Smoke Detection Algorithm based on Negative Sample Mining

Forest fire is one of the most dangerous disasters that threaten the safety of human life and property. In order to detect fire in time, we detect the smoke when the fire breaks out. However, it is still a challenging task due to the variations of smoke in color, texture, shape and the disturbances of smoke-like objects. Therefore, the accuracy of smoke detection is not high, and it is accompanied by a high false positive rate, especially in the real environment. To tackle this problem, this paper proposes a novel model based on Faster Region-based Convolutional Network (R-CNN) which utilizes negative sample mining method. The proposed method allows the model to learn more negative sample features, thereby reducing false positives in smoke detection. The experiments are performed on self-created dataset containing 11958 images which are collected from cameras placed in villages or towns and existing datasets. Compared to other smoke datasets, the self-created dataset is larger and contains complex scenes. The proposed method achieves 94.59% accuracy, 94.35% precision and 5.76% false positive rate on self-created dataset. The results show that the proposed network is better and more robust than previous works.