Deep convolutional neural networks for fire detection in images

Abstract

Detecting fire in images using image processing and computer vision techniques has gained a lot of attention from researchers during the past few years. Indeed, with sufficient accuracy, such systems may outperform traditional fire detection equipment. One of the most promising techniques used in this area is Convolutional Neural Networks (CNNs). However, the previous research on fire detection with CNNs has only been evaluated on balanced datasets, which may give misleading information on real-world performance, where fire is a rare event. Actually, as demonstrated in this paper, it turns out that a traditional CNN performs relatively poorly when evaluated on the more realistically balanced benchmark dataset provided in this paper. We therefore propose to use even deeper Convolutional Neural Networks for fire detection in images, and enhancing these with fine tuning based on a fully connected layer. We use two pretrained state-of-the-art Deep CNNs, VGG16 and Resnet50, to develop our fire detection system. The Deep CNNs are tested on our imbalanced dataset, which we have assembled to replicate real world scenarios. It includes images that are particularly difficult to classify and that are deliberately unbalanced by including significantly more non-fire images than fire images. The dataset has been made available online. Our results show that adding fully connected layers for fine tuning indeed does increase accuracy, however, this also increases training time. Overall, we found that our deeper CNNs give good performance on a more challenging dataset, with Resnet50 slightly outperforming VGG16. These results may thus lead to more successful fire detection systems in practice.