| With the development of aviation technology,the use of drones has become more and more popular.Among them,the use of drones to obtain feature information has become the mainstream method.How to efficiently identify useful information from these massive aerial images,has become a current research hotspot.In recent decades,road recognition has been a hot research direction of many scholars.Roads are not only widely used in areas such as intelligent transportation systems,geological hazard analysis,and urban-rural planning,but also provide significant help for national military operations.Although some progress has been made in road image recognition in aerial photography in recent years,the accuracy of road recognition is affected by several factors.The main factors are as follows:(1)The drone has a high flying height,and the area of the ground feature information it captures is large.The ground resolution of the image is correspondingly reduced,and the edge and texture characteristics of the road information are reduced accordingly.When meters,the area ratio of the feature information captured is 4:1,and the area ratio of the corresponding road information is 4:1.Therefore,the accuracy of 200 meters road recognition may be lower than that of 100 meters road recognition.(2)During network training,the normalization of the data set leads to missing road edges and textures.(3)The background of the road is complex,and the buildings,land,etc,on both sides of the road have similar characteristics to the road.The above problems bring great challenges to the accuracy of road recognition in aerial images.Therefore,this paper is dedicated to studying the improvement of the road recognition rate of drone aerial images.The specific research content is as follows:1.Based on the existing network,road recognition of aerial images is realized by optimizing the network.Based on the UNet network model,the deep separable convolution feature of Mobile Net is used to optimize UNet to form an M-UNet network.The Seg Net network replaces the pre-trained network Res Net50 to replace the original VGG16 to achieve road image feature extraction,and adds a long jump layer structure to perform feature fusion between codecs and reconstruct it into an S-Seg Net network.Through analysis and comparison with D-Link Net networks of different depths,the S-Seg Net network has the best segmentation effect,which can not only extract the advanced features of the road,but alsoextract the detailed features of the edge of the road.M-UNet network is not only suitable for computers with lower configuration,but also has better segmentation.2.In order to reduce the ground resolution of the image due to the high flying height of the drone and the target texture and edge loss caused by the normalization of the dataset during network training,this paper proposes a preprocessing method for image super-resolution reconstruction before network training.Based on the CNN network,the super-resolution reconstruction of the network training data set is achieved by fine-tuning this network.The super-reconstructed data set is experimentally verified on the D-Link Net101 network.Compared with the experimental analysis of the original data set,the over-score data set improves the accuracy of road recognition by improving the edge and texture characteristics of the road,and its average merge ratio and average accuracy are respectively increased by 2.85% and 2.34%.3.Focusing on the convergence speed of the D-Link Net101 network,by adding parameters to the loss function,the convergence speed of the network is slightly improved,and experimental verification is performed on the over-scored data set.Experiments show that the accuracy of road recognition is improved by this method,even if MIOU and MA only improved by 1.28% and 1.26%. |
