Research On UNet Based Remote Sensing Image Segmentation Algorithm

Remote sensing images are films or photographs obtained by remote sensing satellites through imaging techniques.The quantity and quality of high-resolution remote sensing images are enhanced compared to the original images.Remote sensing images contain a large amount of image information and are widely used in various fields,such as urban planning,land use and crop cultivation.Image segmentation is the technique of extracting each specific region or object from the image.Remote sensing image semantic segmentation is an application direction of image semantic segmentation,which is a key technology in remote sensing image processing.In the traditional image segmentation algorithm,the efficiency of manual marking data is low,and the training process needs complex parameter adjustment operation,which leads to the accuracy of the network model is not high.Great progress in image semantic segmentation tasks is due to the rapid rise of deep learning.Using deep convolutional neural networks for image feature extraction not only reduces training time but also effectively improves the performance of semantic segmentation tasks.This paper mainly studies the application of deep learning method in semantic segmentation of high-resolution remote sensing images,and selects U-Net model as the main research object through comparative experiments.This model was initially used for medical image segmentation,with a simple network structure and excellent performance.However,the accuracy of feature segmentation for extracting complex remote sensing images only reached 83.03%.Therefore,the research in this paper is mainly focused on the following points.(1)The comparative analysis of remote sensing image segmentation methods shows that the performance of U-Net is relatively advantageous.This paper further analyzes the problems of U-Net network applied to remote sensing image semantic segmentation,proposes a DUNet(Deep U-Net,DUNet)network structure based on U-Net network,and conducts experiments on the improved DUNet network,and the experimental results under the same dataset showed that DUNet improved the accuracy by 1.04% and m Io U(mean Intersection over Union,m Io U)by 2.19% over U-Net.(2)Remote sensing images contain multiple types of small target objects with an unbalanced distribution of foreground and background information,which makes U-Net segmentation slow and difficult to obtain object features.In this paper,we propose and build a UNet-RS(U-Net for remote sensing images,UNet-RS)network for semantic segmentation of remote sensing images based on depth-separable convolution and attention mechanisms.In order to reduce the impact of parameter increases on network performance,an improved depth-separable convolution module is presented,and it is experimentally demonstrated that the computational effort of replacing the original convolutional structure of U-Net is reduced by a factor of 1.To allow the network to focus on more important information about the image during training,hyperparametric improvement of the attention mechanism is proposed to improve the performance of semantic segmentation.With the same dataset,UNet-RS improves the m Io U by 4.11%compared to U-Net.(3)Remote sensing images contain rich semantic information at different levels,and in order to obtain multi-level semantic features,a W-Net network using dual paths for semantic segmentation of remote sensing images is built.The W-Net network in this paper uses improved FCN(Fully Convolutional Network,FCN)and DUNet to form a dual-path convolution module,and the attention mechanism is combined with the classification structure to form a focused salient region module.After passing the dual path,a multiscale fused feature image that incorporates a global perceptual field,as well as small target object detail information,is obtained.By focusing attention on the features containing multi-scale information,we can focus more on the key information in the image and suppress the influence of irrelevant information on feature extraction,and the number of layers in the network is not too deep,which can effectively identify the same object of different sizes and segment the edges smoothly.In the test experiments,W-Net outperformed the rest of the models with 63.24% m Io U.In the generalizability experiments,W-Ne also outperformed the present state-of-the-art model FarSeg.