A lightweight network for real-time smoke semantic segmentation based on dual paths

There are challenges exist in the segmentation of smoke contours on images currently, the requirements for limited processing resources and low-latency operations based on monitoring platform, and the balance between high accuracy and real-time efficiency of the model performance. Also, smoke always shows to be translucency, resulting in a highly complex mixture of the background and itself, sparse or small smoke is not visually obvious, and its borders are often blurred. Therefore, the task of separating smoke from a single image is challengeable. To overcome the challenges, a dual-path real-time smoke segmentation network based on BiSeNet is adopted in this research, and a PPM to expand the receptive field in the spatial path is added to improve the ability to obtain global information. At the same time, a lightweight ECA channel attention module, included in a context path with fast down-sampling strategy, could reduce the complexity while ensuring the effect of the model. Experimental results show that, on all the self-built dataset, the public synthetic smoke dataset and public videos, the proposed method shows excellent performance, and the detection speed reaches real-time segmentation level, showing high practical value.