Research On Few Shot Classification Algorithm Of Remote Sensing Images Based On Graph Network

The most important thing in remote sensing image interpretation is the recognition of remote sensing image information,which has a large number of applications in the military and civilian fields.The deep learning-based model will rely more on the amount of training sample data,which is in contradiction with its own status quo that it is difficult to label remote sensing image data.Using a small number or even a single sample to make the algorithm have the ability to learn new categories has practical significance in the field of remote sensing image target recognition and analysis.Based on this,this paper applies the few-sample recognition technology to the field of remote sensing image classification,and proposes a few-sample remote sensing image recognition algorithm based on Graph Neural Network（GNN）.The main work and innovations are as follows:（1）Based on the instability of the MAML model in the small-sample recognition task and its inherent limitations on the internal representation of the data and the RN model for non-linear feature fusion of samples of the same category Some information is lost due to the linear superposition method.In this paper,GNN is selected as the prototype model to carry out the research on the small-sample recognition of remote sensing images,A comparative experiment was carried out on the small-sample data set Omniglot,which increased the recognition rate by 0.5%compared to RN.（2）The main thinking in the problem of small sample recognition is to adopt a suitable method to solve the phenomenon of lack of category annotation data.The traditional feedforward network（CNN,RNN）only obtains the information of the sample image itself for comparison and prediction.This article considers the use of GNN’s unique node and edge structure to learn the information expression between samples.Based on this,the GNN is improved to propose the Edge Labeled Graph Network（EGNN）,which expresses the similarity and difference information between nodes through the two-dimensional vector of the edge features of the graph structure,so as to obtain the category prediction results of the nodes.Compared with GNN on the small-sample data set miniImageNet,the recognition accuracy is increased by 0.43%.（3）In the small-sample classification problem,the different distributions of the training set and test set categories will cause the model’s generalization ability to be weak.This paper adopts FMix-based data enhancement technology,and uses Fourier transform on the basis of CutMix method to crop and insert the high-frequency and low-frequency regions,so that the crop template is no longer limited to the rectangular area.Compared with the linear interpolation method,it can effectively retain the information of the semantic distribution in the image.A comparison experiment of sample augmentation performance improvement capabilities was carried out on the small-sample data set miniImageNet,and the improved model FMix-EGNN improved the recognition accuracy of 2.11%compared with EGNN.（4）In the feature extraction stage,the idea of interval enhancement is introduced for"magnification" processing,and gravity and thrust calculations are used to change the Euclidean distance between feature vectors without changing the network infrastructure,and an improved model AugFMix-EGNN is obtained.