Remote Sensing Image Recognition Of Small Samples Based On Deep Learning

With the progress of science and technology,remote sensing technology is in a highly developed stage,and a large amount of remote sensing image data is generated in various industries.The demand for data analysis and utilization of remote sensing images is also increasing.How to efficiently identify and classify remote sensing image targets has become one of the research hotspots in the field of remote sensing images.However,the implementation of deep neural models is based on a large amount of data,and some remote sensing image datasets are different from conventional image datasets.Their inherent characteristics lead to insufficient effective annotation data due to the following factors: for example,objective conditions are limited,and the image is affected by environmental factors such as spatial angle changes and spectral factor proportion transformations;For example,sensitive content and captured images of military targets;For example,due to the particularity of images,the cost of processing a large amount of effective data is too high;The above reasons limit the deep neural network model to insufficient training data,resulting in a significant decrease in the model’s generalization ability to the dataset.As the main research direction of small sample learning,small sample image classification and recognition aims to explore effective methods for small sample remote sensing image classification tasks under the condition of a small amount of data,and complete the recognition task of remote sensing images.In response to the above problems,the research content of this article is as follows:(1)Propose a classification model based on multi-scale features and attention.In response to the recognition problem of remote sensing images containing a large amount of information,considering that the network structure of existing small sample recognition models is too single to perform multi-level feature extraction on images,this chapter is based on the Relationship Net(RN)model and uses a new network structure to replace some layers of the feature embedding module to obtain feature information of different dimensions of the image,A hybrid attention mechanism(MECA)is designed between the embedded module and the relational module,which aims to obtain the correlation between channels through attention,extract rich Semantic information,refine the cross channel input information more carefully,make the network model focus on a finer scale,and eliminate the impact of vector module length weight on accuracy.Under small sample conditions,the model can effectively mine information from remote sensing images containing rich information,thereby improving the classification and recognition accuracy of the model in specific scene tasks.Three datasets were used in the experiment,and the experimental results proved the effectiveness of the model.(2)A small sample data augmentation algorithm based on generative adversarial networks is proposed.In order to solve the problem of overfitting of the model due to the scarcity of data in the same class limited by small samples,this chapter expands the data set quantitatively based on the generation of the countermeasure network model,taking the generation of the countermeasure network model as the research object at the data level.In this algorithm,residual structure and sparse branches are introduced into the generator,and the generator and discriminator are combined in multiple stages,Accelerate the effective combination and transmission of low resolution and high resolution features,and add a new loss function to ensure the consistency of image features and semantics,improve the stability of the generator and model,limit and constrain the generator to gradually generate more realistic training examples,adjust the average distribution of various types in the mapping space,and enhance the identification of support samples and query samples,Further improve the accuracy of the model in small sample remote sensing image classification and recognition tasks.