Semantic Segmentation Of Remote Sensing Images Based On Convolution Neural Network

In recent years,the deep learning developed rapidly,became an important branch of machine learning,and promoted the development of computer vision technology.The deep Convolution neural networks(CNN)has been proved to be one of the effective methods in the field of computer vision,the famous Large Scale Visual Recognition Challenge named ILSVRC is an application based on CNN.The computer vision technology can be frequently applied to object detection,video surveillance,target retrieval and other practical application scenarios.However,the image semantic segmentation is one of the basic technologies of computer vision.The image semantic segmentation technology helps the computer to understand the image or video content,filter the image semantic information,its accuracy directly affects the final outcome of the computer vision task.In other words,a good image semantic segmentation method is very important for the computer to complete visual tasks.The traditional image semantic segmentation method is mainly based on the feature extraction of the image itself,images are divided into different regions,and then extract different characteristics in different regions,finally,the region is merged and the image semantic segmentation task is completed.Easy to know,this method is complicated and difficult to understand,and the images can’t be batch processed.With the continuous improvement of practical application requirements,the traditional image semantic segmentation method do not work for the practical application scene.As the result,researchers try to use the CNN network with excellent feature extraction capabilities to complete the image semantic segmentation tasks.With the development of remote sensing technology,people can obtain a greater number of remote sensing images,the processing of remote sensing images is also becoming more and more complicated.The existing remote sensing image processing methods are mainly rely on the priori knowledge of mankind,we need to spend a lot of artificial resources to mark and identify the remote sensing images so that to complete the image preprocessing task.Based on the above background,we introduce the deep CNN network into the pretreatment of remote sensing image,the semantic segmentation task of remote sensing image is realized based on the deep CNN network.In this paper,a complete convolution neural network is constructed by modifying the network structure of these three deep CNN networks: Alexnet,VGG-16 and GoogLeNet.And then we use the complete convolution neural network to realize the semantic category division of each pixel of remote sensing image.Firstly,we colored the label image and carry on the edge extraction for the remote sensing image.Then,we try three different network training schemes to train the parameters of these three kinds of complete convolution neural networks.Finally,by comparing and analyzing the results of target matching rates and semantic segmentation images,the optimal network training scheme and the optimal semantic segmentation model of the remote sensing image are obtained.According to the experimental results obtained in our paper,it can be considered that it is feasible to apply the deep CNN network to the semantic segmentation task of remote sensing image.