Research On Object Detection Algorithm For UAV Images Based On Deep Learning

As one of the key technologies to improve the autonomous sensing ability of UAV and intelligently process massive data,object detection opens up application scenes of UAV.At present,most of the mainstream algorithms in object detection are based on deep learning methods and convolutional neural network,which have strong ability of feature learning and representation.However,these algorithms are mainly designed for natural scenes,while the effect for detecting in UAV aerial scenes is not acceptable.In addition,object detection algorithm is required to meet the performance characteristics of the UAV like lightweight and real-time processing.To address these problems,a novel object detection network which is named VC-YOLO is proposed.The main research work includes:The characteristics of UAV aerial images are analyzed,such as complex background and high proportion of small objects.The experiments of the benchmark algorithm YOLOv3 on NWPU VHR-10 and Vis Drone datasets show that the complex background and small objects have a negative impact on the accuracy.At the same time,the storage and computational performance of the benchmark algorithm are measured to prepare for the follow-up work.The storage requirement of the benchmark algorithm is improved.A lightweight convolutional neural network REFNet is designed as the backbone network to replace darknet53,which is the backbone of YOLOv3.The receptive field of residual block is enlarged to adapt to the characteristics of aerial images.The accuracy of the benchmark algorithm is improved.In view of the complex background and the small objects in UAV aerial images,a multi-scale attention network with enhanced local feature is applied and Feature Pyramid Network(FPN)of the benchmark algorithm is improved.Finally,through the experimental verification,the parameters of VC-YOLO are 28.6% of YOLOv3,the model size is 28.7% of YOLOv3,and the detection speed is 58 FPS.The detection accuracy of VC-YOLO surpasses YOLOv3 by 2.3 m AP on Vis Drone dataset.