Research On Multi-target Detection Technology Based On Aerial Image

Aerial image has very important application value and broad application prospect in many fields such as resource exploration,disaster disposal and enemy situation reconnaissance,and the application of aerial image depends on the multi-target detection technology of aerial image in many cases.With the rapid development of deep learning technology,multi-target detection technology has made great progress in many fields.However,it is difficult to detect multi-target in aerial images due to the complex background,large noise and large span of target size of aerial images.In this paper,the application of deep learning correlation algorithm in aerial image multi-target detection is studied on the basis of existing algorithms,aiming at the problem of multi-target detection in aerial image and combining with many difficulties in aerial image multi-target detection.At the same time,this paper uses the rotating bounding boxes to detect multiple targets in aerial images.Compared with the horizontal bounding boxes,using the rotating bounding boxes to detect targets can more accurately fit the target position and avoid noise interference.Based on the rotating bounding boxes,this paper introduces several improvement measures on the basis of related directional target detection algorithm,which significantly improves the algorithm performance,the main work of this paper includes the following aspects:Firstly,a multi-target detection algorithm based on multi-scale feature fusion for aerial image is proposed which is DRBox(Detector using Rotatable Bounding Box).On the basis of the directional target detection algorithm DRBox,this paper replaces the feature extraction network with the Res Net101 with stronger feature extraction ability,and uses the multi-scale feature fusion strategy of FSSD(Feature Fusion Single Shot Multibox Detector)to effectively fuse the extracted multi-layer features,so that the features after fusion have the characteristics of both high positioning accuracy of low-level features and strong semantics of high-level features,thus effectively solving the problem of low detection accuracy caused by small target proportion of aerial image.At the same time,aiming at the characteristics of dense target distribution and overlapping labeling frame in aerial images,the idea of anchor frame offset is introduced in the prior anchor frame setting strategy,which further improves the accuracy of the algorithm in detecting small and medium targets.The DRBox algorithm based on multiscale feature fusion has the characteristics of fast target detection algorithm based on regression.It has been verified by experiments that the algorithm has a certain detection accuracy and can obtain the real-time application requirements.Secondly,based on RRPN(Rotation Region Proposal Networks),a multi-scale RRPN aerial image multi-target detection algorithm based on attention is constructed.In this paper,the feature extraction network is replaced as Res Net101,and the feature fusion strategy is introduced to enhance the feature extraction from the network.In order to provide more accurate features for regional recommendation networks,spatial and characteristic channel attention networks are introduced to screen abnormal activation points in the features.At the same time,the position aware RRo IAlign is used for the extraction of RRo I(Rotation Region of Interest)features,which effectively solves the phenomenon of coordinate deviation in the process of feature extraction,and introduces the idea of position perception to further improve the accuracy of features.It has been verified by experiments that the multi-scale RRPN algorithm based on attention has better detection accuracy,and the relevant improvement measures in this paper play a key role in improving the detection accuracy of the algorithm.Finally,the two improved algorithms in this paper are compared with other algorithms,and the differences of detection accuracy and detection speed between the two improved algorithms are compared.The DRBox based on feature fusion has obvious advantages in detection speed,but the detection accuracy still has a large room for improvement,while the multi-scale RRPN algorithm based on attention has a high detection accuracy.At the same time,the two improved algorithms in this paper is tested across data sets and tested on self-collected data sets,which proves that the improved algorithm in this paper has good generalization performance.