Research And Implementation Of Remote Sensing Image Semantic Segmentation Algorithm Based On Deep Learning

The development of aviation technology and sensors provides sufficient conditions for the use of remote sensing images,and the semantic segmentation technology of remote sensing images is a research hotspot,which has important significance in natural resource monitoring,crop extraction,smart city construction and road extraction.In recent years,with the rapid development of artificial intelligence,AI remote sensing has become an active field.In this paper,the deep learning method is used to study the semantic segmentation technology of remote sensing images.Several semantic segmentation algorithms are designed and verified on relevant data sets,and good segmentation results are obtained.The main work is as follows:Firstly,based on the full convolutional neural network,a multi-channel deep convolutional neural network is constructed for semantic segmentation of remote sensing images.Its main innovations are as follows:(1)The classification network Inception V-4 network is taken as the backbone network,and the network structure of the backbone network is modified as the encoder of the full convolutional neural network.Use the Inception convolution block to enhance the width of the network without losing the depth of the network,thereby capturing more abstract features and improving computational efficiency;(2)A modified atrous spatial pyramid pooling(ASPP)module was introduced to extract multi-scale features of the target from different training stages.At the same time,context feature fusion is carried out in the trunk network to preserve more spatial features;(3)An effective decoder network was designed to complete the construction of the whole segmentation network.Experimental results show that this network has good performance and can effectively segment remote sensing images.Secondly,based on the above research foundation and combined with perceptual loss,a remote sensing image semantic segmentation network with dynamic perceptual loss is proposed.The network is improved by using the network developed above,and its innovations are as follows:(1)The perceptual loss module is introduced on the basis of the previous research content,which uses the pretrained VGG19 network as the perceptual loss network;(2)Comparing the changes of the two loss functions in the training process,a dynamic loss fusion strategy is designed to better improve the segmentation details;(3)Simplify the design of the decoder and make the decoder more concise.The experimental results show that the semantic segmentation framework obtained in this research has superior performance,and the segmentation results are more refined,which can be well applied to remote sensing image segmentation.