Bilateral attention network for semantic segmentation

Enhancing network feature representation capabilities and reducing the loss of image details have become focus semantic segmentation task. This work proposes bilateral attention for segmentation. The authors embed two modules in encoder decoder structures . Specifically, high-level features structure integrate all channel maps through dense relationships learned by correlation coefficient module. positively correlated channels promote each other, negatively suppress other. In structure, low-level selectively emphasize edge detail information map position expression is improved fusion to obtain more accurate results Finally, verify effectiveness model, conduct experiments on PASCAL VOC 2012 Cityscapes scene analysis benchmark data sets achieve a mean intersection-over-union 74.92% 66.63%, respectively.