| Optical remote sensing image object detection tasks are currently widely used in various aspects of military reconnaissance,disaster monitoring,and urban planning.It is the prerequisite and basis for further tasks such as remote sensing image analysis.With the continuous development of sensor technology in recent years,the spatial resolution of remote sensing images is getting higher and higher,and more and more detailed information is provided for algorithm processing.Therefore,the task of object detection for your images has become more and more important.In recent years,deep learning methods have been widely used in natural image classification and detection tasks.Researchers have also begun to use deep learning methods to solve remote sensing image object detection tasks.Optical remote sensing images have the characteristics of large object scale changes,uncertain directions,dense object distribution,and extremely complex backgrounds.Therefore,direct application of object detection methods in natural images to remote sensing images cannot achieve good detection results.This article will mainly focus on how to solve the problem of small object detection in remote sensing images.The main research contents include:1.Aiming at the problem that the complex background of remote sensing images causes great interference to small object detection,this paper designs a remote sensing image small object detection network based on context features.One is to expand the candidate frame area to extract the features of the context area,and the other is to use the gating design in LSTM to enable the network to adaptively learn and control the features of the context to improve the performance of the classification results of the candidate area.The features of the small object are here.The process is strengthened to improve the detection performance of small objects in remote sensing images.The effectiveness of the proposed contextual feature fusion network is verified by adding the proposed structure to the RRPN network and the Ro I Transformer network and comparing experiments on the DOTA dataset.2.Aiming at the problem that small objects in remote sensing images often appear in clusters,which makes their detection difficult,this paper designs a remote sensing image small object detection network based on a multi-level attention mechanism.Use Res Net50 as an example to make a feature extraction network,use a dual-channel attention module based on pooling at the low level,and use a globally associated attention module at the high level to construct a multilevel attention network,ensuring that the network calculation is not greatly increased At the same time,it pays more attention to the characteristics of small object monomers in the cluster structure,so as to achieve the goal of improving detection performance.The same network as the previous section is designed for ablation experiments to verify the effectiveness of the multilevel attention network module.3.In view of the huge size of remote sensing images,the usual fixed step cutting method does not solve the problem of large changes in the internal size of the object,but also produces a large number of repeated objects.This article uses multi-scale cutting and improves the training strategy to improve the cutting While generating the number of sub-images,it also uses redundant background information and ensures the input training of multi-scale images,thereby greatly improving the performance of small object detection.In addition,this paper also proposes to use the prior information of different types of objects in the remote sensing image scene to modify the probability output of the detection network to improve the detection performance.Using the overhead angle characteristics of remote sensing images,objects of the same level will not overlap each other,and a post-processing method of improved NMS algorithm is proposed to improve the detection performance without changing the network structure and the amount of calculation.In order to verify the versatility of the pre-processing and post-processing methods proposed in this section,this paper uses different network structures to verify the three remote sensing data sets of DOTA,HRRSD,and NWPU VHR-10.Experiments prove the effectiveness of the proposed algorithm. |
