Semantic Segmentation And Road Extraction From Remote Sensing Images Based On Deep Learning

With the rapid development of science and technology,based on remote sensing data,high-performance processors and algorithms can be used to mine a large amount of geographic information.This geographic information can be widely used in resource management environmental protection,mapping,urban planning,land use,etc.As a very popular research topic in the field of computer vision,deep learning is better than traditional algorithms in various scene experiments.Since the image data acquired by remote sensing technology has made great progress in quantity and quality,this provides conditions for the application of deep learning in the field of remote sensing.Therefore,the use of deep learning algorithms to solve remote sensing image semantic segmentation,road extraction,and other issues have become a mainstream solution.(1)In the application of road extraction,the features of large-scale remote sensing image data are large and complex,and the feature extraction using traditional algorithm is not robust enough.This paper proposes a model based on multi-scale network applied in the remote sensing image scene.Achieve road extraction.The multi-scale network is composed of multiple sub-networks with shared parameters to realize parallel training.The attention module is used to realize the combination of parallel networks.The deconvolution used in traditional upsampling is replaced by bilinear difference,and the parameter quantity of the model is reduced.The road saliency detection model and the edge detection model are trained,and the results of the two models are combined by domain transformation,and the post-processing is realized by the full connection condition random field.Through experiments,the performance of the algorithm is greatly improved compared with the traditional algorithm.(2)This paper proposes a neural network based on a spatial pyramid model to semantically segment remote sensing images.The method combines the dilated convolution with the spatial pyramid model to realize the multi-scale information extraction of the image;the training sample is added to solve the problem of less remote sensing data by scaling,rotating,random clipping,etc.of the training sample;using RGB,IRRGB The two images train two models to fuse the results;the multi-scale image joint detection method improves the model accuracy.Through experiments,the algorithm performs well on the ISPRS segmentation dataset.In addition,in the result of semantic segmentation,the IoU of the vehicle class reaches 90%.Therefore,an algorithm for searching the binarized image connected region is proposed to determine the target position of the vehicle and realize vehicle target detection.After training and comparison,the road extraction algorithm based on the multi-scale network has 87.57%IoU on the self-labeled test set.The semantic segmentation based on the pyramid model is 90.48%on the ISPRS test set.