Research And Implementation Of Vehicle Object Recognition Method Based On UAV Image

Target recognition has always been a research hotspot and difficulty in computer vision.Vehicle target recognition in unmanned aerial vehicle(UAV)images is of great value in both civil and military fields.The vehicle targets in UAV images are generally small,and the details of the target are lost seriously after the network operation,which leads to the poor recognition effect of existing identification algorithms on vehicle targets in UAV images.This paper mainly studies vehicle target recognition algorithms in UAV images.Based on the two most representative algorithms in the field of target recognition,the algorithm is improved according to the characteristics of the data set in this paper.The main work of this paper includes the following parts.(1)A large-scale vehicle target recognition data set based on UAV image is constructed.Large-scale data sets with target category labels and target locations are the basis and core of target recognition by deep learning.The UAV image data set constructed in this paper includes a total of 1,978 images,including nearly 100,000 vehicle targets.(2)Based on region-based Convolution Neural Networks(Faster RCNN),vehicle target recognition in UAV images in this paper is improved.The original Faster RCNN has a poor recognition effect on small vehicle targets in UAV images,especially the incomplete targets.In this paper,different anchor combinations are used to select the optimal anchor size and number,which increases the AP value of network detection by 8.0%;Then,aiming at the problem of detail information loss caused by network operation on small targets,multi-layer feature fusion is added to the network,which improves the detection AP value of the network by 2.1%.The AP test of the improved network increased by 10.1% in total,and the accuracy increased from 80.5% to 90.6%.However,due to the increase of the number of anchor,the speed of the improved network recognition decreased slightly.(3)Based on You Look Only Once(YOLO)network,vehicle target recognition in UAV images in this paper is improved.This paper uses v3 version of YOLO,which fails to detect incomplete targets and partial complete targets in UAV images in this paper.In this paper,k-means++ algorithm is used to carry out K=9 clustering on the Ground Truth of training data set,and the clustering center is used as the initial frame of the network,which improves the network's missed detection of small targets and improves the AP value of the network detection by 4.31%;The non-maximum Suppression(NMS)algorithm was improved by using the Soft non-maximum Suppression(Soft-NMS)algorithm,which improved the error Suppression problem of partial complete targets by the NMS algorithm and increased the AP value of the network by 1.17%.The overall AP value of the improved network increased by 5.48%.In this paper,COWC,VEDAI and Chinese Academy of Sciences CAR are also used to verify and compare the robustness of the Faster RCNN and YOLOv3 networks.The advantage of YOLOv3 network is the accuracy and speed of identification,while the advantage of Faster RCNN lies in the better robustness of the network.