HilbertSCNet: Self-attention networks for small target segmentation of aerial drone images

Abstract

In recent years, many advanced semantic segmentation algorithms have made great progress in common datasets. However, in the drones aerial photography dataset, since some of the segmentation targets are very small in such high-resolution images, these current algorithms are not designed specifically for them resulting in their mediocre performance in the drones aerial photography dataset. To address the problem of difficult segmentation of these small targets, this paper proposes a new semantic segmentation network model-HilbertSCNet: (1) Combine the image dimensionality reduction algorithm of Hilbert curve traversal and the idea of dual pathway to design a new spatial computation module to solve the problem of small target information is easily lost in the downsampling; (2) Add a windowing algorithm in the image dimensionality reduction algorithm of Hilbert curve traversal to solve the problem that the computational complexity of the module is too high, which makes the module cannot be applied to high-resolution feature maps. Experiments show that the proposed network is very effective for segmentation of small targets under high-resolution maps such as drone aerial photography, with certain superiority and generalization. The overall segmentation performance is improved compared to the current new network, including 1.76% better than OCNet in MIoU and 4.55% better than the dedicated drone algorithm RCCT-ASPPNet.