Remote Sensing Image Object Detection And Recognition Based On Object Proposal And Deep Network

Remote sensing image target detection and recognition is a very important branch of remote sensing interpreting technology and it is a technology to obtain specific target areas and categories in remote sensing images.At present,the target detection and recognition algorithms for remote sensing images have the problems of poor generalization performance and low efficiency and accuracy of the algorithms.Therefore,the current remote sensing image object detection and recognition algorithms need further development.Currently,the object proposal methods focus on the most significant and unique locations in the image,greatly reducing the amount of computation for subsequent classification,detection,and other tasks.At the same time,as a deep learning algorithm that has similarities to the multi-layered physical structure of the human learning system,it has achieved the latest results in many problems involving large amounts of data such as classification and reasoning.The research work of this paper is based on the above two points,introducing the proposed algorithm and deep network structure of the target region into the object detection and recognition of remote sensing images.The specific work is as follows:1.A SAR image object recognition algorithm based on deformable convolutional neural network is proposed.A deformable convolution unit is introduced,which can effectively change the position of sampling points and learn the bias of sampling points to enhance the ability of network model transformation.In addition,the network uses the global mean pooling layer instead of the full convolutional layer in CNN,so that the original large number of parameters can be omitted,which can greatly reduce the network size and avoid overfitting.In addition,this chapter proposes a fast detection algorithm based on peak characteristics,which can achieve SAR target detection through the peak feature extraction module and connected area labeling module.2.A remote sensing image object detection algorithm based on YOLO deep network is proposed.The object detection is used as the regression problem solving.The whole map information is used for forecasting during the training and testing process.The generalized representation information of the target learned has a certain universality,which can effectively improve the detection accuracy and recall rate and other indicators.At the same time,it has the advantages of simple network prediction process and fast detection speed.In addition,this chapter proposes an improved inter-frame difference method for the detection of moving targets in high-resolution remote sensing images.It has the advantages of simple implementation,low design complexity and good stability,and it can be used to detect moving targets in high-resolution remote sensing images and has a very good effect.3.A fast object detection algorithm for remote sensing image is proposed.Based on the structured forest edge detection and object proposal algorithm,the edge set is calculated by the image itself,and then the final target detection result is determined by the edge set similarity and candidate frame scoring rules.