Multi-Scale Interactive Fusion Global-Local Network for Smoke Semantic Segmentation

Smoke semantic segmentation in early fire detection faces multiple challenges, with the translucency of smoke leading to boundary blurring, turbulent motion inducing dynamic morphological aberrations, and multi-scale diffusion characteristics in monitoring scenarios. In this paper, we propose a multi-scale global-local interaction fusion network (MIFGLNet) to achieve high-precision real-time smoke segmentation by synergistically optimizing the local texture perception of CNN and the global diffusion modeling capability of Transformer. Firstly, a multi-level feature complementary module (MFCM) is proposed to establish cross-branch interaction channels between the layers of the two-branch encoder, and adaptively fuse the local detail information and global long-range dependency information through a multi-level self-attention mechanism. A lightweight MLP multiscale decoder (MLPMFD) is constructed to replace the traditional inverse convolution with bilinear interpolative convolution to fuse multiscale information and reduce the number of parameters. Finally, an edge-sensitive loss function is designed to optimize the pixel-level classification stability in translucent regions. We conducted extensive experiments on the USS dataset, and the mIoU reaches <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 1. 6 \%}$</tex>, which significantly outperforms existing methods.