预指导的多阶段特征融合的图像语义分割网络.

In view of the current semantic segmentation can not accurately identify image edges and small objects, and simple fusion of multi-stage features will cause information redundancy, confusion and other problems, this paper proposed a Pre-guidanced multi-stage feature fusion network (PGMFFNet). PGMFFNet employed a encoder-decoder structure, at the encoder stage, which used a pre-guidance module to guide the information in each stage. Strengthened the relationship between the features of each stage, and solved the semantic confounding problems in the subsequent fusion process of the features of each stage. At the decoder stage, which used the multi-path up-pyramid sampling module to fuse high-level semantic features, and then used the improved dense void space pyramid pool module to further expand the sensory field of the fused features, and finally fused the feature information of high and low levels to make the segmentation effect of small objects better. This paper verified PGMFFNet on CityScapes open data set,and the mean intersection over union (MIoU) obtained to 78.38%, showing good segmentation effect. [ABSTRACT FROM AUTHOR]