LANet: Local Attention Embedding to Improve the Semantic Segmentation of Remote Sensing Images

The trade-off between feature representation power and spatial localization accuracy is crucial for the dense classification/semantic segmentation of aerial images. High-level features extracted from late layers a neural network are rich in semantic information, yet have blurred details; low-level early contain more pixel-level but isolated noisy. It therefore difficult to bridge gap high due their difference terms physical information content distribution. In this work, we contribute solve problem by enhancing two ways. On one hand, patch attention module (PAM) proposed enhance embedding context based on patch-wise calculation local attention. other an (AEM) enrich focus high-level features. Both modules light-weight can be applied process convolutional networks (CNNs). Experiments show that, integrating into baseline Fully Convolutional Network (FCN), resulting (LANet) greatly improves performance over outperforms methods image datasets.