Research On Road Extraction Method For Remote Sensing Images Based On Improved D-LinkNet

Road are the most common objects in remote sensing images.Extracting road information from the remote images is significance for map updating,city planning and car navigation.Due to the diversification of actual road scenes,for example,an urban road images has the characteristics of complex road network and many ground objects,and the rural roads in an images have characteristics of large span and narrow,so that a single algorithm cannot completely and accurately extract the road under various scenes.The hollow convolution module in the central region of D-LinkNet network model expands the sensitivity field and integrates multi-scale features,thus optimizing the road extraction results.However,through experiments,it is found that in different road scenes,the road extraction results of D-LinkNet network model are quite different,and the results still have the phenomenon of missing extraction and wrong detection.Therefore,the roads in remote sensing images are divided into urban roads and rural roads in this dissertation,and an improved D-LinkNet network model is proposed according to the different characteristics of the two kinds of roads in remote sensing images.The specific improvements are as follows:(1)In view of the characteristics of rural roads,such as large span,narrow,large contrast difference and clear edges,this thesis proposes an improved D-LinkNet network model.Firstly,the hollow space pyramid pooling module DA-ASPP with improved double attention mechanism is added to the hollow space pyramid pooling module ASPP.DA-ASPP can apply the dual attention mechanism to the features of different scales respectively to achieve multi-scale feature extraction and the effect of intensive feature sampling.Then,the DA-ASPP module is added into the central area of D-LinkNet network model to get the improved D-LinkNet network model,so as to increase the accuracy of road information extraction at different scales by a network model.Ablation experiments show that the accuracy of rural road extraction model is significantly improved.Through the algorithm comparison,the overall performance of the network model in this study is much better than the existing mainstream network models.(2)The improved D-LinkNet network model is proposed in this dissertation in view of the complex urban road network,many ground features,and the spectral characteristics of buildings and roads beside main roads are more similar.Firstly,the attention-gate AG is added before the jump connection in the D-LinkNet network model,the superficial information of the attention-gate is selectively integrated into the deep network,thus suppressing some irrelevant feature responses in the shallow feature map,and at the same time,it adds a little extra computation,it not only solves the redundancy problem of the network model jumping connection but also greatly increases the accuracy of extracting target region from network model.Secondly,the decoder is improved,the improved decoder structure consists of transposed convolution and upper sampling layer to avoid the checkerboard effect,in order to make the road extraction results more accurate,this research also uses the joint loss function of Focalloss and Diceloss to replace the loss function of D-LinkNet,so as to promote the accuracy of calculation errors in the training process.It can be seen from the results of ablation and comparison experiments that the accuracy and connectivity of the proposed network model are high,which further proves the overall effectiveness of the new network model.