The automation of fire detection systems can reduce the loss of life and property by allowing a fast and accurate response to fire accidents. Although visual techniques have some advantages over sensor-based methods, conventional image processing-based methods frequently cause false alarms. Recent studies on convolutional neural networks have overcome these limitations and exhibited an outstanding performance in fire detection tasks. Nevertheless, previous studies have only used single-scale feature maps for fire image classification, which are insufficiently robust to fires of various sizes in the images. To address this issue, we propose a multi-scale prediction framework that exploits the feature maps of all the scales obtained by the deeply stacked convolutional layers. To utilize the feature maps of various scales in the final prediction, this paper proposes a feature-squeeze block. The feature-squeeze block squeezes the feature maps spatially and channel-wise to effectively use the information from the multi-scale prediction. Extensive evaluations demonstrate that the proposed method outperforms the state-of-the-art convolutional neural networks-based methods. As a result of the experiment, the proposed method shows 97.89% for F1-score and 0.0227 for false positive rate in the average of evaluations for multiple.

火灾探测系统的自动化可以通过快速，准确地响应火灾事故来减少生命和财产损失。尽管视觉技术相对于基于传感器的方法具有一些优势，但是基于常规图像处理的方法经常会导致错误警报。卷积神经网络的最新研究已经克服了这些限制，并在火灾探测任务中表现出出色的性能。然而，先前的研究仅将单尺度特征图用于火灾图像分类，这对于图像中各种尺寸的火灾都不够鲁棒。为了解决这个问题，我们提出了一种多尺度预测框架，该框架利用由深度堆叠的卷积层获得的所有尺度的特征图。要在最终预测中利用各种比例的特征图，本文提出了一种特征压缩块。特征压缩块在空间和通道方向上挤压特征图，以有效地使用来自多尺度预测的信息。大量评估表明，所提出的方法优于基于卷积神经网络的最新方法。实验结果表明，所提出的方法在F1得分中的平均得分为97.89％，在假阳性率中的得分为0.0227。