Semantic segmentation with step-by-step upsampling of the fusion context

The existing semantic segmentation network deelabv3+ has the problem of weak segmentation ability to small-scale target objects and rough edge segmentation. The method of parallel connection of multiple resolution subnets in HRNet network is introduced. After deeplabv3+ down sampling, the network layers of different sizes were fused with features, and the decode side was fused with up sampling step by step to improve the edge segmentation accuracy. Attention mechanism is added before feature fusion to improve the recognition ability of small object. At the end, the edge is refined again by using CRF random vector field. The test is carried out on Pascal VOC 2012, the experimental results show that: the segmentation is more detailed from the image edge details, the recognition of small objects is more accurate, the Pixel Accuracy (PA) and Mean Intersection over Union (MIOU) are improved compared with deeplabv3+.