CSANet: Cross-Scale Axial Attention Network for Road Segmentation

Road segmentation from remote sensing images is an important task in many applications. However, due to the high density of roads and complex background, are often occluded by trees. This makes accurate road a challenge task. Most existing networks rely on convolutions with small kernels; however, these methods cannot obtain satisfying results because long-range dependencies not captured intrinsic relationships between feature maps at different scales fully exploited. In this paper, deep neural network based cross-scale axial attention mechanism proposed address problem. model enables low-resolution features aggregate global contextual information high-resolution features. Among them, realizes using vertical horizontal sequentially. With strategy, dense can be extremely low computational cost. The effectively combine fine-grained coarse-grained method propagate without losing details. Our achieves IoUs 58.98 65.28 Massachusetts Roads dataset DeepGlobe outperforms other methods.