Semi-supervised wildfire smoke detection based on smoke-aware consistency

Abstract

The semi-transparency property of smoke integrates it highly with the background contextual information in the image, which results in great visual differences in different areas. In addition, the limited annotation of smoke images from real forest scenarios brings more challenges for model training. In this paper, we design a semi-supervised learning strategy, named smoke-aware consistency (SAC), to maintain pixel and context perceptual consistency in different backgrounds. Furthermore, we propose a smoke detection strategy with triple classification assistance for smoke and smoke-like object discrimination. Finally, we simplified the LFNet fire-smoke detection network to LFNet-v2, due to the proposed SAC and triple classification assistance that can perform the functions of some specific module. The extensive experiments validate that the proposed method significantly outperforms state-of-the-art object detection algorithms on wildfire smoke datasets and achieves satisfactory performance under challenging weather conditions.

All-in-one aerial image enhancement network for forest scenes

“…As discussed in the introduction, aerial image enhancement could be helpful in improving the performance of fire detection approaches in forest fire prevention based on drone imagery monitoring. Therefore, we train a smoke detection algorithm ( Wang et al., 2022 ) on the raw dataset. To verify the effectiveness of our method in boosting image detection, we use the results of image dehazing, image motion deblurring, and image compression deblurring as input exemplars for the detection algorithm, respectively.…”