Research On Open Set And Long-tailed Remote Sensing Scene Classification

In recent years,with the rapid development of earth observation technology,remote sensing has entered an unprecedented new stage,and the demand for intelligent earth observation and analysis has become more and more important.Remote sensing scene classification is the basis of many remote sensing applications.With the development of deep learning technology and the emergence of large-scale data sets for remote sensing scene image classification,methods based on deep learning technology that can automatically extract features have achieved great success in remote sensing scene classification tasks.However,the current deep learning technology is often driven by data,and the performance of its model largely depends on the number and quality of samples in the data set.Therefore,most scholars use large-scale,balanced benchmark data sets to conduct research,but the models obtained under such conditions are difficult to put into practical application.This is because remote sensing scene classification data sets under realistic conditions often show the characteristics of open long-tailed distribution,and the performance of deep learning methods based on large-scale balanced data sets will be greatly reduced under this condition.In addition,there are few remote sensing scene classification tasks currently carried out around the distribution of such data.In order to make up for the gaps in this research field,this paper aims to carry out related research on the three major difficulties of integrated solution to the long-tailed distribution,i.e.,class imbalance classification,few-shot learning and open set recognition.First of all,inspired by the meta-metric learning method,we propose a random fine-tuned meta-metric learning model(RF-MML)to solve the problem of class imbalance and few-shot learning.In RF-MML,a novel strategy is proposed to train the model.The strategy includes two stages,namely random episode training and all classes fine-tuning which can successfully improving the performance of meta-learning paradigm on highly imbalanced data by introducing randomness and integrating with all classes fine-tuning.Through experiments,this paper proves the effectiveness of the model on unbalanced data sets and its superiority compared to other stateof-the-art methods.On the basis of RF-MML,this paper also proposes the Attention Prototype Network(APN)to solve the three major difficulties of open long-tail distributed data in an integrated manner.The APN model adds a self-attention mechanism after the embedded features of the traditional CNN,so that the model can extract relevant features across distances,laying a foundation for the feature measurement of subsequent prototypes.In addition,APN also added a prototype measurement module based on deep metric learning.This module maps all samples of all categories into an average prototype in the feature space.And then utilize the distance between the tested sample and the prototype of each category to help classification.This model can effectively alleviate the problem that the deep learning model’s recognition accuracy is severely reduced in the categories with insufficient samples and can better distinguish between known categories and unknown categories.Experiments on the open long-tailed distribution data set prove the superiority of the proposed APN model as compared with other state-of-the-art methods.