An Efficient Deep Learning Algorithm for Fire and Smoke Detection with Limited Data

Detecting smoke and fire from visual scenes is a demanding task, due to the high variance of the color and texture. A number of smoke and fire image classification approaches have been proposed to overcome this problem; however, most of them rely on either rule-based methods or on handcrafted features. We propose a novel deep convolutional neural network algorithm to achieve high-accuracy fire and smoke image detection. Instead of using traditional rectified linear units or tangent functions, we use adaptive piecewise linear units in the hidden layers of the network. We also have created a new small dataset of fire and smoke images to train and evaluate our model. To solve the overfitting problem caused by training the network on a limited dataset, we improve the number of available training images using traditional data augmentation techniques and generative adversarial networks. Experimental results show that the proposed approach achieves high accuracy and a high detection rate, as well as a very low rate of false alarms.