate this process.The filtering of ISS images belongs to the category of image classification essentially.In this paper,an automated ISS image for astrometry classification system combined with the traditional support vector machine and convolution neural network technology is proposed.Then a large number of experimental tests for this system have been carried out.At the same time,several traditional feature extraction methods are used to compare with our methods.The proposed ISS image for astrometry classification system is divided into three parts: ISS image preprocessing,ISS image feature extraction and ISS image classification,in which the latter two parts are the key.Firstly,in the process of ISS image feature extraction,eight kinds of deep convolution networks-CNN-F,CNN-M,CNN-S,CNN-M-128,CNN-M-1024,CNN-M-2048,VeryDeep-16 and VeryDeep-19 trained on the ImageNet 2012 data set are adopted.The input data of these trained CNNs' output layer are used as the feature descriptors of ISS image.In addition,three kinds of traditional feature extraction methods-Hog,LBP and Gabor operators are also adopted to compare with the proposed image feature extraction method.Secondly,in the process of ISS image classification,the classifier used is SVM.However,because all the image need to be divided into three categories and the original SVM only support binary classification,two common multi-classification strategies – one-versus-one and one-versus-rest are applied.Additionally,the common linear function,polynomial function and Gaussian kernel function are selected as the SVM kernel functions respectively.Thirdly,because the ISS image is Vicar format,ordinary tools can't read it directly,it needs special treatment in the ISS image preprocessing stage.So,the storage structure of ISS images is analyzed,and then the function of reading ISS images is written by MATLAB,which lays the foundation for the further feature extraction of ISS images.Finally,through some experiments and analysis,it is found that the classification results obtained by the eight deep convolution networks mentioned in this paper are better than those obtained by using the traditional three image feature extraction methods under the same conditions.Moreover,the results show that the best classification system is the combination of the CNN-S and SVM in which the 2-order polynomial kernel function and 1-v-r strategy are used.The classification accuracy is over 97%.In general,the Cassini ISS image for astrometry classification system proposed in this paper basically achieves the automatic filtering of ISS images,and has achieved good results.At the same time,the system can also be extended to similar works in other space exploration project.