A Novel Encoder-Decoder Structure-based Transformer for Fine-Resolution Remote Sensing Images

Abstract Full convolution neural network (FCN) based on an encoder-decoder structure has become a standard in the semantic segmentation domain. Encoder-decoder architecture is effective means to get finer-grained performance. Encoders constantly extract multilevel features, and then use decoders gradually introduce low-level features into high-level features. Context information critical for accurate segmentation, which main direction of at present. So many efforts have been made make better this kind information, including codec structure, void (expanded convolution), attention mechanism. However, most these schemes are Resnet or other variants FCN, makes it unable rid defective local receptive field itself. In work, we pyramid visual converter (PVT) replace traditional full architecture, design novel more effectively utilize context information.