Semantic Segmentation Algorithm Based On Domain Adaptation And Generative Adversarial Network

With the progress of computer technology,especially the development of deep learning technology,digital image processing and computer vision technology have gradually formed their own system.The semantic segmentation of image has attracted more and more attention because it can be used to assist other computer vision tasks.Image semantic segmentation classifies the image pixel by pixel and assigns semantic labels to all pixels in the image,so as to understand the image content from the pixel level.It has broad application prospects in the fields of automatic driving,object detection and tracking,human-computer interaction,scene recognition and understanding,etc.This paper summarizes the existing research algorithms in the field of semantic segmentation,and deeply realizes that the performance of traditional deep learning-based image semantic segmentation methods depends on the quality of the data set on the one hand.However,under normal circumstances,the acquisition of high-quality pixel-level annotation data is very expensive and time-consuming.On the other hand,in practical applications,the network trained under a specific data set usually performs well only under this data set.When used in similar data sets,the performance will decrease significantly,which limits the semantic segmentation methods in other Application in real-world situations.This paper aims to improve the accuracy of semantic segmentation of remote sensing images.The specific work is as follows:(1)This paper proposes a semantic segmentation algorithm based on improved U-Net network to realize remote sensing road semantic segmentation.Traditional remote sensing image road semantic segmentation methods lack the extraction of multi-level road features,and do not consider the geometric topological structure of the road as a whole,which makes the structure of the final road segmentation result incomplete;at the same time,in the process of image down-sampling feature extraction,the resolution of the image continues to decrease,but the lost spatial information is difficult to recover,making the extracted road edges unclear.In order to solve such problems,this paper proposes an improved U-Net network model that combines context information and attention mechanism.The model extracts and integrates the image context information of the image through context information extraction module,ensure the extraction of the geometric topological structure features of the road;At the same time,the attention mechanism is used to adjust the weights of the underlying features passed by the encoder skip connection to improve the recovery of the missing position information of the segmentation network,thereby improving the segmentation accuracy of the road edge area.In addition,the algorithm was verified on the public data set Deep Globe road extraction data set,which further proved the value of adding an attention mechanism to the codec network.(2)This paper proposes a comprehensive unsupervised remote sensing image semantic segmentation algorithm that combines domain adaptive methods and generative adversarial networks.In the collection of remote sensing images,different satellite photography equipment,light intensity,seasons and other conditions will affect the data set.The difference in image resolution causes the most obvious interference to the semantic segmentation of remote sensing images among them.As a result,the performance of semantic segmentation of remote sensing images under different training sets is degraded.In response to the above problems,this paper introduces the super-resolution method into the semantic segmentation network,and assists the feature extraction and information recovery of the semantic segmentation network by integrating the features of different resolutions extracted during the image super-resolution process,thereby reducing the domain shift of the data set caused by the image resolution;The domain adaptive method combined generative adversarial networks eliminates the difference between the source domain and the target domain,and finally realizes the unsupervised semantic segmentation method under the target domain data set.(3)Verify the unsupervised semantic segmentation model proposed in this paper in the remote sensing road segmentation dataset,as well as the traditional feature-based domain adaptive model,the output space-based domain adaptive model,and the image-based domain adaptive model for road segmentation effect.The experimental results show that the superresolution adversarial domain adaptive network proposed in this paper has better segmentation performance in unsupervised remote sensing semantic segmentation.It further proves the auxiliary effect of the image super-resolution method on the semantic segmentation network of the image.