Real-Time Semantic Segmentation of Aerial Images Based on Dual-Feature Attention Networks

With the widespread use of unmanned aerial vehicles (UAVs), the segmentation of their aerial images has been widely applied. However, fast semantic segmentation of high-resolution UAV aerial images still faces several challenges: limited hardware processing resources and high real-time requirements of UAV platforms, the balance between model accuracy and real-time efficiency, and the poor inference ability of models for small target classes in aerial images. To address these issues, this paper proposes a lightweight remote sensing image bilateral segmentation network (RS-BiseNet) that achieves good inference accuracy and real-time performance for aerial images. Specifically, inspired by the concept of receptive fields, a deep separable fusion pyramid pooling module (DSFPPM) is proposed to rapidly extract multi-scale contextual information from different receptive fields. Furthermore, a dual feature attention module (DFAM) and a feature fusion module (FFM) are proposed based on the channel attention mechanism to extract fine-grained details. RS-BiseNet was evaluated on the ISPRS Potsdam dataset and obtained an improvement of 2.96% MIoU and 1.59% OA over Aerial-BiSeNet.