Research On Few-shot Detection And Recognition In Optical Remote Sensing Image

As remote sensing technology becomes more and more developed,remote sensing image data is becoming more and more available.People use the technology of intelligent interpretation of remote sensing image to analyze the hidden information of the remote sensing image,which is widely applied in many fields such as environmental monito’ring,geological research,national defense and military and so on.Remote sensing image target detection and recognition transforms remote sensing images from image information to readable intellectualized information.At present,remote sensing image target detection recognition based on deep learning can locate and classify targets in remote sensing images by learning features of a large number of training samples.However,on the one handodue to the imaging mechanism of remote sensing images,it costs a lot to label remote sensing images;on the other hand,the number of samples available in some specific scenes is extremely scarce,which makes it impossible to provide a large number of labeled training samples for the remote sensing target detection and recognition algorithm based on deep learning,and then the target detection and recognition effectiveness decreases seriously.In addition,the study of remote sensing image target detection and recognition technology in few-shot scenarios can reduce the dependence of deep learning algorithms on massive labeling data,enable the model to acquire automatic labeling capability,and push the remote sensing image target detection and recognition technology to a higher level of"artificial intelligence".Based on this,the reseagrch on few-shot optical remote sensing image target detection and recognition technology is carried out.Researching on the few-shot target detection and recognition algorithm based on meta-learning and metric learning provides theoretical and methodological support for few-shot remote sensing target detection and recognition under the scarcity of labeled training samples.The main research contents of the paper are as follows.1.The remote sensing target recognition dataset RSD-FSC and the remote sensing target detection dataset RSDD are constructed by integrating data sources such as open source remote sensing image datasets on the Internet and Google Earth raw images.The RSDD contains remote sensing images of 10 types of remote sensing targets and the corresponding annotation files,and the RSD-FSC contains slices of 21 different types of remote sensing targets.They provide the basis for the research of few-shot target detection and recognition technology in remote sensing scenes.2.The mainstream few-shot image classification algorithms are studied and learned deeply.From the perspectivest of metric method and training mode,we propose a few-shot remote sensing target recognition method RS-DN4 based on metric learning,which alleviates the problem of small inter-class differences and large intra-class differences under few-shot conditions by metric learning,and improves the generalization ability of the model by using episodic training method,so as to realize few-shot remote sensing target recognition in image level.3.A comprehensive analysis and research on current few-shot target detection algorithms and deep learning-based target detection algorithms is conducted.Combined with RS-DN4,a few-shot target detection algorithm based on metric learning is proposed.It uses metric learning to guide the RPN network to generate high-quality region candidate frames in the few-shot scenarios,and designs a multi-task loss function to improve the performance of the model in few-shot scenarios.