A scale-cross non-local network with higher-level semantics guidance for smoke segmentation

Abstract

Smoke semantic segmentation (SSS) is particularly challenging task due to the various patterns of the target itself, which are caused by the characteristics of smoke, like, non-rigid, translucent, fuzzy, environment-sensitive, and so forth. This paper tailor-makes the Scale-Cross Non-Local Network (SCNN) for Smoke Segmentation, aiming to accurately locate the position of smoke in complex scenes. While non-local enjoys the bonus of the excellent competence in modeling long-range contextual dependencies acquired by self-attention, the constraint on single-scale input and the suitability for low-resolution feature erode its capability in information representation. To address these issues, we bespoke a Scale-Cross Non-Local (SCNL) module to better integrate local features with global dependencies. In practical scenes, diverse non-smoke objects sharing similarity with smoke pose great obstacles to accurate location of smoke. As a solution, we design a Pyramid Irregular Convolution (PIC) module containing rich high-level semantic to further refine the feature representation of segmentation task. By supervising classification task, the high-level semantics obtained can guide the segmentation feature to correct semantic errors at the image level and alleviate the issue of between-class similarity. To assess its generalization ability, we empirically evaluate our SCNN on extensive synthetic and real data. Experimental results demonstrate that SCNN achieves state-of-the-art performance, exhibiting enhanced smoke localization, accuracy in boundary detection, and a significant reduction in the false segmentation rate for smoke-like objects.