Multiscale Fusion Convolutional Network in Real-time Semantic Segmentation

Abstract

To achieve semantic segmentation tasks in practical applications such as autonomous driving, networks need to efficiently process high-resolution images while maintaining high accuracy. This requires methods to effectively fuse spatial information in high-resolution images with semantic information in low-resolution images. To address this, this paper proposes a Multi-scale Fusion Convolutional Network (MFCNet) based on a single-branch network structure. In order to simultaneously handle information at different scales and assist the network in capturing a wide range of contextual information, separable Multi-Scale Convolution Modules (MSCM) are introduced to enable the network to obtain richer and more comprehensive feature representations. Additionally, considering that shallow-level information is difficult to directly restore resolution, a Dual-Attention Fusion Module (DAFM) is designed, introducing two attention mechanisms to respectively weight feature maps at different resolutions. Experimental results demonstrate that MFCNet achieves outstanding performance in real-time semantic segmentation tasks.