Object Detection In Remote Sensing Images Based On Improved Faster R-CNN

Object detection in remote sensing images is a branch of traditional object detection.It is one of the core problems in remote sensing images processing.As the basis of remote sensing images segmentation,scene classification,remote sensing information automatic extraction and other tasks,it has been widely used in military and civilian fields.In recent years,with the development of carrying platform and sensor technology,the spatial resolution of remote sensing images is increasing,and the visual difference between remote sensing image and natural image is decreasing.More and more computer vision methods can be applied to object detection in remote sensing images,but the problems of complex background,small target and target rotation in remote sensing images still lead to low detection accuracy and efficiency.In order to solve these problems,this paper proposes an improved Faster R-CNN detection model which fuses attention mechanism,and verifies the effectiveness of this method through experiments.The specific work of this paper is as follows:(1)Preprocessing the public NWPU_VHR-10’ data set:firstly,the data is expanded,and the number of samples is increased by flipping,panning,scaling deformation.Secondly,the low-quality images in the data set are reconstructed by using the Enhanced Depth Super-Resolution(EDSR)network.Finally,the data set of 3000 images is divided into training set,verification set and test set according to 3:1:1.(2)Construct a feature extraction network with attention mechanism:in the selected feature extraction network ResNet-101,attention mechanism module is introduced,which aims to obtain more detection target information and suppress useless background information during image feature extraction,so as to solve the problem of complex background and small target caused by large field of vision of remote sensing images.(3)The original Faster R-CNN detection model is improved:replace the feature extraction network;the concept of cross-correlation in signal analysis is introduced,and the cross-correlation suppression strategy is designed in this paper.The core idea is to use the cross-correlation between the target distribution to screen the candidate boxes to reduce the false alarm rate;Softer Non-Maximum Suppression is used to replace the traditional Non-Maximum Suppression in order to adapt to the problem of target rotation in remote sensing images and improve the performance of the detector.In order to prove the effectiveness of this method,two groups of comparative experiments are carried out:the first group is the Ablation Experiment of EDSR module,CBAM module,cross-correlation suppression module and Softer Non-Maximum Suppression module used in this paper.The results show that each module can improve the average detection accuracy to a certain extent,which proves the effectiveness of each module proposed in this paper;the second group is the comparison and analysis of this method and other existing methods on NWPU_VHR-10 data set,which proves that this model has batter adaptability to the remote sensing images scene with complex background and object rotation than other methods.