Research On Object Detection Technologies For Remote Sensing Images

As the developing of the remote sensing imaging technologies,the acquisition of remote sensing images is more convenient.As a research hotspot in the field of remote sensing images analysis,object detection aims at quickly locating the specific object from the large-scale remote sensing images,which is widely used in traffic management,city planning,and military monitoring.The abundance of information in remote sensing images is both an opportunity and a challenge,which not only enriches the source for object detection,but also increases the difficulty of object feature extraction.Taking object detection in remote sensing images as task,this paper deeply studies the object detection technologies in remote sensing images on the basis of fully summarizing and analyzing the existing problems,basic principles as well as related works.In this paper,a region-enhanced convolutional neural network is proposed for the object detection of remote sensing images.First,to settle the restraint of cluttered background to the feature extraction,a region-enhanced two-branch convolutional neural network is introduced to enhance the feature of object regions and weaken the interference of background.Second,the saliency information is adopted with the pixel-level loss function to optimize the extraction of saliency information and enhance the object regions in feature maps.Finally,a multi-layer fusion strategy is utilized in the proposed method for exploiting the contextual information,which enriches the information in feature maps and then improves the object recognize ability of the proposed network.Experimental results show that the proposed method is capable of enhancing the object feature and suppressing the interference of background in complex scene,thus obtain satisfied detection accuracy.A self-attention-based object detection method for remote sensing images is also implemented in this paper.First,considering that the self-attention mechanism is able to increase the weight of the object regions and enhance the response of object,a simple but effective self-attention module is designed to enhance the object regions in feature maps.Second,according to the information difference between multi-scale feature maps,a hierarchical self-attention information extraction network isconstructed,and then the architecture of multi-scale feature maps is implemented with the dilated convolutional technology to build an efficient object detection network.Experimental results show that the implemented method is able to effectively strengthen the feature of object regions and improve the detection accuracy.