Image Recognition Method And Application Of Galaxy Pair Based On Multi-view Clustering

As one of the important research fields of data mining and artificial intelligence,image recognition technology has been greatly developed and is also widely used,including medicine,aerospace,communication and so on.Galactic merger is the main form of the evolution of cosmic celestial bodies,so the search and identification of galaxy pair is one of the main goals of the sky survey project.Due to the limitations of observation technology and the brightness and size of target galaxies,the characteristic information provided by a single image is relatively limited.Therefore,in the thesis,multi-band image recognition methods for galaxy pairs in SDSS are studied and multi-view spectral clustering frameworks are designed.By capturing the underlying internal relationship structure of features and features,features and views,and views and views,the clustering performance is improved.And at the same time,it provides strong evidence for astronomers to study the possible physical mechanisms of galaxy pairs.The main research contents are as follows:(1)SDSS image feature extraction method based on statistics and connected domains.Since the FITS image format is different from the common format,such as png and jpg,the method firstly reconstructs the target image based on the idea of statistics and connected domains;secondly,it extracts the features in the target region of the reconstructed image,and these feature data are collated.(2)Regularized multi-view spectral clustering based on adaptive weights and manifold mapping(MuSC).The algorithm first uses the idea of information entropy to explore the consistent and complementary information within and between views,and calculates the contribution of each feature and each view of multi-view data to the optimal clustering.The greater the contribution,the greater the assigned weight,and vice versa.Secondly,the manifold space and multi-view spectral clustering are combined,and the fusion and representation of multiple views are completed in the manifold space,which enhances the graph cut effect of multi-view spectral clustering,and further improves the clustering effect.Finally,extensive experiments are conducted on multiple multi-view datasets.The proposed algorithm outperforms the comparison algorithms in four evaluation indicators and the results demonstrate the efficiency of the proposed clustering algorithm.Experiments are carried out on standard data set of SDSS images and the proposed algorithm exhibits significant superiority by four evaluation indicators.The performance of the regularized multi-view spectral clustering algorithm based on adaptive weights and manifold mapping are further verified.(3)A prototype system of multi-view clustering analysis based on M-MuSC is designed and implemented.In order to fully mine the consistent and complementary information of multiple views,a multivariate MIC-based weight learning strategy is deployed on the MuSC algorithm,and an advanced MuSC algorithm M-MuSC is proposed.On this basis,a multi-view clustering analysis prototype system based on M-MuSC is designed and implemented,and the structure,functional modules,interface design and operation results of the system are introduced and analyzed.