Research On Semantic Segmentation Method Of Remote Sensing Image

High resolution remote sensing images include relatively complex semantic information and easier confusing objectives.Semantic segmentation of high-resolution remote sensing images is an important and challenging task: such as different types of objects(such as rivers,roads)in remote sensing images,usually have highly similar appearance in the middle and interference identification;sometimes object It will be blocked,and the existence of various noises during the acquisition of images and transmission images,the noise is considered to be invalid in the image,so the input image also causes damage to the division.The convolutional neural network has its own characteristics of its improved image recognition performance throughout the various image processing research.Semantic segmentation of the convolutional neural network is to automatically obtain the meaning of each pixel in the image,followed by classifying and identifying the pixels,and is now gradually starting to optimize the extension of image segmentation accuracy..However,remote sensing image semantic division tasks are cumbersome,as this network depth increases and deepens,the problem of gradient disappearance will be more obvious,such as small objective deletion,and different images are easier to roughen the edges.In this regard,this study analyzes the existing depth network learning model to perform semantic semantic semantic segmentation of remote sensing images by analyzing the existing convolutionary network model,and proposes two optimization models based on convolutional neural networks.At the same time,the comparative experiment of a variety of different semantic segmentation models in remote sensing images is mainly completed as follows:(1)Research on High Resolution Remote Sensing Image Variation Detection Detection Algorithm Based on Diffusion Convolution and Channel Attachment Mechanism.The U-NET network module is used as the basic network extraction feature,and the DENSE-NET intensive module is used to enhance the U-NET,and the focus of densely consuming channels is introduced into the basic convolutional unit,and the Dense-Net is respectively The PSP NET,U-NET model performs comparative experiments,complete the semantic segmentation of the fused reward,the experiment proves that the DENSE-NET model segmentation effect is more ideal for other two network models,and the experimental results are verified.The effectiveness and robustness of research.(2)Study on the Earthwork of Works Based on Deep Convolution Neural Network Calm.First,a semantic segmentation network that deals high-resolution remote sensing images is proposed in the Subject Class CRF algorithm.Use the depth neural network architecture to complete the Internet construction,through 85 items of updated training to improve the generalization of the Internet,by using K-Means algorithm with Iso Data,using the two sets of data sets based on ENVI4.8-based experimental environments Building a network model handling semantic segmentation tasks for comparison experiments,verifying the depth convolutional neural network combined with CRF algorithm models in combination with CRF algorithm.(3)Research on Remote Sensing Image Demographic Method Based on Asymmetric Roll.Asymmetric Convolution Blocks,ACBs are added as CNN,and then the characteristic expression of the square volume core is enhanced by using one-dimensional non-symmetrical convolution to enhance the characteristics of the square.Training to higher precision,that is,the ACB replaces the original square core,to construct an asymmetric convolution network(AC-NET),the AC-NET network model and the De EPlabv3 network model,SE NET Network model,U-NET network model,PSP NET network model,and CE NET network model to compare experiments,which proof AC-NET can train network to higher accuracy based on its own validity,and simulation results confirm the effectiveness of this study.Sex and robustness.