Research On Semantic Segmentation Of Remote Sensing Image Based On Convolutional Neural Network

With the rapid development of modern remote sensing technology in China,the exploration ability of remote sensing satellites has grown,the traditional remote sensing image segmentation method which is unable to meet accurate real-time automatic extraction of remote sensing image data.Semantic segmentation methods of natural images have advanced significantly in recent years,thanks to the widespread use of deep learning technologies in the field of computer vision.Because the boundaries between various items in remote sensing photos are easily obscured by the huge number of tiny and medium scale objects,existing natural image semantic segmentation methods cannot be directly applied to remote sensing images.As a result,this study investigates the semantic segmentation model of remote sensing photos,taking into account the peculiarities of remote sensing images as well as the semantic segmentation approach of natural images.The following is the main research project:(1)Aiming at the problems of edge confusion caused by multiple objects gathering in remote sensing images,unclear segmentation of small scale objects,and insufficient acquisition of global information in semantic segmentation process,a remote sensing image semantic segmentation algorithm DU-net based on mixed attention and full-scale skip connection network was proposed.In this algorithm,U-net3+is used as the basic network,and full-scale layer-hopping network is used as the feature extraction network.The depth supervision in the original algorithm is abandoned,the association between feature and attention mechanism is established,and finally the process of semantic segmentation is realized.The experimental results show that the DU-net algorithm has a significant improvement over the classical algorithm under different indexes,and improves the image edge segmentation quality and the accuracy of the algorithm for the segmentation of small scale targets.(2)A lightweight semantic segmentation algorithm for remote sensing images based on interaction between global and local features,EFLG-Net,was proposed in response to problems found in most current semantic segmentation models of remote sensing images,such as slow training speed,many network layers,and a large number of parameters.The feature extraction network was EfficientNetBO,and the algorithm introduced the global feature path,established the connection between global and local features,improved the convolution module MBConv in the original algorithm,proposed a new module FU-MBconv,optimized the network structure and parameters,and then connected to the global feature path through deconvolution operations.Finally,the semantic segmentation procedure is completed.The EFLG-Net technique improves model parameter size,training time,and model correctness,according to experimental data.