Research On Several Types Of Remote Sensing Image Semantic Segmentation Methods Based On Transforme

Semantic segmentation of remote sensing images is an important component of remote sensing image processing,widely used in fields such as urban planning,military target detection,and geological disaster assessment.With the development of remote sensing acquisition technology,high-resolution remote sensing images have emerged,which brings rich texture information of ground objects and also makes the background of ground objects more complex.The problem of identical ground object categories but different external features,as well as different categories but highly similar external features,can occur in images,making it difficult for traditional image segmentation methods to effectively segment based on manually designed methods.Deep learning can adaptively learn image features at different levels and effectively solve this problem.The Transformer model can capture remote dependencies and has good performance in the field of image segmentation.Therefore,this article mainly studies several Transformer based semantic segmentation methods for remote sensing images,and the main work is as follows:In response to the problem of insufficient global contextual information in convolutional networks,this thesis proposes a Class Attention Network(CANet),which extracts contextual dependency information from class,spatial,and channel levels through class attention module,kernel attention module,and channel attention module.The proposed class attention improves the segmentation effect by improving the feature representation of category information on the channel,enhancing its ability to perceive and differentiate internal information categories.CANet achieved optimal and suboptimal results on the Potsdam and Vaihingen remote sensing datasets of ISPRS,and validated the effectiveness of class attention in ablation experiments.In the field of remote sensing image segmentation,there is little research on the application of Transformer’s image semantic segmentation methods.In this thesis,several Transformer based segmentation methods are mainly Transfer learning to remote sensing images.These methods have been tested on the two remote sensing datasets mentioned above and achieved good segmentation results overall,making them suitable for remote sensing segmentation tasks.Compared with convolutional segmentation networks and CANet,the segmentation performance indicators are superior to convolutional networks,but their extraction and fusion of local information in segmentation performance are insufficient.Transformer has shown great potential in remote sensing image segmentation,and future research can draw on the design ideas of Seg Former,Swin,and Twins to achieve better segmentation results.