Research On Semantic Segmentation Algorithm Of Urban Remote Sensing Image Based On Improved SegNet

The high-resolution urban remote sensing image has complex details and rich content,which can accurately reflect the surface information.Using deep learning technology to perform semantic segmentation on urban remote sensing images can continuously and dynamically detect the information of urban roads,water bodies and buildings,which is conducive to a comprehensive and efficient understanding of the development of the city.This paper focuses on the difficulty of segmentation of high-resolution urban remote sensing images due to complex backgrounds and different target scales.The following studies are carried out:(1)Aiming at the problem of small data set samples and unbalanced category distribution.First,the original image and the label image are cropped with a certain step size,and then the data set is augmented by operations such as mirror flip,Gaussian noise,translation transformation,and light adjustment.At the same time,over-sampling of small samples in the data set can effectively avoid over-fitting in the training process and improve the generalization ability of the network.(2)On the basis of the application of SegNet and U-Net network for the research on the semantic segmentation of urban remote sensing images,based on the SegNet model,this paper proposes a layer-by-layer feature fusion network model U-SegNet.The parameter bilinear interpolation method restores the spatial information of the feature map layer by layer,and at the same time introduces skip connections to fuse the results of the deep feature map upsampling with the corresponding size of the shallow feature map in the encoder,which effectively solves the segmentation of the SegNet network.The sparseness phenomenon in,enhances the segmentation effect.(3)In view of the existence of targets of different scales in urban remote sensing images,the U-SegNet network cannot improve the segmentation accuracy of small targets.On the basis of U-SegNet,DAU-SegNet is proposed,which deepens the network by introducing residual blocks to characterize higher-dimensional feature information,and at the same time introduces the atrous spatial pyramid pooling,and uses hole convolutions with different hole rates in parallel.Using different scales of receptive fields to achieve multi-scale feature extraction.(4)Inspired by ensemble learning,the four learners of SegNet,U-Net,U-SegNet and DAU-SegNet are model-fused.The segmentation result map output by each learner is voted pixel by pixel,and the category with the most votes is taken as the category of the pixel,and the segmentation results are integrated and optimized to improve the accuracy of the segmentation.The experimental results show that the DAU-SegNet network proposed in this paper has achieved a good segmentation effect on the public CCF data set.