Fire Recognition Based On Multi-Channel Convolutional Neural Network

In recent years, fire recognition methods have received more and more attention in the fields of academy and industry. Current sensor-based recognition methods rely heavily on the external physical signals, which will probably reduce the recognition precision if the external environment changes dramatically. With the rapid development of high-definition camera, the methods based on image feature extraction provide another solution which tries to conduct pattern recognition for the monitoring video. However, these methods couldn't be widely and successfully applied to fire detection due to two deficiencies: (1) there are too many interference items like lamplight and car highlight in the room or tunnel, which will disturb the recognition performance largely; (2) The features depend on much prior knowledge about flame and smoke, and there lacks a universal and automatic extraction method for various fire scenes. As a breakthrough in pattern recognition, deep learning is capable of exploring the useful information from raw data, and can automatically provide accurate recognition results. Therefore, based on deep learning idea, a novel fire recognition method based on multi-channel convolutional neural network is proposed in this paper to overcome the deficiencies mentioned above. First, three channel colorful images are constructed as the input of convolutional neural network; Second, the hidden layers with multiple-layer convolution and pooling are constructed, and simultaneously, the model parameters are find tuned by using back propagation; Finally, softmax method is used to conduct the classification about fire recognition. To save the training time, we utilize GPU to construct training and test models. From public fire dataset and Internet, we collect 7000 images for training and 4494 images for test, and then run experiments with the comparison of four baseline methods including deep neural network, support vector machine based on scale-invariant feature transform feature, stack auto-encoder and deep belief network. The experimental results show that the proposed method is more capable of restoring the features of input image by means of hidden output figure, and for various flame scenes and types, the proposed method can reach 98% or more classification accuracy, getting improvement of around 2% than the traditional feature-based method. Also, the proposed method always outperforms other Deep Learning methods in terms of ROC curve, recall rate, precision rate and F1-score.