Research On Remote Sensing Image Segmentation Method Based On Deep Neural Network Adversarial Learning

Image segmentation technology divides the image into different types of regions according to the internal characteristics of the image,such as edges,textures,and colors.Each region belongs to the same category and has consistency or similarity.Remote sensing image target segmentation is the segmentation of images into regions with actual semantic information.It is one of the important applications in image processing.Through remote sensing image processing,objects such as buildings,roads,coastlines,and vegetation can be extracted for urban construction,agricultural development,disaster monitoring,and other applications that provide guidance,which has high application value.Remote sensing data has the characteristics of large scale,dense targets and different scales,complex lighting environment,annotating images requires a lot of time and labor costs,remote sensing image target segmentation still faces many challenges.In allusion to complex remote sensing image targets,the current image segmentation methods based on mainstream deep neural networks have repeated convolution pooling operations,resulting in the segmentation of targets that have problems such as missing small targets,poor target pixel continuity,and missing edge details.And in the case of large differences in data distribution and lack of corresponding labels,a good segmentation result cannot be obtained.Based on the above analysis,on the basis of the existing deep neural network,this thesis explores to solve the problems of the mainstream deep segmentation model in the field of remote sensing image segmentation due to the loss of small targets,the loss of edge details,and the distribution difference between different data.The specific research content is as follows:(1)For small targets can't be completely classified,occluded target unable to be extracted and details missing in the current deep convolution network,a remote sensing image segmentation method based on multi-level channel attention is proposed on the basis of deep convolution coding-decoding network.This method first adds channel attention mechanism to the network coding stage,and obtains more effective features through self-learning to solve the problem of target occlusion in remote sensing images.Secondly,feature map fusion of channel attention is applied on different scales to extract abundant context information and deal with target scale changes.This model improves the problem of small target segmentation and improves the performance of segmentation.The experiment results show that the proposed model has higher accuracy of target segmentation and better segmentation results on small targets and occluded targets.Good results have been achieved in target segmentation of remote sensing images with limited training data,complex and diverse backgrounds and large scale changes.It shows that the model can be applied to target segmentation of complex remote sensing images.(2)In view of the insufficient retention of small target information and the lack of edge information in the previous part of the work,it is not only necessary to consider the segmentation map can contain more feature information,but also hope to improve the continuity of the long-distance space to obtain a smooth edge segmentation map,this thesis proposes a conditional generative adversarial network with multilevel channel attention mechanism to segment remote objects.The model includes a generating network G and a discriminant network D.G is a segmentation model with a multi-level channel attention mechanism.It is the same as the segmentation model in the first part.This model constructs a channel attention mechanism through self-study to deal with the segmentation of buildings of different sizes,especially for small building information.The segmentation result produced by G is used as the input of D.Since G incorporates the channel attention features containing multi-level jump connections,G has stronger feature extraction capabilities at all scales,and on this basis,G and D The adversarial training for the remote sensing data is more accurate modeling of the distribution of remote sensing data.Compared with other comparison methods,it pays more attention to the spatial consistency and continuity of the data.Experiments show that it can bring smoother and more complete edge details to the segmentation result map,which improves model segmentation performance.(3)Given the two problems currently faced by remote sensing image segmentation,one is that when a well-trained model on one data set is transferred to another data set,due to the large distribution difference between the data domains,accurate segmentation cannot be achieved.The second is that when the two domains are not only different between domains,but also when the image of one of the domains has no label data,it is impossible to directly perform supervised training.Aiming at the above two problems,a domain adaptive remote sensing image segmentation method combined with adversarial learning is proposed.When the labeled source domain data and the unlabeled target domain data are sent to the segmentation network,the discriminant network is added for adversarial learning,minimize the difference in prediction results between the two domains.To further enhance the domain adaptation model,a two-layer domain adaptation network is constructed to deal with the problem of domain differences in complex remote sensing images.Experiments show that when the data of the source domain and the target domain are quite different,and the label of the target domain is not used,it still has a good segmentation performance.