SCU-Net: Semantic Segmentation Network for Learning Channel Information on Remote Sensing Images

Abstract

Extracting detailed information from remote sensing images is an important direction in semantic segmentation. Not only the amounts of parameters and calculations of the network model in the learning process but also the prediction effect after learning must be considered. This paper designs a new module, the upsampling convolution-deconvolution module (CDeConv). On the basis of CDeConv, a convolutional neural network (CNN) with a channel attention mechanism for semantic segmentation is proposed as a channel upsampling network (SCU-Net). SCU-Net has been verified by experiments. The mean intersection-over-union (MIOU) of the SCU-Net-102-A model reaches 55.84%, the pixel accuracy is 91.53%, and the frequency weighted intersection-over-union (FWIU) is 85.83%. Compared with some of the state-of-the-art methods, SCU-Net can learn more detailed information in the channel and has better generalization capabilities.