基于卷积神经网络的烟雾检测

针对已提出的很多烟雾检测方法中都是基于手工制作的特征或者使用原始图片直接作为神经网络的输入，减少了深度学习的鲁棒性。为解决这些问题，提出一种基于卷积神经网络(convolutional neural network，CNN)的烟雾检测方法。使用图片归一化方式消除光照的影响，利用烟雾颜色检测烟雾候选区域，CNN自动提取烟雾候选区域的特征，进行烟雾识别，根据分类结果得到报警信号。针对烟雾产生初期烟雾区域相对较小的问题，利用扩大候选区域的策略提高烟雾检测的及时性。由于训练数据少或不平衡引起的过度拟合，使用数据增强技术从原始数据集生成更多训练样本解决该问题。实验结果表明，该方法能有效地检测烟雾，且具有更高的准确率和更好的鲁棒性。

Smoke detection based on convolutional neural network

Researchers have proposed many smoke detection methods, but most of them are based on hand-crafted features or use raw images directly as input to the neural network, which reduces the robustness of deep learning. To solve these problems, this paper proposes a smoke detection algorithm based on the convolutional neural networks (CNN). Firstly, the image normalization method is adopted to remove the influence of illumination, and the suspected smoke areas are detected using smoke color model. Then,CNN automatically extracts the features of the smoke candidate area for smoke recognition,and the alarm signal is obtained according to the classification results. Aiming at the problem that the smoke area is relatively small in the early stage of smoke generation,the strategy of expanding the suspected areas is used to improve the timeliness of smoke detection. In order to solve the overfitting caused by insufficient training samples or imbalance, more training samples are generated from the original data set by using data augmentation technology. The experimental results show that the method can effectively detect smoke, and has higher accuracy and better robustness.