Research On Semantic Segmentation Of Remote Sensing Image Combine Superpixel And Transformer

As an important research direction in the field of remote sensing image processing.Semantic segmentation of remote sensing images has been widely used in land resource management,military target recognition and other fields.At present,the semantic segmentation model based on traditional neural network can’t extract the features of small objects in remote sensing images with higher dimensions,which leads to higher segmentation error rate and lower segmentation accuracy.In view of the above problems,a method combining superpixel and transformer model is proposed to segment remote sensing images.Firstly,the parameter grid search method is used to arrange and combine each optional parameter.The evaluation method with minimizing segmentation inaccuracy as the core is used to transform.The parameter search objective into finding the parameter combination corresponding to the lowest expected risk in different segmentation results.So as select the optimal parameters needed by SLIC superpixel algorithm.Secondly,the original remote sensing images are clustered by superpixel segmentation to obtain initial patches.And the results of superpixel segmentation are vector embedded by using full convolution network,which is input into the improved transformer model suitable for semantic segmentation.Finally,using the characteristics of the sequence model,we can input a range of region vector sequences,output the region label sequences in the same order as the input sequence.And the transform semantic segmentation tasks into classification tasks.In this process,the self-attention mechanism is used to realize contingency filtering.Compared with pure convolution network,the attention mechanism can pay more attention to the inherent feature relations and topological relations of regions,which can improve the segmentation accuracy.Experimental results show that 3 accuracy of the algorithm is 0.843,0.882,0.818,The overall accuracy is 0.849.Compared with U-net and FCN networks,it has increased by 2% and 8.7% respectively.Experiments shows that the superpixel and transformer model is superior to the other networks on segmentation accuracy.It shows the potential of this method in high-resolution remote sensing image segmentation.This paper has 36 figures,8 tables and 59 references...