Smoke Detection Based on Dark Channel and Convolutional Neural Networks

Smoke is an important sign of fire and could enable early fire detection. However, it could be hard to discriminate smoke in images because of the irregular shapes and density variation of the smoke. Background interference could also influence the performance of smoke detection methods. Moreover, it is difficult to collect large scale smoke dataset and the dataset used to train the classifier for smoke identification is usually severely imbalanced. To address these problems, a solution combining dark channel image input and a relative concise convolutional neural network (CNN) was developed. The dark channel of an image could well enhance the difference between the smoke and background. The relative concise CNN could be efficiently trained on small dataset. Furthermore, data augmentation techniques have been employed to generate more training samples and alleviate the influence from small dataset. To deal with the data imbalance issue, we apply weighted softmax loss to highlight the contribution of the samples from the minority class. Extensive experiments have verified that our method has superior performance against the other smoke detection algorithms.