基于混合卷积神经网络的火灾识别研究

Research on Fire Identification Based on Hybrid Convolutional Neural Network

图像识别是实现火灾预警的重要手段之一。针对传统方法存在的检测精度低、难以识别小目标等问题,提出了一种基于混合卷积神经网络(CNN)的火灾识别方法。为了丰富模型提取的特征信息,充分利用不同尺度下的特征,文中提出的混合网络结构(HybridNet)包含两路特征提取器。首先,通过其中一路特征提取器提取图像中的深层语义信息,另一路特征提取器提取图像的浅层上下文信息,通过池化操作使两路特征提取器提取的特征图大小得以匹配。为了进一步实现特征之间的融合,提高模型的小目标识别性能,通过自编码器对特征进行降维处理,剔除冗余信息保留关键特征,实现多尺度特征的融合。最后,融合特征经过分类器得到分类结果。实验结果表明,提出的混合CNN优于现有的识别方法,在FireDetectData和Mivia数据集上分别取得了96.82%和97.96%的准确率。

Image identification is one of the important means to realize fire early warning. A fire identification method based on hybrid convolutional neural network(CNN) is proposed to solve the problems of low detection accuracy and difficulty in identifying small targets in traditional methods. In order to enrich the feature information extracted by the model and make full use of the features at different scales,the hybrid network structure(HybridNet) proposed contains two-way feature extractor. Firstly,one of the feature extractors is used to extract the deep semantic information and the other one to extract the shallow context information of the image. The feature map size extracted by the two-way feature extractor is matched by the pooling operation. In order to further fuse different features and improve the small target recognition performance of the model,the features are reduced by the self-encoder to eliminate the redundant information and retain the key features to achieve the fusion of multi-scale features. Finally,the fusion feature is classified by the classifier. The experiment shows that the proposed hybrid CNN is better than the existing recognition methods,and the accuracy of 96.82% and 97.96% is achieved on the FireDetectData and Mivia datasets respectively.