Abstract: A novel Swin Transformer semantic segmentation optimization algorithm with a codec structure is proposed to solve the problems such as difficulty in identifying the edge of carbon sliding plate by semantic segmentation model,large interference in complex background,and serious feature information loss. Firstly,the backbone network adopts U-shaped codec structure to realize multi-scale information fusion.Secondly,the attention local enhancement module is added to expand the sensing field and improve the model generalization ability. Then,the upsampling structure with data correlation is used to enhance the quality of upsampling,eliminate the impact of resolution on prediction results,and improve image reconstruction capability. Finally,the skip connection is replaced by the residual path to make the semantic information in the codec structure more closely connected and improve the training efficiency. The experimental results show that the Swin Transformer semantic segmentation algorithm improves the measurement prediction accuracy by 3.63 %,and the average accuracy of pixel classification in all categories is improved by 7.29 %. The research results confirm the superiority and robustness of the Swin Transformer semantic segmentation model in identifying and handling carbon slide tasks.

Key words: carbon sliding plate, Swin Transformer, local enhanced sensing, image reconstruction, semantic segmentation