Fine-grained Classification For Aircrafts In Optical Remote Sensing Images

With the rapid development of remote sensing technologies,the collection of high-resolution remote sensing images provides us with a large amount of remote sensing in-formation,which contains many aircraft targets.The aircraft targets are roughly similar in structure and shape,but they vary widely in local parts.It is difficult to distinguish differ-ent types of aircraft targets for humans as well as artificial intelligence models.Therefore,the fine-grained classification for aircraft targets in remote sensing images plays an impor-tant role in airport management,route monitoring,and aircraft target screening.There-fore,this paper aims to augment the fine-grained dataset,analyze the model,propose a keypoint prediction technique,and build a network for fine-grained aircraft classificaiton.Our methods improve the recognition efficiency of fine-grained classification of aircraft targets in remote sensing images.Firstly,this paper proposes a data augmentation method based on Interest Points of Feature for the problem of small amount of data and high cost of labeling in fine-grained datasets.By looking for local areas that are influencing the classifier,the method augu-ment the data for the fine-grained dataset and alleviate the over-fitting problem.To some extent,the classification accuracy also has been improved on many fine-grained datasets.Specially,we earn a performance boost of 8.47%on the fine-grained aircraft dataset.For the categories which model confuse,we use the confusion graph,Grad-CAM,and Part-Grad-CAM methods to analyze the causes of confusion from coarse to fine at the category level,target individual level,and local part level.The most important for fine-grained classification is the feature of local part.This paper proposes a key point prediction model based on deep convolutional neural network to predict key points,which provides a basis for the acquisition of key parts.In the key point prediction,this paper uses transfer learning and designs a specific classifier and loss function for the key point prediction network.Experiments show that the key point prediction model based on deep convolutional neural network can predict the key points precisely.Due to the lack of intermediate semantic features in traditional convolutional neural network,this paper proposes a fine-grained classification algorithm based on part-based fully connected layer and concatenation fully connected layer.This method adds part-based fully connected layers and a concatenation fully-connected layer in the traditional convolutional neural network to link low-level image features and advanced object in-formation,and fills the intermediate semantic features by learning specific part features,and adjusts the importance of each part according to the classifier.This paper also com-bines the key point prediction method with this method to form a two-stage process for the fine-grained classification of aircraft targets.Experiments show that this method im-proves the performance of aircraft 's classification by nearly 13 percentage compared with the traditional classification model.