Remote sensing semantic segmentation with convolution neural network using attention mechanism

Semantic image segmentation is an essential part of remote sensing image processing because accurate understanding of the ground information is the first step in obtaining useful knowledge of surface coverage. The popular semantic segmentation convolutional neural network model (DeepLab v3+) cannot electively use attention information, resulting in coarse segmentation boundaries. In this work, a new type of bottleneck using attention information which can extract semantic information and more abundant features from images is proposed. Compared with original network, the model using new bottleneck finely segments the target regions, solves the problem of segmentation boundary roughness better, leading to higher mloU and accuracy. Experimental results based on the dataset in the ISPRS benchmark on urban object classification show bringing attention model into semantic segmentation neural network improves performance.