Research On Convolutional Neural Network Based Object Detection For Remote Sensing Image

Remote sensing images have the characteristics of large image size,small ratio of target area to image area and large amount of data.The spatial information contained has high research value in major fields such as national security,resource monitoring,urban management and so on.With the continuous improvement of aerospace technology and image sensor technology,the collection of remote sensing images has gradually formed the characteristics of all-day,all-weather,and global scope.The rapid increase of data volume has put forward higher requirements for the corresponding image interpretation technology.Object detection is an important part of image interpretation.We aim to realize object detection for remote sensing image by using deep learning methods,and modify the algorithm structure to improve the accuracy of detectors.The main work of this paper is divided into three parts.First,Faster R-CNN is used to detect multiple classes of objects,thereby exploring the working principle and shortcomings of the two-stage structure detector.Aiming at the situation that there is oriented labeling in the remote sensing image dataset,we design R-GIoU based on rotated bounding boxes to replace IoU.By calculating the minimum circumscribed polygon,the relative position and angle difference between two rotated bounding boxes are more strictly limited.The experimental results show that using R-GIoU can improve the detection accuracy by 2%.Secondly,a feature fusion method based on dense connection and global information extraction is designed to improve the detection accuracy of plane objects.plane target detection on remote sensing image is a typical subject in this field.The proposed method consists of top-down and bottom-up feature fusion branches,which can effectively superimpose the target's high-level semantic information and lowlevel structure information to the output feature maps and increase the amount of information of the output feature maps.Dimension reduction process is added in the fusion structure to reduce the dimension of the output feature maps and reduce the computational complexity of the subsequent network.The experimental results show that our feature fusion method improves the detection accuracy while reducing the depth of feature extraction network,and achieves 94.20% on the DOTA plane dataset.Thirdly,a detection network based on NAS-FPN is constructed to improve the detection accuracy of dense and small targets.In this paper,the feature fusion structure NAS-FPN acquired by the neural architecture search method is introduced as the feature fusion module,and Faster R-CNN is used as the detector.According to the area ratio and the total target numbers in different class,five classes of targets are selected from the DOTA dataset and a dense small target dataset is constructed.NASFPN has good reusability,different detection networks can be constructed by controlling the number of multiplexing and output feature dimensions.We achieve 10% higher AP than by increasing the number of multiplexing and output feature dimensions,the best result with ResNet-50 d is 74% which is 19% higher than Faster R-CNN with ResNet-101,13% higher than FPN.