基于卷积神经网络的火焰识别  

Flame recognition based on convolution neural network

作　　者：段锁林[1] 刘福 高仁洲 王一凡[2] 潘礼正[1] DUAN Suo-lin;LIU Fu;GAO Ren-zhou;WANG Yi-fan;PAN Li-zheng(School of Mechanical Engineering,Changzhou University,Changzhou 213164,China;Mechanical and Electrical College,Changzhou Vocational Institute of Textile and Garment,Changzhou 213164,China)

机构地区：[1]常州大学机械工程学院,江苏常州213164 [2]常州纺织服装职业技术学院机电学院,江苏常州213164

出　　处：《计算机工程与设计》2019年第11期3288-3292,3298,共6页Computer Engineering and Design

基　　金：江苏省科技支撑计划基金项目(社会发展)(BEK2013671);江苏省高等学校自然科学研究面上基金项目(18KJB460001)

摘　　要：针对卷积神经网络在火焰识别应用中受复杂环境背景影响无法充分提取火焰特征的问题,提出一种多通道输入结合策略。在网络输入层,根据LMDB数据源制作过程中提供的火焰区域坐标获取火焰区域,对火焰区域的RGB这3个通道图像分别做灰度和二值化处理,结合形成9通道的三维数据作为网络的输入;提出一种改进的Relu激活函数,使用两个参数分别控制正负区域斜率,弥补原Relu函数负区域为0强制引入稀疏性的缺点,通过减小Relu正区域斜率,平衡特征数量,降低过拟合风险。实验中重新构建卷积神经网络模型,设置网络参数,通过实验获取最佳Relu改进参数,实验结果表明,该方法对火焰识别精度有显著提高。For the problem that the convolution neural networks cannot fully extract flame characteristics due to complex environmental backgrounds in flame recognition applications,a multi-channel input strategy was proposed.In the input layer of the network,the flame region was obtained according to the coordinates of the flame region provided in the process of making the LMDB data source,three channel images of RGB in flame region were processed with gray scale and binarization respectively,and the3 Ddata resulted from the combination of 9 channels played the role of the input of the network.An improved Relu activation function was proposed,in which two parameters were used to control the slope of positive and negative regions,respectively.It not only made up for the disadvantage of introducing sparsity into the negative region of the original Relu function for 0,but also balanced the number of features and reduced the risk of overfitting by reducing the slope of the positive region of Relu.In the experiment,the convolution neural network model is re-constructed,the network parameters are set up and the best Relu improved parameters are obtained.Experimental results show that the proposed method can improve the accuracy of flame recognition significantly.

关 键 词：卷积神经网络 特征提取 火焰识别 多通道输入 Relu函数 

分 类 号：TP39[自动化与计算机技术—计算机应用技术]