Research On Building Extraction Method Based On Full Convolution Neural Network

In recent years,the spatial resolution of remote sensing images is continuously improved,it reaches the decimeter level.In the high resolution remote sensing images,the ground information is more abundant.The building is one of the important components in the surface target information.How to quickly and accurately extract the buildings in the image is always a hot issue.The traditional building extraction algorithm is mainly based on the processing of pixel spectral information to obtain the desired features.However,due to the large amount of remote sensing images and its high update speed,manual interpretation and pixel-based extraction methods are no longer applicable.It is necessary to explore new methods for extracting information in the high resolution remote sensing image.In recent years,deep learning develops rapidly,and it has achieved amazing results in image recognition that are difficult to previous classification algorithms.This method can learn more advanced features from a small amount of preprocessed or unprocessed raw data,and it has been widely used in technical fields such as license plate detection,driverless,target detection and tracking.This thesis studied the existing remote sensing image classification algorithms at home and abroad,and based on the deep learning theory,draws on the U-Net and SegNet decoding ideas and the dilation convolution method and multiscale semantic segmentation to construct a new convolutional neural network structure.The proposed network is trained on the building dataset to verify and evaluate its effectiveness.The main research contents and results of this thesis are as follows:(1)Based on the traditional convolutional neural network and the idea of coding and decoding,this thesis built a building scene segmentation network(BSSNet),and establishes a high resolution aerial orthophoto dataset,which is used to train and verify the convolutional neural network,and data augmentation methods are used to extend dataset for small sample sizes and uneven distribution between classes in the dataset.(2)Using multiple continuous convolutions with different dilation rates between encoding and decoding,it solves the problem that the common convolutional neural network has a small receptive field,the extracted target boundary is blurred,and the result is incomplete.At the same time,the partial convolutions form a series and parallel structure,and the features of different scales are extracted and merged to obtain good performance.(3)For the problem that the boundary of the building extracted from the convolutional neural network model is not regular enough,and the result is incomplete,the post processing operation of the image is performed by using the fully connected condition random field,guide filtering and mathematical morphology to achieve the best prediction.(4)In order to verify the effectiveness of the proposed method,this method and other convolutional neural networks were qualitatively and quantitatively analyzed on the same test dataset.The results show that the classification results of this method are better than other models,and the building boundaries are relatively regular,with low noise and certain practical value.