Target Detection From High-resolution Remote Sensing Image Based On Regional Convolution Neural Network

In recent years,with the successful implementation of "high-resolution special project" in China,the number of high-resolution satellites has gradually increased,the spatial resolution of optical remote sensing images has rapidly promoted,and the amount of information has also soared.The improvement of image resolution and the increase of image information has provided more refined geospatial information,but also puts forward higher requirements for various remote sensing image processing technologies.High-resolution remote sensing image target detection task has become one of the important research fields in remote sensing image processing because it plays an important role in the construction of smart city.The traditional high-resolution remote sensing image target detection methods have poor adaptability to the target and rely too much on human assistance.Therefore,the research of efficient target detection algorithm is of great significance to the information retrieval of high-resolution remote sensing image.This paper is based on the Faster R-CNN,one of the regional convolution neural network algorithms,to study the target detection of high-resolution remote sensing images.The main research contents are as follows:(1)Building detection datasets of high-resolution remote sensing image is constructed.In view of the fact that there are few public data sets for building detection,this paper labels the remote sensing image with a resolution of 1 meter in the urban area of Chongqing in the form of manual labeling,and establishes a high-resolution remote sensing image building detection data set.(2)Pointing at the particularity of high-resolution remote sensing image,a remote sensing image detection algorithm based on residual network and clustering idea is proposed.By using resnet50 as the feature extraction network,the depth of the network is deepened and the ability of feature extraction is improved;K-means++ algorithm is used to cluster the dimensions of the real annotation boxes in the data set,and the appropriate preset candidate box super parameters are obtained to enhance the inclusiveness of the algorithm to the target size;ROI align is used to replace ROI Pooling algorithm to eliminate the error caused by the quantization operation on the feature of the candidate region;In addition,in order to reduce the missed detection rate of overlapping targets,soft NMS algorithm is used to replace NMS algorithm for post-processing of proposal box.The detection accuracy of the optimized algorithm is improved in UCAS-AOD aircraft data set and self-built building data set.(3)In order to further improve the detection accuracy of complex background and dense targets,this paper proposes a remote sensing image target detection algorithm based on dilated convolution and feature fusion.Through the introduction of dilated convolution,a new feature extraction network DResNet50 is proposed.At the same time,based on the idea of feature fusion,the feature extraction network is multi-layer feature fusion,making full use of the shallow location information and deep semantic information to improve the detection performance of the algorithm.Through the model comparison experiment and ablation experiment,the effectiveness of the optimization algorithm for remote sensing image detection of complex background and dense targets is verified.(4)Based on the user requirement analysis,the target detection platform of remote sensing image target detection is designed and realized.Combined with Python programming language,QT creator development platform and a series of interface development tools,the development of visual target detection platform is completed.The platform provides the display of detection results and information browsing.By testing the experimental data on the interface,the usability of the interface is verified.