Research On The Improved Method Of Semantic Segmentation Of Aerial Remote Sensing Image Based On Fully Convolutional Networks

The processing of high-resolution remote sensing images has a great impact in the military and commercial fields.The improvement of the resol ution and resolution of remote sensing images has driven the development of various related industries.Such aspects as earthquake disaster relief,glacial melting assessment,transportation and navigation have great demands on the segmentation and extraction of remote sensing images.Automatically identifying building targets from remote sensing images and obtaining accurate edge contour information has been an important endeavor of digital mapping to achieve automation and intelligence.Despite this,the degree of automation of the analysis of satellite remote sensing images is still very low,so in recent years,deep learning methods have been fully developed in the field of remote sensing image segmentation due to their high degree of intelligence.Based on the task of semantic segmentation of high-resolution remote sensing images,this paper studies the semantic segmentation methods of aerial remote sensing images,uses deep neural network feature learning and powerful representation capabilities to extract the features of the images,and at the same time targets multi-scale aerial remote sensing images The characteristics of multi-angle,multi-angle,complex background,small target,etc.realize the optimization and improvement of existing model methods to obtain higher segmentation accuracy.The main research content of this article includes the following aspects.First of all,the development history of image segmentation and the application of deep learning in the field of image segmentation are compr ehensively summarized.The current status of remote sensing image segmentation and the theoretical basis of general segmentation methods are discussed.The composition of convolutional neural networks is introduced.Elements.Next,four different mainstream fully convolutional network structures are discussed,and four model methods are constructed on high-resolution aerial remote sensing images.Certain image preprocessing operations are performed on aerial remote sensing images to segment the quality of different methods.Evaluation and comparison with model performance verify the effectiveness of the constructed model on remote sensing images.Then,based on the existing model,the segmentation method is improved and optimized according to the characteristics of the remote sensing image,based on the selected basic model framework,analysis,design,and implementation of model performance optimization.The improvement includes the following aspects: first,image enhancement and preprocessing of the initial image for the multi-directionality of the remote sensing image,and secondly,an improved multi-scale fusion technique is proposed for the multi-scale nature of aerial remote sensing image to improve the model’s ability to different scales The ability to express the target is compared and discussed with the existing multi-scale fusion technology.Based on this,a set of improved methods for extracting high-level semantic information from the encoder output are proposed.Then for the shortcomings of higher resolution and slower training speed of remote sensing images,this paper introduces batch normalization technology.The goal is to further improve the training convergence speed and generalization ability of the proposed network through normalization technology.The loss function function was improved,and an auxiliary loss function was introduced to increase the training speed of the network.Finally,for the characteristics of aerial remote sensing image with complex background and many small targets,an attention mechanism is introduced.The characteristics of the current attention mechanism are discussed.Based on this,the attention learning mechanism is improved.Increase the network’s attention to the target building.Finally,this paper compares the existing mainstream segmentation model with the improved method proposed in this paper,and then evaluates the segmentation performance of different methods.At the same time,it compares with the improved segmentation algorithm of aerial images that also performs well in the same field.The advantages and disadvantages of the proposed method verify the effectiveness of the improved method.