Remote Sensing Image Segmentation Based On Fuzzy Conditional Random Field Deep Learning

The visual effect of high-resolution remote sensing images is intuitive and contains rich geometric information about ground objects,which has been widely used in environmental monitoring,urban construction,map drawing,resource management,and other fields.However,due to the high-resolution remote sensing image in the imaging process is vulnerable to noise,atmospheric radiation,and many other factors,often exist "the same object with a different spectrum" and " a foreign object in the same spectrum" the uncertainty of the phenomenon,and the types and spatial information of high-resolution remote sensing images are complex and diverse.This poses a serious challenge to the task of feature information extraction.Therefore,how to solve the uncertainty problem in remote sensing images,giving full consideration to the rich context information in the image,and accurately segmenting high-resolution remote sensing images has been an important problem to be solved for a long time.In recent years,deep learning technology has made great progress.Because of its advantages of intelligence,deep learning shows a more perfect segmentation effect than conventional image segmentation methods in the segmentation of multiple different scenes.Therefore,this paper mainly studies semantic segmentation in high-resolution remote sensing images based on deep learning method.The main work is summarized as follows:(1)Due to the uncertainty of high-resolution remote sensing images affected by noise and the strong correlation between the context information in the images,the feature extraction of high-resolution remote sensing images using traditional image segmentation methods lacks robustness.Therefore,a new collaborative neural network model based on fuzzy deep learning conditional random field is proposed in this paper.To solve the fine segmentation problem of high-resolution remote sensing image.Firstly,a fuzzy segmentation network model is designed to obtain effective feature information,and fuzzy logic unit is introduced to deal with fuzziness and uncertainty of high-resolution remote sensing images.In addition,in order to learn geospatial scene representation,we explore a conditional random field model that integrates spectrum and spatial context to reduce the effect of spectral changes and highlight the details of the information in the image.In the conditional random field model,we encode the spectral context and the spatial context using the unary and pairwise potentials respectively.Finally,the effectiveness and superiority of the proposed model are verified on three data sets.The experimental results show that the proposed model has better subdivision and generalization performance compared with the classification results of relevant existing models.(2)In order to further explore remote sensing image segmentation based on deep learning and solve the problems faced by high-resolution remote sensing image segmentation,a fuzzy convolutional neural network based on multi-scale features is proposed.Firstly,a fuzzy learning module is designed to solve the uncertainty problem of high-resolution remote sensing image segmentation under the influence of noise.The membership degree in Gaussian form is used to constrain multiple parameter descriptions,and corresponding fuzzy semantic markers are assigned to each feature point,which are integrated into the skip connection part of the network.In addition,considering the lack of shallow network features and incomplete acquisition of context information,the improved the atrous spatial pyramid pooling structure is aggregated in the network model to expand the receptive field and make full use of spatial context information to capture small target features.Secondly,a complete residual module is used to replace the original residual module to reduce the risk of network gradient disappearance.At the same time,in order to enhance the learning process of segmented objects,a new loss function is improved and proposed according to focus loss and classification of small and medium targets in highresolution remote sensing images.Finally,the conditional random field model was used to process the detailed information.Experimental results show that this model can significantly improve the segmentation performance of high-resolution remote sensing images.