Research On Water Remote Sensing Image Segmentation Algorithm Based On Improved UNet

In recent years,computer hardware and deep learning algorithms have made great progress.From face recognition to environmental monitoring,deep learning which made human society and ecological civilization promoted has entered people’s daily life.In order to alleviate conflict between human development and water resources,research proposes two improvement schemes for UNet model:(1)MSFM-UNet water body segmentation model.The improvement measures are as follows:First,part of the traditional convolution in encoding network and decoding network is replaced by dilated convolution to enlarge the receptive field.Second,transposed convolution is used for upsampling to replace traditional linear interpolation.Third,a multi-scale fusion module is designed by using the Phantom convolution and SENet modules.This module is applied to the output end of network to process and integrate complex ground object information and improve the segmentation accuracy of the network.(2)PM-UNet water body segmentation model.The improvement measures are as follows:First,traditional convolution in the backbone is replaced by dilated convolution to enlarge the receptive field.Second,transposed convolution is used for upsampling.Third,traditional convolution and expansive convolution are cascaded,and then a multi-scale convolution module is built to replace some traditional convolution modules in the decoding network by using the design idea of the residual network.Fourth,a parallel residual attention mechanism is designed using modules such as asymmetric convolution and pooling operation.This module is utilized in output end of network to enhance water feature information and suppress interference of background information.The experimental results show that the accuracy of the model is improved.Among them,the mean intersection over union,mean pixel accuracy and overall accuracy of MSFM-UNet reach 96.41%,98.35%and 99.35%respectively,the indexes of PM-UNet also reach 96.23%,98.30%and 99.33%respectively,all of which are superior to other comparison algorithms.Therefore,the improved algorithm can be well applied to water body segmentation task.