Research On Remote Sensing Image Information Extraction Method Based On Deep Learning

Remote sensing technology is a modern detection technology emerging in the 1960 s.Over the years,it has provided stable and detailed data for land use status analysis,agricultural pest monitoring,urban planning,and other civil fields.It has more important significance for military fields like military target detection,battlefield environment simulation.After years of development,remote sensing technology has made great progress in improving spatial resolution,temporal resolution,spectral resolution,and other key technical links.Remote sensing technology has acquired the capability of high spectral,high spatial resolution and all-weather earth observation.With the continuous improvement of the three resolutions,the remote sensing data has witnessed explosive growth.How to quickly and effectively extract the information in the remote sensing data,reduce the information redundancy and improve the information utilization rate has become an urgent problem to be solved.Especially in recent years,with the rapid development of high-resolution remote sensing satellites at home and abroad,the amount of high-resolution remote sensing image data is increasing at the terabyte level every day.The automatic information extraction of highresolution remote sensing images has become a research hotspot concerned by the market and researchers.As the world’s ten breakthrough technology in 2013,deep learning made a breakthrough in the field of multiple,especially image classification,target detection,semantic segmentation,video analysis,speech recognition,machine translation,automatic driving under the support of computer hardware and machine learning theory.Deep learning is essentially a process of fitting complex high-dimensional functions through artificial neural networks.Through experimental verification,the deep artificial neural network has a strong ability to abstract features of images or sequence data.In addition,the fast computing ability of the new Graphics Processing Unit provides the necessary support for the development of deep learning algorithms.The fully trained deep learning model can quickly execute the functions defined by the model,and its fast execution speed,strong generalization ability,and high stability have become incomparable advantages over other methods.In recent years,deep learning technology has been widely used to extract information from high-resolution remote sensing images,and gratifying achievements have been made in remote sensing scene classification,remote sensing target detection,remote sensing image description,remote sensing image segmentation,and other fields.However,there are still some problems.In this paper,remote sensing scene classification,remote sensing target detection,and remote sensing image description are explored and attempted.The research work and contributions of this paper mainly include the following aspects:In the research of remote sensing scene classification,a remote sensing scene classification method based on integrated convolutional neural network is proposed.When the deep convolutional neural network is used to classify the optical remote sensing scene images,a large amount of running time is needed to obtain a high classification accuracy.Reducing the number of layers of the network can improve the classification speed,but reduce the accuracy at the same time.In order to integrate the advantages of the high speed of the shallow network and high accuracy of the deep network and improve the efficiency of remote sensing image scene classification,a remote sensing image scene classification algorithm based on integrated convolutional neural network is proposed.First,the back-propagation network was constructed to measure the complexity of scene images.Then,according to the complexity level of images,the convolutional neural network suitable for the complexity was selected adaptively.Finally,the convolutional neural network is used to classify images and complete the classification process of remote sensing image scenes.The proposed algorithm can improve the classification accuracy and speed.In the study of remote sensing target detection,a fast detection method for multiclass rotating targets is proposed.For remote sensing image target detection,many existing algorithms mark the object with a classic box whose four sides are parallel to the axes.There is no problem when marking an airplane,oil tank,or other square targets with a classic box,but there is a lot of backgrounds reduce the recall rate when marking ships,ports,or other rectangular targets with a classic box.Some researchers have proposed a target detection method that can mark arbitrary rotating ships.However,this method has only one target category and the detection speed is slow.Therefore,this paper designs a detection model that can simultaneously detect 15 kinds of targets and mark arbitrary rotating targets.In terms of remote sensing image caption,there are few related pieces of research.This paper designs a remote sensing image caption method combining attention mechanism,Convolutional Neural Network and Recurrent Neural Network based on the natural image caption algorithm.Feature extraction part imports feature fusion technology.In addition,the attention mechanism enables the model to automatically focus on the image regions related to the generated words when using the Recurrent Neural Network to generate description statements.This method has made progress in the evaluation scores of several evaluation indexes.