Extracting Urban Building Tops From Remote Sensing Images Based On Augment Feature Pyramid Network

With the rapid development of sensor technology and mounting platforms,the spatial resolution of remote sensing images has gradually increased.High-resolution images have provided clearer and richer surface information,with buildings accounting for more than 80% of urban surface information.Most of the traditional buildings extraction methods require artificially designed features or parameter selection.The process is time-consuming and labor-intensive,and the deep semantic features in the image cannot be obtained.In recent years,more and more deep learning algorithms have been applied to remote sensing image processing,in which convolutional neural networks can automatically learn advanced features from the given data,so as to perform accurate target extraction.However,during the forward propagation process of the most existing networks,the information loss of the feature map is relatively large,the information fusion of the target building at different scales is slightly insufficient,and the location of the extracted building boundary is not accurate and clear.Therefore,the accurate extraction of buildings from high-resolution remote sensing images is still the focus of current research.In view of the above problems,this paper proposed an aerial images urban buildings top extraction method based on Augment Feature Pyramid Network(AFPN)by investigating and analyzing the domestic and foreign research status of remote sensing images buildings extraction,and experiments were conducted on building datasets including Chicago,Kitsap,etc.The main work is as follows:(1)Taking the existing ResNet and FPN structure as the basic framework,the visual attention mechanism and the atrous convolution group are introduced to construct an Augment Feature Pyramid Network without pooling layers.In the early training of the network,through the choice of optimizer and batch size,and the addition of the dropout layer,a set of optimization solutions was explored.(2)Aiming at the problems of jagged and mosaic-like buildings boundaries extracted by the proposed model.Respectively,the method of corner detection and fitting is used to regularize the boundaries of straight-line buildings,and the grouped Douglas-Peucker algorithm is used to regularize the boundaries of arc-shaped buildings.(3)Experimental samples were selected from the INRIA ortho-photo images buildings dataset,and three datasets containing different regions were constructed through images cropping and data enhancement,respectively,to evaluate the accuracy of the model in this paper and research the impact of training data differences on model performance.Then,this paper trained the constructed AFPN model on the produced building datasets and analyzed the experimental results.At the same time,in order to verify the effectiveness of the AFPN model,two classic networks,ResNet-34 and FCN-8s,are selected for training on the same dataset,and compared with the experimental results of the proposed method.The research shows that the model constructed in this paper extracts buildings top with high accuracy,less noise and complete graphics.The boundaries regularization method can effectively reduce or eliminate jagged and mosaic shapes,simplifying redundant details on the boundaries.And the AFPN model is superior to ResNet-34 and FCN-8s in terms of prediction accuracy and visual effect,and the accuracy index values on the prediction dataset are all above 80%,the average accuracy rate is over 93%,which has certain practical value.