GLFFNet: A Global and Local Features Fusion Network with Biencoder for Remote Sensing Image Segmentation

Abstract

In recent years, semantic segmentation of high-resolution remote sensing images has been gradually applied to many important scenes. However, with the rapid development of remote sensing data acquisition technology, the existing image data processing methods are facing major challenges. Especially in the accuracy of extraction and the integrity of the edges of objects, there are often problems such as small objects being assimilated by large objects. In order to solve the above problems, based on the excellent performance of Transformer, convolution and its variants, and feature pyramids in the field of deep learning image segmentation, we designed two encoders with excellent performance to extract global high-order interactive features and low-order local feature information. These encoders are then used as the backbone to construct a global and local feature fusion network with a dual encoder (GLFFNet) to effectively complete the segmentation of remote sensing images. Furthermore, a new auxiliary training module is proposed that uses the semantic attention layer to process the extracted feature maps separately, adjust the losses, and more specifically optimize each encoder of the backbone, thus optimizing the training process of the entire network. A large number of experiments show that our model achieves 87.96% mIoU on the Potsdam dataset and 80.42% mIoU on the GID dataset, and it has superior performance compared with some state-of-the-art methods on semantic segmentation tasks in the field of remote sensing.