LightNet+: Boosted Light-Weighted Network for Smoke Semantic Segmentation

Smoke is a crucial indicator of early fires and gas leaks. By segmentation of smoke in the image, the detailed information such as smoke direction of spread, and source location can be obtained. Considering the popularity of video surveillance systems, smoke segmentation is of great significance. This paper uses an efficient boosted light-weighted network for smoke semantic segmentation with only 0.53M network parameters. Firstly, we propose a novel Smoke Feature Extractor (SFE) to improve the capability of smoke feature representation in encoders. The SFE achieves scale invariance by repeatedly stacking Multi-scale Foreground Enhancement Module (MSFEM) at different coding stages. The MSFEM increases the field of view of features by down-sampling the feature maps, and fusing information from different scale spaces into weights to enhance smoke foreground information. Secondly, we propose a novel Attention-guided Coupled Feature Fusion Module (ACFFM) that introduces Self-Refinement Coefficients (SRCs) generated from cross-layer fusion to weight the original layer images. This two-stage fusion approach effectively utilizes information from different scales to alleviate the impact of scale changes and performs hierarchical decoding. ACFFM serves as a step-by-step recovery model that guides cross-level feature mapping. It retains the original feature size to capture local information, while effectively preventing small objects from being ignored. Finally, we propose a Smoke Feature Decoder (SFD) and utilize a Global Coefficient Path to further aggregate feature expression capabilities. The smoke feature decoder fuses the outputs from different levels of ACFFM and weights them with deep global semantic attention coefficients. Experimental results on synthetic and forest smoke datasets show the effectiveness and superiority of our proposed method.