Structural feature optimization method for road extraction
In previous research on remote sensing road extraction,the road structure characteristics of the entire input image were often overlooked,making it difficult to generate a complete road network for complex and multi-road areas.First,a strip pool-ing module(SPM)was designed to effectively expand the backbone's receptive field.Specifically,SPM consists of two paths,focus-ing on encoding long-distance context along the horizontal or vertical spatial dimensions and encoding global horizontal and vertical information at each spatial position of the aggregated feature maps.This enhances the ability to capture long-distance spatial depen-dencies and utilize inter-channel dependencies.Considering the diversity of road scales,a cascade multi-scale attention enhance-ment(CMSAE)module was proposed,which uses spatial attention residual blocks on multi-scale features to capture long-distance dependencies and introduces channel attention layers to optimize multi-scale feature fusion.The goal is to address the issues of dis-continuous road extraction and jagged boundary recognition present in existing methods and aggregate spatial details and semantic information for continuous roads.Experimental validation showed that,compared to various algorithms,the proposed method im-proved Precision,Recall,IoU,and ACC by 3.05,2.12,3.43,and 1.85 precentage,respectively,outperforming the comparative algo-rithms.This demonstrates the effectiveness of this method in addressing the problem of discontinuity in road extraction tasks.
NETL
NSTL
万方数据
