| Remotely sensed image of city planning,land use and construction of geographic information system plays an important role,with the continuous development of deep learning and convolutional neural network,the requirements for accurate and effective remotely sensed image segmentation is becoming more and more intense,the traditional remotely sensed image segmentation is mainly depend on the underlying features and texture features of remotely sensed images,in the image segmentation accuracy is not high and low robustness,which cannot meet the accuracy requirements of remotely sensed image segmentation.Therefore,this paper conducts research on the semantic segmentation algorithm of remotely sensed images.In order to better improve the accuracy of remotely sensed image segmentation,this paper builds three paths based on Res Net18 basic network,and proposes a semantic segmentation model of remotely sensed image based on multi-path feature fusion.Based on this model,Vaihingen,Potsdam,WHU and other public data sets are used to conduct remotely sensed images semantic segmentation algorithm research and experiment;and use the global accuracy(GA),weighted intersection over union(WIOU)and mean intersection over union(MIOU)and other indexes to the overall segmentation effect of the experimental model make an evaluation.The specific research contents are as follows:(1)Experiment and analysis of four classical depth segmentation models.Based on FCN,Seg Net,Deep Lab V3+(Res Net18),Deep Lab V3+(Mobile Net V2)and other four models in Vaihingen,Potsdam,WHU building and other public data sets to verify the effectiveness of the proposed multi-path feature fusion model;(2)Build and experiment a multipath feature fusion model.Build the model based on Res Net18 network and combine spatial information path,semantic information path and attention path;Among them,the spatial information path is a combination of high-resolution and low-resolution features,the spatial pyramid is used for the semantic information path,and the attention path uses channel attention,spatial attention and pyramid attention.The fusion module merges the features extracted from different paths;(3)Ablation experiment and analysis of multi-path feature fusion model.In order to verify the superiority of the model,the proposed multi-path feature fusion experiment was used for semantic information path and attention path ablation experiment;(4)Comparison experiment of semantic information module of multi-path feature fusion model.In further verify the effectiveness of the new model,this paper conducts experiments on the dilation rated of pyramid module to further explore the influence of dilation rated on the new model.The experimental results are as follows:(1)The global accuracy of the proposed model is5.537%,1.428%,1.348% and 2.670% higher than that of FCN,Seg Net,Deep Lab V3+(Res Net18)and Deep Lab V3+(Mobile Net V2)on the three data sets,respectively;Compared with the four classical models,the weighted intersection over union is 8.988%,2.112%,2.539% and 4.294%higher on average.The mean intersection over union was 10.361%,2.457%,1.770% and 5.274%higher than that of the four classical models,respectively;(2)In the ablation experiment,the global accuracy of the proposed model was 0.453%,0.287%,0.393%,1.379%,0.329% and 0.867%higher than that of experiments 1 to 6(ablation experiment).Weighted intersection over union were 1.369%,1.225%,0.936%,3.321%,0.825% and 1.438% higher than those in experiments 1to 6,respectively.The mean intersection over union was 1.205%,0.674%,2.162%,2.901%,2.574%and 2.193% higher than that in experiments 1 to 6,respectively;(3)In the comparison experiment of dilation rated,the segmentation accuracy of the model proposed in this paper is better than that of experiment 7(comparison experiment of dilation rated).The global accuracy,weighted intersection over union and mean intersection over union are 0.183%,0.364% and 1.182% higher than those of experiment 7,respectively.The experimental results showed that:(1)The new model proposed in this paper,the multipath feature fusion model,has a better segmentation effect than the four classical segmentation models: FCN,Seg Net,Deep Lab V3+(Res Net18)and Deep Lab V3+(Mobile Net V2);(2)Ablation experiment,the absence of either semantic information path or attention path in the new model will affect the segmentation effect;(3)In terms of comparison experiment of dilation rated,when the new model only changes the void ratio,the overall segmentation effect of the model with a smaller void ratio is not good,while the new model with a larger void ratio has a better semantic segmentation effect.The segmentation effect based on multi-path feature fusion model proposed in this paper is better than that of classical network,which provides a new choice for semantic segmentation of remotely sensed images. |
