Research On Semantic Segmentation Algorithm Of UAV Aerial Photography Image

Due to the advantages of convenience,speed and low cost,UAV technology has been widely applied in various industrial fields and emergency handling tasks.The UAV aerial image play a huge role in the process of high efficiency operation.The research on semantic segmentation of UAV images is helpful for the extraction of UAV aerial photography image information,the monitoring of the flight environment and the completion of UAV flight operations.However,in practical UAV image semantic segmentation tasks,the high resolution of aerial images often makes it difficult for the network to achieve high speed,and the characteristics of multi-scale and high complexity of targets in aerial images add difficulties to the semantic segmentation task of aerial images.Therefore,this article focuses on the semantic segmentation task of drone images and conducts research in the following four aspects using deep learning technology:(1)Aiming at the problem of low semantic segmentation accuracy of tiny targets in UAV aerial image,a bilateral semantic segmentation network T-BiseNetv2 is proposed.The proposed segmentation network is based on BiseNetv2,and the Transformer structure is introduced into the semantic branch to improve the network’s ability to extract the global context information of the image.And the detail branch and feature fusion module are reconstructed,which improves the network’s ability to pay attention to the details of the image.The experiment results on the standard dataset UAVID show that the MIoU of the verification set reaches 69.52%,which is about 2.2%higher than original BiseNetv2 network.For the small target category "Human",the IoU of T-BiseNetv2 proposed about 4.3%higher than UNetFormer,which indicates the feasibility of the proposed algorithm.(2)Aiming at the problems of multi-scale targets in UAV images and low segmentation accuracy of existing networks for small objects,a U-shaped lightweight semantic segmentation network EFU-Net is proposed.In order to extract feature maps of different scales and model multi-level local information in the decoder,this paper designs a gather-and-expansion module based on inverted bottleneck structure and applies it to the decoder of EFU-Net.To learn more generalized fusion features,the weighted fusion method is used to fuse the feature maps of the corresponding stages of the expansion path and the compression path.To improve the segmentation accuracy of the network,the auxiliary segmentation head is added to the decoder stage of the network,and the loss value of the training stage is increased.Finally,our EFU-Net is analyzed on the standard dataset UAVID.The experimental results show that the segmentation ability of EFU-Net for multi-scale targets is better than T-BiseNetv2.On the other hand,the EFUNet network can also meet real-time requirements.(3)The semantic segmentation dataset of UAV images in campus scene is constructed.In order to realize the semantic segmentation task of campus scenes with UAV images,this paper uses DJI M300 UAV equipped with Zen Si H20T camera to shoot video sequences in the course of UAV flight,including images of winter and summer scenes.Among them,179 images were selected to construct UAVCampus,a semantic segmentation dataset of UAV images in the campus scene,including 130 training sets and 49 verification sets,with image resolution of 4096×2160.The EFU-Net network was trained on this data set,and 88.68%MIoU was obtained on the verification set.Finally,the video sequence shot by DJI M300 UAV was segmtioned with the weight obtained from training on UAVCampus dataset,and the robustness of EFU-Net network was verified.(4)In order to verify the effectiveness of our EFU-Net in the actual flight of UAV,a Tello UAV real-time semantic segmentation platform was built.We utilize the DJI Tello UAV to build a real-time semantic segmentation platform for UAV images,enabling real-time communication between the UAV and the ground segmentation platform.The real-time semantic segmentation task for UAV images is carried out in a real campus environment,and the EFU-Net is verified in practice.The experimental results indicate that the EFU-Net can achieve real-time semantic segmentation in real environments.