Semantic Segmentation of Remote Sensing Architectural Images Based on GAN and UNet3+ Model

Abstract

Semantic segmentation of remote sensing building images can provide important data support for urban planning and resource management. It also plays a crucial role in assessing building density, monitoring urban expansion, and optimizing traffic planning. In recent times, with the continuous integration of computer vision and deep learning, Convolutional Neural Networks (CNNs) have achieved outstanding results in semantic segmentation tasks for remote sensing images. Although deep CNNs can significantly improve the accuracy of semantic segmentation for remote sensing images, some network models used for segmentation tasks still have limitations, such as low segmentation precision and inadequate feature extraction. In this paper, we propose an adversarial semantic segmentation network based on Generative Adversarial Networks (GANs). To better extract the features and semantics of buildings in remote sensing images, we introduce the UNet3+ network as the segmentation network of the adversarial network for the first time and make improvements to the UNet3+ network. We add the scSE (Spatial Channel Squeeze and Excitation) attention mechanism to the network, the scSE attention mechanism enhances the network’s perception of different channel features by considering their correlations in the channel dimension, allowing it to capture fine-grained details and coarse-grained semantics at the full scale. In this paper, we conduct experiments on the Inria Aerial Image Labeling dataset, and the results show that our method outperforms other network models mentioned in the paper in terms of performance.