Research On Remote Sensing Target Detection Technology Based On YOLOv5

In recent years,due to the development of satellite and UAV technology and remote sensing technology,the spatial resolution of remote sensing images has even reached the sub-metre level,and the detailed information depicted in the images has become more and more detailed and complex,which makes the demand for high-precision multi-target detection based on remote sensing images more and more urgent.In the remote sensing images captured by satellites and UAVs,the object targets are small,dense and have a rotating angle,and most conventional target detection algorithms based on deep learning are horizontal frame detection algorithms,which are difficult to effectively detect object targets in complex scenes.Although the existing rotating target detection algorithms have improved detection accuracy,their model memory is too large and the detection speed is only about 20 fps at the fastest,which makes it difficult to port to mobile devices such as satellites and UAVs with limited computing resources and realise real-time detection.Furthermore,the loss function of existing rotating target detection algorithms cannot accurately calculate the difference between the real frame and the predicted frame,leaving some room for improvement in rotating target detection.In this paper,we propose a lightweight rotating remote sensing target detector that can be deployed in mobile devices with limited computing resources,achieving a better balance between detection speed and accuracy.To address the problem of excessive amount of parameters in the existing target detection model,this paper builds up a lightweight DN-YOLOv5 based on the YOLOv5 baseline algorithm and lightly modifies the CSP1＿X module of the backbone network.As the detection capability declines relatively after the lightweight improvement,this paper draws on the idea of Densen Net and appropriately increases the number of CSP1＿X modules after dense connection to strengthen the feature reuse of the network and compensate for the reduced detection accuracy after light weighting.In addition,this paper also uses the K-means++ clustering algorithm to re-optimise the anchor frame for clustering and replace the H-swish activation function suitable for fast detection.Experimental results show that the detection speed of DN-YOLOv5 is significantly improved and the model memory is significantly reduced.To address the problem that conventional target detection algorithms are difficult to apply to complex scenes with small,dense object targets and rotating angles,this paper introduces the angle parameter θ based on DN-YOLOv5,adopts the OBB annotation method,and adopts RIo U and R-NMS operations for the rotation detector to achieve the function of rotating target detection.Aiming at the critical problem existing in the existing rotating target detection algorithm,the conventional loss function is difficult to solve,so this paper proposes an improved loss function: R-Focal-EIo U Loss.Firstly,R-EIo U Loss sets corresponding penalty terms for overlap rate,center point distance,and rectangular box edge length.At the same time,in order to focus more on high-quality anchor boxes during regression optimization,Focal terms are added to R-EIo U Loss based on the idea of Focal Loss.The experimental results show that the rotating target detector with improved loss function overcomes the critical problem,and the detection accuracy is further improved,which fully meets the requirements of terminal deployment.