Research On High Resolution Remote Sensing Road Image Extraction Technology Based On Deep Learning

In recent years,roads as an important part of the ground information,to obtain accurate remote sensing road images is of great value.In the road detection of high-resolution telemetry road images,the traditional remote detection road image extraction method obtains poor road integrity due to the complexity of the remote sensing images.Deep learning in latest years has powerful feature extraction capabilities and can be effective in obtaining road detection results with the high accuracy in common situations,but further research is needed for complex urban roads and small discontinuous road detection.Thus,this dissertation constructs a highresolution remote sensor road extraction model based on a deep learning approach for extracting road information from remote sensor imaging.The specific work is studied as follows:(1)A remote sensing road extraction algorithm based on an improved U-Net network is proposed to address the problem that the road network in a city context is easily obscured by buildings and green belts,resulting in discontinuity and low accuracy of the extracted roads.The algorithm consists of three parts: the encoder,the bridging network and the decoder.Firstly,in the encoder,a modified Res Net34 network combining the advantages of DO-Convolution and Leaky Re LU activation functions to increase the ability to extract global features from images and to speed up the training process of the network.Secondly,a context acquisition module is designed to bridge the encoder and decoder,reducing the loss of key feature values.Next,the strengths of the CBAM attention mechanism are incorporated in the decoder to focus more comprehensively on the information of the road and refine the effect of road edge extraction.Finally,the experiments on the Deep Globe dataset indicate that the proposed method compares well with mainstream network models such as U-Net and EDRNet in terms of performance,extracts road information much better,raises the precision of road segmentation in an urban context,and verifies the usefulness of this method.(2)A dual-decoder DSU-Net network for remote sensing road detection is suggested to solve the issues of the small roads with discontinuities and the geometric similarities between roads and their surroundings,which leads to road omission and misclassification.First of all,in the encoder,the improvement of Res Net50 combines the advantages of the SEnet attention mechanism to focus more on the channel dimension of the image in the image feature extraction stage,and to obtain the key features of the road image based on the correlation of each feature channel dimension,for improving the feature extraction capability of the encoding network.Then,a multi-scale convolutional extraction module is designed in the bridging network to expand the image receptive field and to decrease the loss of feature information.Further,a dual decoder is engineered to extract information in different dimensions.The small decoder acquires low-dimensional feature information,and the large decoder makes integration of the high-dimensional feature information passed from the multi-scale convolution module and integrates the feature information from the small decoder to promote the accuracy of road classification and refine the effect of road classification.In conclusion,experiments on the Massachusetts road dataset and the Deep Globe road dataset demonstrate that the algorithm network model outperforms the U-Net network model in every index,and the proposed method addresses the problem of extracting road discontinuities caused by small-sized road edge occlusion and the misclassification problem caused by the geometric similarity between the road and its surroundings in complex situations,verifying the efficient and universal applicability of the algorithm.