Class Attention Network for Semantic Segmentation of Remote Sensing Images

원문 초록

Semantic segmentation in remote sensing images is beneficial to detect objects and understand the scene in earth observation. However, classical networks always failed to obtain an accuracy segmentation map in remote sensing images due to the imbalanced labels. In this paper, we proposed a novel class attention module and decomposition-fusion strategy to cope with imbalanced labels. Based on this motivation, we investigate related architecture and strategy by follows. (1) we build a class attention module to generate multi-class attention maps, which forces the network to keep attention to small sample categories instead of being flooded by large sample data. (2) we introduce salient detection, which decomposes semantic segmentation into multi-class salient detection and then fuses them to produce a segmentation map. Extensive experiments on popular benchmarks (e.g., US3D dataset) show that our approach can serve as an efficient plug-and-play module or strategy in the previous scene parsing networks to help them cope with the problem of imbalance labels in remote sensing images.