A label-relevance multi-direction interaction network with enhanced deformable convolution for forest smoke recognition

Abstract

Forest fires pose a significant threat to both the economy and ecology, causing extensive damage. Smoke serves as a crucial indicator of forest fires, often appearing before the actual fire. However, existing methods for smoke recognition are susceptible to missed alarms and false alarms due to the diverse nature of smoke and the presence of various continuous interferences in complex real-world scenarios. To tackle these challenges, this paper proposes a label-relevance multi-direction interaction network with enhanced deformable convolution. Firstly, to ensure the extraction of robust features, we propose an enhanced deformable convolution module that breaks away from fixed geometric structures and incorporates interval information. Secondly, to prevent high-response high-level features from overshadowing low-level features during the feature interaction process, we introduce a multi-directional feature interaction module that obtains complementary features from different convolution layers. Lastly, to leverage the relevance and pixel distribution information in the label image, we propose a new loss term based on generative adversarial learning. This loss term measures the distribution similarity between the network's predictions and the ground truth. Through extensive experiments, we demonstrate that our model accurately estimates smoke pixels and outperforms existing smoke recognition methods.