Research On Fine-grained Classification Based On Aircraft Target In Remote Sensing Image

With the development of remote sensing technology and the increase of ways to obtain remote sensing data,remote sensing data show the trend of quantification and diversification.Remote sensing image classification is an important research topic in the field of remote sensing.The research of aircraft targets in remote sensing images is widely used in military reconnaissance and other fields.Aircraft data sets tend to have higher intra-class variations and smaller inter-class differences,which puts forward higher requirements for current classification algorithms and requires more fine-grained classification accuracy.Therefore,this article focuses on fine-grained classification based on aircraft target in remote sensing images,from the perspectives of image preprocessing,high-dimensional features of images,and obfuscation of the community,through the analysis of results from different angles,solutions from different angles are proposed,and multiple solutions are combined to improve the classification accuracy of remote sensing image aircraft targets.First of all,this article aims at the characteristics of the small proportion of the target objects and the relatively large noise in the remote sensing image data set,puts forward the data preprocessing method based on the background segmentation before the image feature extraction,using semantic segmentation technology to extract the feature information more efficiently.In different feature extraction methods,the pre-background segmentation data preprocessing is proved to be effective.Among them,on the aircraft fine-grained data set,in the deep network classification experiment,the classification accuracy was improved by 3.9% compared with that without preprocessing.Secondly,this paper focuses on that the high-dimensional features extracted by the deep convolutional neural network show clustering,but the range of each cluster is uncertain,and the cluster center is not obvious enough,and there is a range that coincides with the clustering areas of other categories.We propose a method of feature metrics to restrict the learnt features in the feature space.To adjust the learning process so that the features belonging to the same category are as close as possible in the feature space,and the features belonging to different categories are as far away as possible from the feature space,and finally the purpose of improving the fine-grained classification accuracy of the aircraft target is achieved.In the aircraft data set,the classification accuracy was improved by 2.4% compared with that without feature constraint method.Finally,this paper uses the visual features of the confusion in the image to construct the confusion tree of the aircraft dataset and the confusion community with a strong degree of confusion.Then,based on the causes of visual confusion in different confusion communities,the method of restricting the number of engine features and the method of constraining angular features are proposed to ensure that the overall classification accuracy of the community is improved or unchanged,while integrating different methods to improve the classification accuracy on the entire data set.Finally,in the overall aircraft data set,the classification accuracy improved 5.5 %.