基于数据增强的卷积神经网络火灾识别

当前图像识别采用的普遍方法是卷积神经网络方法,但该方法依赖于大数据集,在样本不足时会出现过拟合问题。针对以上问题,根据火灾的背景复杂性和卷积神经网络自动学习特征的优点,提出一种基于数据增强的卷积神经网络火灾识别方法。对少量火灾图片引入数据增强技术,通过搭建一个3层卷积池化层和一个全连接层自动提取火灾特征,使用softmax分类器输出。仿真实验结果表明:原始数据测试集的识别率为95%,损失值发散,提出方法使测试集损失值收敛到0.2,改善了过拟合的问题;对数据增强减少过拟合的原因进行分析,表明对小样本使用卷积神经网络具有重要意义。

Convolution Neural Network Based on Data Enhancement for Fire Identification

Currently, the common method for image recognition is the convolutional neural network method, but this method relies on big data sets, and over-fitting problems may occur in small samples. To solve above problems, In this paper, according to the background complexity in fire occurrence and the advantages of automatic learning features of convolutional neural network, a convolutional neural fire identification method based on data enhancement is proposed,Data enhancement technology was introduced for a few fire pictures, and fire features were automatically extracted by constructing a 3-layer convolutional pooling layer and a full connection layer, and finally the softmax classifier was used for output. The simulation results show that although the recognition rate of the original data test set is 95%, the loss value diverges. The method proposed in this paper makes the loss value of the test set converge to 0.2, which increases the reliability of the test recognition rate and is of great significance for the image recognition of small samples in convolutional neural network.