Research On Object Detection Algorithm Of UAV Image Based On Deep Learning

UAV technology has important application value in fields such as agriculture,surveying and mapping,and military affairs.It obtains important video image data through aerial photography.However,UAV aerial images have the characteristics of high resolution,complex background information,uneven illumination changes,and small scale of objects of interest.Therefore,directly using existing object detection methods for the detection of objects of interest in aerial images will lead to poor detection result.Based on the above considerations,this paper proposes an improved UAV target detection algorithm based on YOLOv5 to focus on solving the detection problems of high resolution aerial images and small targets.This paper mainly optimizes the three parts of data preprocessing,feature extraction and network detection head to enhance the detection model’s ability to perceive small targets in aerial images.The main research contents are as follows:This paper proposes a small target detection method for UAV aerial images based on improved YOLOv5.Aiming at the high resolution of UAV aerial images,and the small shooting target can be learned with less feature information,A data enhancement method using a magnifying glass is proposed to improve the model’s ability to learn the feature information of small targets by increasing the relative size of small targets in the picture,and initially improve the detection effect of small targets.Secondly,for the problem that the three detection branches in the YOLOv5 model are not enough to cover all small targets,and its given default preset anchor box is not suitable for the detection of drone aerial image datasets,This paper proposes a reconstruction detection branch method to reduce missed detection by increasing the model’s capture of small targets of 12 to 24 pixels in aerial images;at the same time,we optimize the preset anchor frame to adapt to the actual situation of many small targets in aerial images.The detection effect of small targets is further improved.Experiments on the Visdrone dataset show that the accuracy of bicycles,motorcycles,and pedestrians in small objects is increased by 11.7%,6.2%and 4.8%,respectively,and the recall rate is increased by 1.2%,4.4% and 3.5%,respectively.This method can effectively improve the detection accuracy and recall rate of small objects.In this paper,we propose an object detection method for UAV images based on cs SE attention and refinement decoupling head.Aiming at the insufficient ability of the backbone network to learn small target features,the cs SE attention module is introduced in the backbone network.By adding the cs SE attention module to the lower layer of the backbone network,the model pays more attention to the information of the target to be detected rather than the background information.It effectively enhances the backbone network’s ability to learn small target feature information.Aiming at the problem that the detection head of YOLOv5 s cannot optimally express classification and positioning at the same time,the decoupled head is introduced in the head network to separate the classification task and the positioning task,and improve the overall detection accuracy of the target.The experimental results on the Visdrone data set show that the overall detection accuracy of UAV-YOLO has increased by 10.6%,the recall rate has increased by 3.0%,and the m AP has increased by 5.5%.Since that there is no visual operation interface for UAV detection at present,and the detection results are inconvenient to observe and compare the detection results of different models,this paper develops a UAV aerial image detection system to facilitate the detection of images and videos,and to compare the detection results of different models.