Deep Learning Based Object Detection And Recognition Methods On High Resolution Optical Remote Sensing Image

The role of space remote sensing in national economic development,national defense modernization and sustainable development of human society is becoming more and more important.The application-oriented concept of space-ground integration is more and more fully embodied in the design and manufacture of remote sensing satellite systems.While remote sensing is striving for higher spectral resolution,spatial optical resolution and temporal resolution,the application of more targeted remote sensing small satellites has also been rapidly developed and the human's ability of acquiring remote sensing image data acquired is increasing.Although remote sensing image data can be obtained quickly and easily,the ability of human to process large amounts of data is far from being able to keep up with the ability to acquire data,which makes manual analysis and interpretation more and more difficult.In recent years,the deep learning technology based on big data has broken the limitations of traditional methods in the field of computer vision and made breakthroughs,which provides a new technology method to solve the problems faced by the practical application of optical remote sensing image detection and recognition.Therefore,this paper considers convolutional neural network as a new technology approach and faces the challenging application background of remote sensing image target detection and recognition and studies the high-resolution optical remote sensing image target detection and recognition technology based on deep learning.The main work of this paper is as follows:1.Basic theory research on remote sensing image target detection and recognition based on deep learning.Deep learning theory is the premise of researching remote sensing image target detection and recognition technology,so this paper introduces the basic principle and technical details of convolutional neural network,mainly including network model,network structure and network training and summarizes the method of optimizing and evaluating the performance of deep learning network.On this basis,the application characteristics and depth learning adaptability of remote sensing image are analyzed,which lays a foundation for the research of remote sensing image target detection and recognition based on convolution neural network.2.Deep learning network model analysis for target detection and recognition.Aiming at the problem of object detection and recognition in optical remote sensing image,applying deep learning to the object detection and recognition of optical remote sensing image.It is necessary to study the deep convolutional neural network for target detection and recognition.In this paper,two typical convolutional neural network models are studied in detail,namely,regional extraction and regression-based deep learning networks and experimental verification and in-depth analysis of these two types of deep learning models are carried out through data sets,which provides the basis for learning optical remote sensing image target detection and recognition methods.3.Remote sensing image target detection based on transfer learning convolutional neural network.Aiming at the complex background of large-format optical remote sensing images and the relatively small target size and the lack of data for labeled training samples,this paper proposes a remote sensing image target detection and recognition method based on transfer learning deep convolutional neural network: firstly,the improved deep learning network is pre-trained under the large data set of natural images,and the pre-trained model is transferred to the remote sensing target domain to solve the problem of lacking remote sensing image training samples.Then,based on the transfer learning deep convolutional neural network for target location and identification,aiming at solving the problem of accuracy and speed of remote sensing image target detection.Finally,the large-scale optical remote sensing image is used to verify the effectiveness of the proposed algorithm.4.Target pose extraction of optical remote sensing image based on convolutional neural network.Aiming at the problems existing in the traditional pose estimation of optical remote sensing image target,this paper integrates the target pose information extraction process into the target detection and recognition process based on deep learning and studies the target pose extraction method based on deep convolutional neural network.The proposed remote sensing image target pose extraction network Di-CNN can make full use of the multi-layer features of the convolutional neural network and return the target position and attitude information in an end-to-end manner,which further improves the accuracy of target detection.The effectiveness of the proposed method is verified by experiments on remote sensing image datasets.