Semantic Segmentation Of UAV Remote Sensing Image Based On Deep Learning

Compared with general image semantic segmentation tasks,UAV remote sensing image segmentation is complex,the data set for segmentation training is smaller and the number of data sets is small,and the similarity between each segmentation scene of the data set is higher..The development of drone technology is very rapid,and it has begun to be used in production and life on a large scale.The demand for image processing technology for UAV remote sensing images is increasing day by day.Using traditional methods to build a remote sensing image semantic segmentation model has complex design and poor generalization ability.In recent years,the rapid development of deep learning and continuous breakthroughs in the fields of computer vision and artificial intelligence have also pointed out a new direction for the research on the semantic segmentation of UAV remote sensing images.The experimental part of this article uses public data sets,mainly including the Potsdam data set provided by the International Photogrammetry and Remote Sensing Association.The research content of this article mainly focuses on the following two aspects:1.Semantic segmentation of UAV remote sensing image based on U-Net network.U-Net is a typical encoder-decoder structure network.The encoder part mainly completes feature extraction through downsampling,and the decoder part mainly completes the up-sampling task: restoring the resolution of the image.The low-level semantic information is retained in the network,and the features are spliced during upsampling.Therefore,the U-Net network can obtain more spatial information and low-level semantic features than the full convolutional network(FCN).In this paper,Res Net34 is used instead of VGG-16 as the basic structure of the network,and the residual block is introduced to deepen the number of network layers and improve network performance.The attention mechanism can simulate the process of human brain processing visual signals,and can make the network pay more attention to the area of interest for the target task by weighting the attention features.The attention mechanism module can strengthen the attention to features and improve the performance of a certain feature of the network.In order to strengthen the spatial and channel features of U-Net,the attention mechanism is introduced.2.Build a semantic segmentation model for UAV remote sensing images with full convolutional network as the core.Convolutional neural network is the predecessor of full convolutional network.The difference between the two is that the full convolutional network replaces the former fully connected layer and replaces it with a fully convolutional layer.This paper uses Res Net50 instead of VGG-16 to build a full convolutional network model.The full convolutional network loses a lot of detailed information during downsampling,so the low-level semantic features and high-level semantic features are added during the upsampling operation.Considering that although the model incorporates 8 times upsampling features,it is not sensitive to multi-scale information.Therefore,a new hollow convolution pooling pyramid is designed and embedded in the model.The hollow convolution can be used without reducing the resolution.Enlarging the receptive field and convolving multiple holes in parallel can obtain different receptive fields,so that the model has the ability of multi-scale target recognition.Although the cross-entropy loss function can effectively guide the segmentation,the MIOU index of the model has not been improved.In order to improve the network evaluation index: MIOU,a multi-category loss function based on the intersection-to-union ratio(IOU): Lovasz Softmax Loss function is selected to improve the performance of the model.