Dilated Residual Network Based on Dual Expectation Maximization Attention for Semantic Segmentation of Remote Sensing Images

Compared with common RGB images, remote sensing images (RSIs) have larger size and lower spatial resolution. RSIs are usually cropped into sub-images for training convolutional neural networks (CNNs), which loses amounts of context information, thus limiting the extraction of feature interdependencies and reducing the accuracy of semantic segmentation. In this paper, a novel dilated residual network based on dual expectation maximization attention (DE-MANet) is proposed for semantic segmentation of RSIs. In specific, we append a dual expectation maximization attention (DEMA) module on top of the dilated CNN. The spatial expectation maximization attention (SEMA) can model spatial feature interdependencies to acquire rich long-range contextual information. The channel expectation maximization attention (CEMA) enhances discriminant ability of channel-wise feature representations through extracting the channel dependencies. We evaluate the model on the dataset released in the Tianzhi Cup Artificial Intelligence Challenge and achieve 85.60% pixel accuracy and 69.00% mean intersection over union (mIoU).