Research On Semantic Segmentation Algorithm Of Remote Sensing Image Based On Transformer

In the semantic segmentation of remote sensing images,due to the complexity of remote sensing image imaging and the significant differences in sample distribution among different ground objects,it is challenging to accurately segment multiple target ground objects with varying proportions under such conditions.At the same time,there are often multiple very similar and dense ground feature categories at the edges of remote sensing images.This impedes the semantic segmentation task of remote sensing images and quickly leads to incorrect pixel-level classification.Therefore,improving the segmentation effect of remote sensing images is still a research frontier and technical difficulty in this field.This paper proposes a multi-attention-guided algorithm based on CBAM(Convolutional Block Attention Module)and Transformer to address these issues.It constructs a remote sensing image semantic segmentation network model CTrans MAtt-Net to complete the semantic segmentation task of remote sensing images.In response to the complex imaging of remote sensing images and significant differences in the distribution of different ground object samples,in order to solve the problem of segmentation difficulty caused by the varying proportion of different ground object types,this network model proposes a method that combines a Multi-headed Self-Attention mechanism and a CBAM to extract features of target ground objects in long and short distances.By using the Multi-headed Self-Attention mechanism,the encoded representation information in different feature subspaces is added,To enhance the expression ability of the network model,enabling it to capture the association of ground feature features based on global contextual information and leveraging the advantages of CBAM,it can enhance the attention ability to key local features at a lightweight cost.In addition,to address the issue of interference from similar features in remote sensing images edges,this network model uses a combination of a Multi-headed Self-Attention mechanism and progressive upsampling to address the issue of interference from similar features at edge details.By leveraging the excellent parallelism of the Multi-headed Self-attention mechanism,the extracted abstract features are correlated and focused with the target feature information,Complete pixel-level classification while accelerating segmentation speed.At the same time,this processing method uses the cross-processing operation of convolution and bilinear interpolation.It uses the inductive bias of convolution as the error compensation of bilinear interpolation to reduce the interference between different objects as much as possible to predict image edges accurately.In addition,CTrans MAtt-Net also utilizes Online Hard Example Mining algorithms to achieve training optimization.Finally,this article evaluated the performance of the network model on Accuracy,Io U,and F1-score metrics.Compared to mainstream network models such as PSPNet,Deep Labv3+,DANet,DNLNet,Point Rend,DPT,and SETR,the CTrans MAtt-Net proposed in this article can achieve high segmentation accuracy on the Vaihingen and Potsdam datasets.