A multiscale fuzzy dual-domain attention network for urban remote sensing image segmentation

Semantic segmentation of high-resolution remote sensing images plays an important role in the remote sensing community. However, many indistinguishable objects are prevalent within urban remote sensing images, and some objects belonging to the same class are different and many objects that do not belong to the same class are similar. These tricky objects make the images exhibit low-interclass variance and high-intraclass variance, which significantly limits segmentation performance. Therefore, a fresh insight was presented to alleviate this issue by incorporating the fuzzy pattern recognition method and deep-learning method. Specifically, we proposed a multiscale fuzzy dual-domain attention network (MFDAN). In MFDAN, a two-dimensional Gaussian fuzzy learning module is proposed to eliminate those factors that influence the intraclass and interclass variance. In addition, a dual-domain attention module is proposed to derive more informative semantic representations in the channel and spatial domains, respectively. These two modules will be integrated in a multiscale perspective. Extensive experiments on the benchmark datasets illustrate qualitatively and quantitatively that the proposed MFDAN is competitive.