An Efficient Smoke Detection Approach Based on Dual-Channel Neural Network

Smoke is the key feature of fire detection in its early age. Thus, an efficient smoke detection approach (i. e. an accurate and rapid method) is essential important to prevent fires. However, it is difficult to obtain an efficient method due to non-obvious details and monotonous color of smoke images. Moreover, traditional methods of smoke detection based on CNN contains lots of parameters and operations, which severely influents the computing efficiency in practical applications. Thus, we propose an efficient dual-channel neural network (EDCNN) on the basis of the state-of-the-art DCNN. Concretely, we use the linear inverted bottleneck (LIB) to replace the traditional convolution layers on DCNN to build a light weight deep neural network. The introduction of the LIB block can efficiently trade off between accuracy and latency. Moreover, ReLU6 is used as the activation function, because it is more suitable for low-precision hardware devices. We then use some experimental results to demonstrate the effectiveness of EDCNN compared with the competitors for smoke detection in terms of model and computational complexity.