Lightweight Semantic Segmentation of UAV Images Based on CNN

UAV aerial survey technology has become an important means of quickly obtaining critical information about specific locations. To address the problem of low segmentation accuracy and low real-time performance of traditional semantic segmentation models on UAV platforms with lower computational capabilities. We designed a lightweight bilateral image semantic segmentation network based on BiseNet. Inspired by the concept of receptive field, we propose a deeply separable fusion pyramid pool module. Rapid extraction of multi-scale contextual information from different receptive fields. Then We based on the channel attention mechanism, the dual feature attention module and feature fusion module are proposed. BiseNet achieved 81.35% average crossover to MIoU and 89.52% overall pixel accuracy OA on the Potsdam datasets. The MIoU and OA on the Vaihingen datasets were 71.82% and 87.89%. BiseNet shows good performance in segmenting high-resolution aerial images.