Remote Sensing Image Object Detection Based On Improved YOLO Series Algorithm

Remote sensing image detection,as a frontier and hotspot in current target detection,has important significance and applications in vehicle detection,air patrol,military,navigation,salvage and so on.However,compared with ordinary natural optical images,remote sensing images generally vary greatly in target scale and are easily affected by complex environmental factors such as light and weather.In addition,difficult samples in remote sensing images have poor perception ability to local information,which puts forward higher detection requirements for the model.In order to improve the detection accuracy of remote sensing images in complex environments,this paper improves YOLOv3 and YOLOv5 respectively.The main work of this paper is as follows:Firstly,aiming at the problem that remote sensing images are easy to be affected by complex environment,which leads to low detection accuracy,an improved remote sensing image object detection algorithm RS-YOLOv3 is proposed.The first and last residual modules of the backbone network Darknet53 are replaced by Res Ne St Block modules to realize information interaction across feature map groups.In this paper,a Multi-dimensional information interaction Polarized Self-Attention(M-PSA)module is proposed to fully consider the importance of polarized channel interactions for detailed information.Embed it in the Res Ne St Block module;A new and efficient residual module is redesigned to alleviate the feature loss problem caused by the original residual module’s construction of short connections in low-dimensional feature space,and improve the computational efficiency.An improved serialization feature enhancement structure is introduced to realize feature enhancement.The detection accuracy and convergence speed of the target regression box are enhanced by replacing the CIo U loss function.Based on RSOD data set,self-built data set is added to make up for the problem of insufficient data in complex background.Second,aiming at the problem of detection difficulty caused by the weak information perception ability of difficult samples in remote sensing images,an improved remote sensing image detection algorithm YOLOv5-AQ is proposed.An Aggregate context information awareness module(Aggregate CIAM)is designed.Firstly,dynamic convolution is used to replace the traditional common convolution in this module to seek a balance between the performance of the model and the computational load.Then,an efficient Quadruplet Attention(QA)module is proposed to capture the intersecting dimensions through a new four-branch structure and enhance the dependence between dimensions.In order to avoid higher memory access cost and computing energy consumption caused by feature redundancy,the Aggregate CIAM structure was optimized by one-time aggregation.Finally,Transformer is introduced into YOLOv5 s backbone network to enhance the global information sensing ability of the model.The experimental results show that compared with other classical detection algorithms,RS-YOLOv3 has a better detection effect on remote sensing images in complex environments,and the detection speed is consistent with that of YOLOv3,which meets the requirements of accuracy and real-time performance.At the same time,YOLOv5-AQ is less affected by the environment,and the detection accuracy is generally higher.Generally speaking,difficult samples in complex environment have limited influence on the model.Figure [35] Table [8] Reference [63]...