Semantic Segmentation Algorithm of Remote Sensing Images Based on Improved Panoptic

Semantic segmentation of images can play an important role in land use, building extraction, road extraction and vehicle detection using remote sensing images. In the scenario of applying Panoptic FPN algorithm for remote sensing images, we found that the decoder of this algorithm is not robust in feature extraction and cannot do weighted enhancement of key spaces and channels, and its encoder loses a lot of high-dimensional semantic features when simply fusing semantic information at all levels by simple addition. To address these two problems, we propose the Se-Resnext encoder based on the attention mechanism and the Concat-based feature fusion mechanism, respectively, and verify the effectiveness of the methods through experiments. The accuracy of semantic segmentation is improved in both suichang dataset and posdam dataset.