Research On Small Object Detection Method Of UAV Aerial Image Based On Deep Learning

Using images obtained by unmanned aerial vehicle(UAV)for object detection is of great significance in fields such as power inspection,traffic monitoring,crop monitoring,and forest fire warning.However,UAV aerial images are often affected by various factors such as shooting height,angle,and weather conditions,and the targets to be detected are small in size,densely distributed,difficult to distinguish from the surrounding background,and difficult to improve detection accuracy.To address these issues,this thesis proposes a deep learning-based method for UAV aerial image detection.The main work of the thesis is as follows:(1)Construct a drone aerial image dataset.To address the problem of insufficient images under complex conditions in existing drone datasets,the Visdrone dataset is used as the basic dataset.In addition,drone-captured images in nighttime and rainy/foggy weather are used as expansion images,and samples that are not suitable for annotation and training are removed.Then,the collected images are expanded by rotation,mirroring,adding noise,etc.to enhance the model’s generalization performance.Using the Label Img image annotation software,the seven types of objects in the collected images are manually selected and boxed,and all.XML files are integrated to obtain datasets in both VOC and YOLO formats.(2)Using four algorithms,Faster R-CNN,SSD,YOLOv5,and YOLOv7,training and testing were conducted on the constructed UAV aerial image dataset.A total of six sets of test results were obtained.After comparative analysis,it was found that the YOLOv7 algorithm performed better on the UAV aerial image dataset.(3)In order to further improve the performance of YOLOv7 on UAV aerial image datasets,the network structure of YOLOv7 has been improved.Compared with the original model,the m AP of the improved model has increased by 3.6 percentages in total.To enhance the saliency of small object features,CBAM and Sim AM attention mechanisms were respectively fused into the MP module and SPPCSPC module of YOLOv7.To achieve the fusion of contextual information between feature maps of different scales,a multi-scale feature fusion network based on RS-ELAN was proposed,which adopted a non-layer-crossing skip connection to preserve more small target information and maintain gradient stability during continuous convolution operations of the feature maps in the ELAN module.To enhance the contribution of high-quality anchor frames to the regression loss,the Focal-EIOU loss function was introduced to assign larger weights to high-quality anchor frames.This thesis analyzes the characteristics of small object in UAV aerial images,builds a object dataset,compares commonly used object detection algorithms,improves the YOLOv7 detection algorithm with good performance,tests the improved YOLOv7 small object detection model for UAV aerial images,and the actual test results show that the improved target detection model has significantly improved performance.It has practical significance for identifying and locating small object in UAV aerial images,and has certain reference value for small object detection in other fields.