Application Of Deep Generative Adversarial Networks In New Building Detection

The statistics and management of urban buildings is a very important task for government departments.But as the city expands,the work of rapidly expanding the number and variety of buildings becomes more and more difficult.The use of new technologies to solve the statistics and management of urban buildings is an inevitable trend of the times.The main research goal of this paper is how to quickly and accurately find new buildings in the city to facilitate the management of staff.The use of high-resolution remote sensing satellite imagery to identify urban buildings has been a hot topic in recent years.In the past,some experts and scholars tried to detect buildings based on the color,reflectivity,shape and other characteristics of buildings in remote sensing satellite images.However,these algorithms have very high quality requirements for data sets,and they are somewhat weak in the face of satellite images,deviations,distortions,and tilt problems caused by different angles.In order to improve these problems,this paper proposes a new building detection algorithm based on Deep Generative Adversarial Networks.First,the two satellite image channels will be directly superimposed together to see a double channel image.This new building detection problem can be simplified to a segmentation problem.In this way,the newly added building parts are directly divided without separately identifying the buildings of different periods for subtraction.The deep-generation generation network in the Adversarial network uses a 28-layer full-convolution neural network to segment new buildings.This network inputs the superimposed images and standard labels in training,and then performs multiple volumes.Product,downsampling and upsampling,and finally get a two-dimensional image.The discriminant network uses a 5-layer full convolutional neural network,and the input is a segmentation image generated by the network.The image is then subjected to multiple convolution operations to obtain a vector,and finally a normalized value of 0-1 is calculated to evaluate the quality of the input image.This judgment result is weighted into the loss function of the generated network.In the training,the network is fixedly generated first,the discriminant networkis trained,and then the discriminant network is fixed,and the training network is trained.The discriminating ability of the discriminating network is continuously strengthened in repeated iterations,and the generating network is continuously provided with a new loss function,and the generating network can utilize the updated The loss function trains better results.The experiment uses the public data set in the Ali Tianchi Big Data Competition.The channel stretching algorithm is used to make the numerical distribution more uniform.At the same time,the random set selection and stretching are used in the training set construction process to increase the randomness of the training set.Sex.Finally,the experimental comparison shows that the F1 scores of the proposed algorithm are 5.6% and 13.6% higher than those of U-Net and FCN,respectively.In this competition,he also won the second place in the final.