Edge-Enhanced GCIFFNet: A Multiclass Semantic Segmentation Network Based on Edge Enhancement and Multiscale Attention Mechanism

Abstract

In recent years, remote sensing images (RSIs) have witnessed significant improvements in both quality and quantity. With the application of deep-learning techniques, these RSIs can be more effectively utilized to harnessed to aid in environment monitoring and urban planning. Semantic segmentation, as a common task in RSIs processing, confronts numerous challenges, including inaccurate classification, fuzzy boundaries, and other problems. This article proposes a novel semantic segmentation network known as the edge-enhanced global contextual information guided feature fusion network to address these challenges. This network consists of an edge-enhanced part and a backbone network part. First, in the encoding stage, the recurrent criss-cross attention block is employed, which incorporates spatial attention, mechanisms to capture global information. Second, in the decoding stage, a channel attention residual block module is proposed to facilitate the fusion of high-level and low-level features. Moreover, we enhance the network's ability to extract edge information during training by sharing parameters between the backbone and employing a specialized loss function. The network proposed in this article utilizes both channel attention and spatial attention at different stages, effectively utilizing edge information. Finally, we conduct experiments using the Yinchuan dataset and the LoveDA dataset. The experimental results show that the proposed network demonstrates excellent performance on both datasets.

Keywords

• Attention
• convolutional neural network (CNN)
• edge segmentation
• remote sensing images (RSIs)
• semantic segmentation