Research On Urban Road Segmentation Algorithm Based On Deeplabv3+ Remote Sensing Image

Semantic image segmentation technology has attracted more attention and become a hot research topic in the field of computer vision.Its research results can be effectively applied to map reconstruction,face recognition and unmanned driving.At present,there are many existing semantic segmentation networks.With the continuous development and improvement of semantic segmentation technology,the segmentation accuracy on public datasets is also increasing.With the increase of remote sensing image acquisition methods,the demand is increasing.How to effectively improve the segmentation accuracy of high-resolution remote sensing urban road image is the main problem in this paper.After studying the difficulty of classification accuracy caused by the complexity,morphological diversity and texture diversity of road types in remote sensing urban road images,this paper analyzes and studies the feasibility of existing segmentation algorithms.The interactive image segmentation technique of the algorithm is applied to the segmentation process of visible light image and remote sensing urban road image.The experimental results show that the segmentation accuracy of visible light images is much higher than that of remote sensing images because of the high resolution and complex road features of remote sensing urban road images.A lot of texture and detail information interfere with the segmentation.In order to effectively improve the segmentation accuracy of high-resolution remote sensing urban road image,this paper further proposes the segmentation method based on Deeplabv3+ network model to segment it.Firstly,the semantic segmentation experiment is performed by using Deeplabv3+ network based on BN(Batch Normalization)layer.Then the network structure is improved,and the semantic segmentation experiment based on the Deeplabv3+ network under the GN(Group Normalization)layer is used.The experimental results show that the segmentation result based on Deeplabv3+ network model is obviously better than the interactive segmentation technology based on graph cut algorithm.Compared with Deeplabv3+ network using BN layer,deeplabv3+ network using GN layer has higher segmentation precision and maximum after improvement.The advantage is that the segmentation accuracy is not affected by the size of the batchsize.