Using Improved DeepLabV3+to Extract Ground Road from Remote Sensing High-Resolution Image
In view of the problems that the current image segmentation algorithm has low accuracy and serious fractures when implementing road extraction from remote sensing images,a model R-DeepLabV3+is proposed based on the DeepLabV3+model and is suitable for remote sensing road extraction.In the feature extraction network part of the encoder,the focus down sampling layer is used instead of the standard convolution kernel to compress the in-put feature map without feature loss.At the same time,a dense connection mechanism is introduced to replace the residual connection to enrich the detailed feature content in the deep feature map.At the end of the encoder,a fast feature pyramid pooling-cross-order local network with depth-separable convolution is used to implement the fusion of multi-receptive field feature map.In the decoder part,a coordinate attention mechanism is introduced to calculate the distance between the road and the background Feature weights to improve model learning efficiency.Experimental results show that this method has significant advantages in extracting a complete road structure,Mean Intersection over Union(MIoU)and Accuracy(Acc),and can implement high-precision road extraction based on remote sensing images.
NETL
NSTL
万方数据
