Few-shot Ship Target Recognition Based On Graph Convolution Network In Remote Sensing Image

As a hotspot in the field of remote sensing image processing and pattern recognition,remote sensing ship image target recognition has important military value.However,it is difficult to obtain remote sensing ship images and the sample capacity is relatively little,which easily leads to the problem of few-shot.In recent years,many studies related to few-shot learning have been developed in the field of deep learning,including transfer learning,attribute learning,metric learning,and meta-learning and so on.On the recognition of few-shot ship of remote sensing image,the abovementioned few-shot learning has the disadvantages that the range of relevant data is limited and the migration effect is difficult to guarantee.Therefore,aiming at the above problems and combined with the structural characteristics of ship targets,the contents of this study are as follows:(1)Research scheme: Different from the end-to-end structure of mapping from the bottom layer(pixel)to the high layer(category)directly in the conventional method,the middle layer is embedded between the bottom layer(pixel)and the high layer(category).By introducing the middle layer structure elements,the available data range and its characteristic information are defined.Furthermore,the traditional convolution network(CNN)is used to map the bottom layer to the middle layer,and the graph convolution network(GCN)is used to map the middle layer to the top layer.(2)Extraction of structural elements: By analyzing the structural characteristics of the ship,considering the generality,differentiation and significance,the middle layer structural elements are defined: the outline of the ship and the upper components of the ship(ship Island,apron).In addition,the spatial positioning information of each structural element in the ship target area is extracted.The specific contents include:using Mask R-CNN to extract the feature of ship contour,and using Yolo-v5 to extract the features of the upper components of the ship.(3)Few-shot ship target recognition based on GCN: Under the technical framework of GCN,two models of ship structure diagram construction are proposed based on the various structural elements of the middle layer: the structure diagram construction based on the original spatial information and the structure diagram construction based on the high-dimensional spatial information.This enables traditional ship image data types to be applied to graph convolutional neural networks that process non-Euclidean data types,and to improve the anti-interference ability of the graph construction stage by extracting features from different dimensions of nodes.By designing graph convolution operator and reasonable graph network hierarchy,the spatial distribution characteristics of structural elements are effectively extracted,and the accuracy of ship target recognition is increased to 91.21%.Finally,according to this study results,a remote sensing ship target recognition platform based on graph convolution network is designed.Through the combination of few-shot learning technology and graph convolution network technology,the limitation of existing fewshot target recognition method framework is broken.