| Object detection and recognition based on aerial images of aerial platforms can be applied to military,civil and other fields.Compared with traditional object detection tasks,aerial images have characteristics of large viewing field,small targets,various distributions,complex background texture and so on,which greatly increases the training difficulty of deep learning models and affects their detection performance.In this paper,according to the characteristics of aerial images,relying on the project of ground detection and identification of an aircraft industrial company,a UAV ground object detection system for aerial platforms is developed,which effectively improves the object detection accuracy in aerial images.It has been put into practical application,and has strong practical value.The main contents and innovations of this paper are as follows:(1)In aerial images,small scale objects could be easily ignored in the process of model prediction,while a large number of undetected objects can be caused by the down sampling feature extraction in dense target areas.In this paper,a bi-directional feature fusion pyramid model BDFPN is designed,which can ensure the effective feature fusion of multi-scale feature maps,and improves the accuracy.At the same time,a rotated bounding box regression method is used in our network.And design regression balance and relation loss function,which effectively improves the accuracy of the precision of bounding box regression,and proposes the rotated object detection model in aerial images.In the rotating object dataset HRSC2016 and UCASAOD,the accuracy reaches 94.28% and 96.20% respectively.In the horizontal target detection task,the accuracy of VOC2007 test set is 85.4%,which is 4.2% higher than that of Retina Net.(2)To handle the problems of small proportion of aerial target and complex background texture,this paper designs and implements feature map optimizer FAM and IAM based on attention mechanism,constructing feature optimization network.Combining with rotated object positioning mask to retain the important features of small target in feature map,and reduces the situation of missing small target in aerial image.Two different attention optimization models are designed for single-stage detectors and two-stage detectors respectively,which effectively solve the problem of low dimensional feature loss with the increase of network depth.In the HRSC2016 dataset,m AP is increased by 4.75% in the single-stage detector based on Retina Net,and 2.76% improvement in the DOTA dataset based on FPN.(3)Aiming at the practical application requirements of aerial image object detection,this paper designs a ground target detection system,which is applicable for military reconnaissance,engineering simulation and other complex scenes and proposes a training and detection method for large-scale aerial images.Combined with the above optimization strategies,integrating data processing,model training,model switching,object detection and other functions,this system supports the deployment of Ground Workstation,and explore the UAV airborne.It can automatically detect and identify specific targets,achieve a high degree of stability and real-time,through experimental testing and practical application,the system can effectively detect the UAV aerial images of vehicles,ships and other targets,verifying the overall performance and practicability of the system.The aerial image ground target detection system developed in this paper has been deployed and applied in an aircraft industrial company.It has good practicability and economic value to analyze and test the real aerial image of UAV and has made an important contribution to solve the problem of ground object detection in aerial image. |
PDF Full Text Request