Boundary Detector Network with Squeeze-and-Excitation Network for Semantic Segmentation of Remote Sensing Images

Abstract

With the rapid development of deep learning, remote sensing image interpretation using semantic segmentation has been widely used in various fields. Deep convolutional Neural network (DCNN) is an important tool for semantic segmentation. However, increasing the complexity of the network and deepening the neural network not only can improve the accuracy of the network, but also lead to the overfitting problem. Based on this point, this paper used the squeeze-and-excitation (SE) modules to optimize the network. SE modules are superimposed after each layer of the ResNet-101 network. This method not only improves network performance but does not lose small target information too much. In this experiment, RGB and nDSM multi-channels of high resolution remote sensing (RS) images with multi-semantic information are used. Experimental results show that the accuracy of the proposed method in the training set and verification set can reach 98.11% and 98.45%, and the model is in continuous convergence. OA, Precision, Recall, F1 and MIOU can reach 85.73%, 84.64%, 85.76%, 85.67%, and 75.47% respectively.