Semantic segmentation of UAV remote sensing images based on edge feature fusing and multi-level upsampling integrated with Deeplabv3.

Deeplabv3+ currently is the most representative semantic segmentation model. However, Deeplabv3+ tends to ignore targets of small size and usually fails to identify precise segmentation boundaries in the UAV remote sensing image segmentation task. To handle these problems, this paper proposes a semantic segmentation algorithm of UAV remote sensing images based on edge feature fusing and multi-level upsampling integrated with Deeplabv3+ (EMNet). EMNet uses MobileNetV2 as its backbone and adds an edge detection branch in the encoder to provide edge information for semantic segmentation. In the decoder, a multi-level upsampling method is designed to retain high-level semantic information (e.g., the target's location and boundary information). The experimental results show that the mIoU and mPA of EMNet improved over Deeplabv3+ by 7.11% and 6.93% on the dataset UAVid, and by 0.52% and 0.22% on the dataset ISPRS Vaihingen.