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

Satellite remote sensing is an important means for people to observe the surface conditions.With the continuous improvement of my country's high-resolution satellite observation capabilities,the amount of high-resolution remote sensing image data has increased significantly,and how to accurately and quickly extract objects such as surface buildings has become a research hotspot.In recent years,deep learning,especially the great success of deep convolutional neural networks in the field of natural image processing,has also prompted its expansion to the analysis and processing of remote sensing images,and promoted high-resolution remote sensing images including semantic segmentation.Development of related tasks.This article mainly focuses on the method of semantic segmentation of remote sensing images based on convolutional neural network.The main research contents are as follows:In order to solve the problems of low efficiency and insufficient segmentation accuracy of traditional remote sensing image segmentation methods in complex scenes.UNET with a codec structure is used as the image segmentation model.UNET extracts the high-level semantics in the image by down-sampling and up-sampling the input image,so that the model has a good segmentation effect.However,the effect of edge segmentation for smaller objects contained in the image is not good.In response to this problem,a model structure based on the combination of Feature Pyramid Networks(FPN)and UNET is proposed to improve the ability of the image segmentation model to integrate multi-scale information,and at the same time replace the loss function with a better capture The boundary label relaxation loss function of the target edge improves the segmentation effect of the UNET model on the target boundary.Experimental results show that the improved method effectively improves the accuracy of semantic segmentation,and better alleviates the problems of poor segmentation of small-scale targets and inaccurate segmentation of target edges.Semantic segmentation is not only to achieve pixel-level classification,but also to achieve the positioning of foreground objects.Due to the inherent translation invariance and insensitivity to space of convolutional neural networks,it is difficult for image segmentation networks to locate foreground objects.In response to this problem,this article adds an attention mechanism module to the improved UNET network model.The CBAM module learns the channel attention and spatial attention of the feature map respectively,and the channel attention can emphasize For the important details of each feature map,spatial attention can better help the network locate the target location.By introducing the attention mechanism module,UNET achieves a more precise and accurate segmentation effect.At the same time,for irregular buildings and objects appearing in the image,the model also adds a Deformable Conv Network(DCN)to the back-end feature extraction network.DCN adds an offset to each convolution kernel.To improve the model's ability to model the deformation of irregular objects.By using the above methods to construct the image segmentation model,and experiment on the remote sensing image data set,the results show that the method used in this paper effectively improves the accuracy of semantic segmentation,and better alleviates the problem of small-scale target edge segmentation.Good question.The target edge segmentation can be more accurate,and the segmentation accuracy of the deformed object is also better improved,so that the MIOU of the model reaches 0.87,the MPA reaches 0.95,and the loss is reduced to 0.15.