Research On Classification Of Marine Vessels Based On Feature Fusion

The classification of vessels in the sea is an important issue considered for achieving maritime safety and traffic control.It has the general application prospects in both civil and military industries.Compared with other target recognition,the visual recognition of the vessel in the sea is more affected by environmental factors,such as illumination,perspective,scale and background,so its classification is more difficult than others.Tlhrough the research on the classification methods of existing marine vessels,it is found that the following problems exist:firstly,how to extract the features of images to better represent the target has always been one of the key issues in the field;secondly,most of the existing methods only select a single feature or a simple concatenation feature to describe the image,these methods greatly reduce the complementary efficiency of features.The feature structure fusion method uses a unified metric to define the structure matrix of features,and can effectively fuse different features.However,it is based on the unsupervised learning method and the feature category information is not considered in the feature fusion process,so the perfornmance of classification needs to be improved;finally,in the case of large dimension of original features,how to quickly and effectively construct the feature structure with the consideration of the category information of the feature is also a key problem.In view of the above problems,this thesis has carried out the related research and experiments,the main contents are as follows(1)For extracting the better features to describe images,we have firstly carried out some experiments in this work.Different traditional features and convolutional neural network features of VAIS dataset have been extracted respectively.Combining with SVM classifier,the classification results of single feature have been obtained.After comparative analysis,two kinds of features with the higher expressiveness have been selected.(2)Using the structure fusion method based on locally preserved projection,we have carried out the feature structure fusion experiment on the data set.When constructing the internal structure of the feature,the square structure was respectively constructed by the chi-square distance or the Euclidean distance.Their experimental results were compared and analyzed.(3)Because the structural fusion method of locally preserved projection does not incorporate feature category information,we propose a method of combining structural fusion idea with supervised linear discriminant analysis,and carry out related experiments.(4)Aiming at the problem that the original feature dimension is large,the computation of the internal structure is complex,the storage capacity is huge,and the information of the feature category is still not enough for class distinguish ability,a new idea of constructing the internal structure of the feature is put forward in this paper.The method mainly includes tlhree aspects:one is the construction of the weight matrix of the same category feature,the other is the weight matrix construction of the different category features,and the third is the weight matrix construction after the feature weighting.This method not only guarantees a high degree of uniformity of intra-class structure,but also distinguishes the class information of inter-class structure.At the same time,it greatly reduces the computational complexity in the construction process.(5)For the inefficiency computation of the linear discriminant analysis algorithm,we combine the spectral regression discriminant analysis algorithm with the new structure fusion idea in this paper.Experiments on the VAIS dataset show that the proposed method can reduce the original 102400-dimensional features to 5 dimensions,the classification accuracy of visible images can reach 88.53%.and the infrared image can reach 74.86%at daytime,the paired image can reach 88.34%.The results show that the improved method not only reduces the feature dimension greatly,but also integrates the infommation of the feature class structure well,and achieves the better classification results than the structure fusion method or linear discriminant analysis.(6)A software for marine vessels classification has been developed on the MATLAB platfom..The model parameters of the structural fusion process can be set and trained in the interactive interface.A single vessel image can also be selected for classification.