To solve the problem that multi-scale information cannot be effectively aggregated and used and the boundaries are blurred in Segformer network,a semantic segmentation network based on cascaded attention and boundary prediction is proposed.In the decoder part of Segformer,by using cascaded attention,multi-scale feature information is effectively aggregated and boundary prediction is performed by aggregating multi-scale features to provide assistance for semantic segmentation tasks.In the gradient update part,gradient surgery is added to reduce the interference to training by gradient conflicts between tasks caused by adding auxiliary tasks and speed up model convergence.Experiments are conducted on ADE20k dataset and Cityscapes dataset.By increasing the calculation amount of 2.69 M parameters and 24.67 G,the average intersection over union ratio of the network is improved by 2.38％,proving the effectiveness of the proposed method.
关键词
Segformer/级联注意力/辅助任务/边界预测/梯度手术
Key words
Segformer/cascaded attention/auxiliary tasks/boundary prediction/gradient surgery
维普
万方数据