HRSF-Net: A High-Resolution Strong Fusion Network for Pixel-Level Classification of the Thin-Stripped Target for Remote Sensing System

Abstract

High-resolution pixel-level classification of the roads and rivers in the remote sensing system has extremely important application value and has been a research focus which is received extensive attention from the remote sensing society. In recent years, deep convolutional neural networks (DCNNs) have been used in the pixel-level classification of remote sensing images, which has shown extraordinary performance. However, the traditional DCNNs mostly produce discontinuous and incomplete pixel-level classification results when dealing with thin-stripped roads and rivers. To solve the above problem, we put forward a high-resolution strong fusion network (abbreviated as HRSF-Net) which can keep the feature map at high resolution and minimize the texture information loss of the thin-stripped target caused by multiple downsampling operations. In addition, a pixel relationship enhancement and dual-channel attention (PRE-DCA) module is proposed to fully explore the strong correlation between the thin-stripped target pixels, and a hetero-resolution fusion (HRF) module is also proposed to better fuse the feature maps with different resolutions. The proposed HRSF-Net is examined on the two public remote sensing datasets. The ablation experimental result verifies the effectiveness of each module of the HRSF-Net. The comparative experimental result shows that the HRSF-Net has achieved mIoU of 79.05% and 64.46% on the two datasets, respectively, which both outperform some advanced pixel-level classification methods.