Research On Road Information Extraction Of High-Resolution Remote Sensing Image Based On Deep Learning

High resolution satellite remote sensing image plays an important role in urban and rural planning,military reconnaissance,environmental monitoring,intelligent transportation and disaster emergency response.As an important part of remote sensing image,road is a great significance in real life.With the continuous progress of urbanization and the rapid updating of road network,it is necessary to extract the latest road information from remote sensing images in order to effectively support road space applications.In the extraction of road information from remote sensing images,there are still many difficulties in the accurate extraction of road information due to the long and thin shape,huge span,high complexity and interconnection of roads.Based on the deep learning method,this dissertation builds a high-resolution remote sensing image road information extraction model,which uses the high-efficiency learning ability and feature expression ability of the deep learning network model to extract the complex road information in the remote sensing image data.The specific research work is as follows:(1)Aiming at the problem that the road information in the remote sensing image is complex and difficult to extract,a UNet road information extraction model(B-UNet)based on bilinear pooling is proposed.The B-UNet model introduces a central module based on bilinear pooling between the encoder and decoder,which makes the network model make full use of the global spatial feature information of the image,better obtain the second-order information,spatial long-distance information,and the correlation of different feature channels.Moreover,the B-UNet model adds a Batch Normalization layer between the convolution layer and the activation layer of the network,which improves the training speed and generalization ability of the network.In order to solve the problem of insufficient publicly available datasets for road extraction task,this dissertation built a road information extraction dataset of high-resolution remote sensing image.B-UNet model and UNet model are trained on the dataset.The experimental results show that the B-UNet model can extract the road information well,and each evaluation index is better than UNet model.(2)To solve the problem of unbalanced distribution of positive and negative samples(small proportion of Road area and large proportion of background),the sum of BEC loss function and Dice loss function is taken as the loss function of the model training.(3)In view of the problem that there are a few road information interruptions and omissions in the extraction results of B-UNet model,this dissertation first introduces the residual structure of Res Net34 network which has completed pre-training into BUNet model to build B-Res UNet model.The transfer learning technology is used to improve the accuracy of information extraction.The residual structure can make the network better integrate the feature information of different layers and enhance the generalization ability of the network.Then,the parallel multi-scale dilated convolution module is added to the bridge part of the model to build BD-Res UNet model.The parallel dilated convolution structure can increase the receptive field of the network,realize the full fusion of multi-scale features,effectively improving the problem of road information interruption or missing detection.(4)In this dissertation,through the research and analysis of the mainstream network model training acceleration methods,it is found that Adam algorithm has excellent performance for tasks with large data scale and multiple network parameters.Adam algorithm is chosen as the training acceleration method of BD-Res UNet model.Finally,the network model proposed in this dissertation is compared with the current excellent image segmentation network model.The experimental results show that the proposed network model has stronger comprehensive performance,which proves the advanced and effective method of improving the network model.