Lightweight smoke segmentation algorithm based on improved DeepLabV3+

Abstract

The fires can be effectively detected by monitoring fire smoke. However, existing fire smoke segmentation algorithms have not performed well on the small smoke and edges of large smoke. This article proposed an improved lightweight DeepLabV3+ smoke segmentation algorithm based on deep learning, which effectively detects smoke. The feature extraction network of the DeeplabV3+ algorithm was replaced, which reduced the number of parameters. This improvement enhanced the algorithm's ability to extract smoke features and segment smoke. The convolutional block attention module (CBAM) was added to the encoder module to enhance the algorithm's perception of small smoke. This improvement enhanced the algorithm's segmentation capability for smoke in complex backgrounds, and effectively alleviate smoke mis-segmentation. Experimental results on the test set show a noticeable gain up to 6.46% in smoke intersection over union(sIoU), 4.28% in mean intersection over union(mIoU), and 1.72% in mean pixel accuracy (mPA), respectively. Moreover, the improved algorithm's weight size is only 10.76% of the original algorithm's weight size. The experimental results show that the improved smoke segmentation algorithm, which has higher smoke segmentation accuracy, shorter training time, and a smaller model size compared to the original DeepLabV3+ algorithm. The improved smoke segmentation algorithm is more suitable for real-time smoke monitoring tasks.

Keywords

• deep learning
• segmentation algorithms
• deeplabv3+ algorithms
• fire smoke
• convolutional block attention module