Research On Road Recognition Of Remote Sensing Image Based On Generating Countermeasure Network

With the development of remote sensing technology,higher resolution of remote sensing data can be obtained.Remote sensing has the characteristics of fast information acquisition,short cycle,wide detection range and less limited conditions,so remote sensing images are widely used in various fields,such as surveying land resources and updating urban road network.As an important part of urban construction,the accurate extraction and identification of road can bring great significance to the rapid update of urban road network,traffic development,precise automobile navigation,urban planning and the update of basic geographic information database.Therefore,accurate identification and extraction of road becomes particularly important.Since the 1970 s,researches on extracting road information from remote sensing images have been carried out at home and abroad.However,the manifestation of road in high-resolution remote sensing images is relatively complex,which makes the extraction and identification of high-resolution remote sensing road very difficult.At present,most road extraction technologies still have the problems of poor algorithm robustness and low recognition accuracy to some extent.Therefore,remote sensing image road extraction and recognition is still a hot topic in the industry,with very important academic value and application prospect.Based on the analysis and research on the existing shallow machine learning method to extract road,on the basis of combining the characteristics of high resolution remote sensing images,an improved method to extract the Bayes road segmentation,in considering the remote sensing images,the geometrical characteristics of the road can present a more specification and easily affected by environment factors such as the noise is very sensitive and algorithms in image segmentation with the Bayes increased the mean shift filtering processing,can mean shift filtering to eliminate noise images of the local color close smoothly,and can keep the edge contour,well after filtering the image than the original image is more easy to recognize.Next,Bayes classifier is used to identify the filtered image according to the color features of pixels,and the general shape of the road can be segmented.Finally,linear morphological filter is used to process the preliminary segmentation of road image again,which can further remove irregular regions and obtain more accurate road segmentation image.Although this method tries its best to reduce the influence of ambient noise and can solve the problem of road extraction in complex scenes to a certain extent,its accuracy rate is still not high,and its robustness is not strong and its generalization ability is insufficient.In remote sensing images,the road is banded,but its geometric shape is changeable,and it does not have fixed color or texture features.Therefore,it is very difficult to extract road features,which makes the traditional shallow machine learning method relying on artificial feature extraction difficult to obtain a good road extraction effect in complex scenes.In recent years,deep learning and convolutional neural network have developed vigorously,and the use of convolutional network for feature extraction has surpassed the artificial feature extraction method in many aspects.In view of this,this paper further studies the road extraction method based on convolutional neural network.The local features of image were obtained through the downward sampling of convolutional network.The road and non-road were classified by the full connection layer.The activation function sigmoid was introduced to control the size of the pixel value output by the full connection layer between 0 and 1,and thereby represent the probability of road and non-road.The effect of convolutional network is improved than Bayes before,but it is still not perfect.The convolution network blurs the edge information in the process of down sampling.To solve this problem,this article also proposed another deep network model: emergent against the way of neural network identification,its network architecture consists of encoder and decoder networks,in the process of each layer of the sample will be combined with the previous layer information of convolution and pooling,its core idea is to make generation network and network to form a mutual "game",the concept of up and down sampling influence each other each other constraints,end up with a dynamic stable status;In addition,the residual network mechanism is introduced into the generated network G to optimize the original GAN network.Both of these deep learning methods can eliminate the impact on noise and enhance the robustness and generalization ability of the code.At the same time,the segmentation precision of generative antagonism network can also maintain a good level of road recognition in complex scenes,which is a good classifier.