DAFNet: Dual-branch Atrous fusion enhanced network

Semantic segmentation is one of the crucial research directions of computer vision. The accuracy, fine-grained, and speed of segmentation results restrict the practical application of semantic segmentation. In order to take full advantage of the spatial information in the high-level semantic features and the global vision in the low-level semantics features. This paper proposes a novel network structure framework DAFNet based on the pyramid structure. In view of the poor segmentation effect of PSPNet on the original basis, by designing a dual-path structure, adding an atrous space enhancement module and a cross-feature fusion module, DAFNet has improved the segmentation accuracy of the original network. In addition, in order to have better segmentation speed and make the model have a lighter application scenario range, this paper utilizes Mobilenetv2 as the backbone network. Experimental results show that DAFNet has achieved satisfactory results on the Pascal VOC 2012 val dataset.