Research On Semantic Segmentation Of Remote Sensing Image Based On DeepLabv3+

Semantic segmentation has received extensive research and attention in the field of computer vision at present.It uses pixels to distinguish objects in images to automatically identify objects.Semantic segmentation of remote sensing images is the difficulty in semantic image segmentation.In remote sensing,Existing diversified feature information and objects with different status quo have caused many difficulties for segmentation.Remote sensing usually refers to the detection of long-distance targets through the sensors on the aircraft,and then the specific categories can be identified at the same time through semantic segmentation to achieve semantic segmentation.On the premise of deep convolutional neural network,a new network innovation scheme is proposed based on Deeplabv3+ network.The main work of this paper is as follows:(1)Semantic segmentation of high-resolution remote sensing image data sets: Through the processing of public data sets,a data set with a large amount of information is obtained.At the same time,the related process of establishing remote sensing image semantic segmentation data set is explained in detail,including picture cutting,image preprocessing,data set division and so on.(2)A small-scale data set is used to compare several classic networks used in this article,and a method is proposed for the problem that remote sensing images are often large in size and need to be spliced.Compared with direct stitching,the accuracy is improved by 2-3%.(3)Deal with the problem that the remote sensing image semantic segmentation accuracy is poor in processing small objects: Through the analysis of the image semantic segmentation network of Deeplabv3+,this paper proposes a semantic segmentation network based on Deeplabv3+,and several improvements have been made to the network:(a).Use the self-attention mechanism to improve the model,so that the picture can be distinguished from the priority during the segmentation process,and know which part of the picture is the most important;(b).Using the convolutional split method makes the network have more convolution Layers can increase the depth and width of convolution and improve feature acquisition capabilities.The research results show that the optimized model has an accuracy of 95.14% on the Inria public data set,which is an increase of 8.85% relative to the accuracy of FCN,and an increase of 0.46% for Deeplabv3+.At the same time,for objects with fewer categories in the classification target,the IOU increased by 3.8%.