摘要
针对森林火灾背景复杂,图像拟合方式欠佳,使特征描述不充分而导致卷积神经网络(CNN)识别率低,卷积核随机初始化导致迭代次数增多等问题,提出了一种三通道拟合的改进卷积神经网络林火识别算法。通过调整三原色(RGB)三通道比例拟合样本图像,寻求火焰和背景对比度最大的优化参数组合,以解决识别率降低的问题;同时采用主成分分析(PCA)算法初始化卷积核来改进模型,提高迭代速率。实验结果表明:所提算法能有效地提高识别率,加快迭代速度,识别率达98. 5%。
Aiming at the problem that background of forest fires is complex,image fitting mode is subpar,which make feature description is inadequate,resulting in low recognition rate of convolutional neural network( CNN) and random initialization of convolution kernel results in increasing iteration number. A three-channel fitting improved CNN forest fire identification algorithm is proposed. By adjusting the RGB three-channel ratio fitting sample images,the optimal combination of the contrast between the flame and the background is searched to solve the problem of recognition rate reducing. Principal component analysis( PCA) algorithm is used to initialize the convolution kernel to improve the model for enhancement of the iteration rate. The experimental results show that the algorithm can improve the recognition rate and boost the iteration speed,the recognition rate reaches 98. 5 %.
作者
张海波
赵运基
张新良
ZHANG Haibo;ZHAO Yunji;ZHANG Xinliang(College of Electrical Engineering and Automation,Henan University of Technology,Jiaozuo 454000,China)
出处
《传感器与微系统》
CSCD
2020年第11期134-136,140,共4页
Transducer and Microsystem Technologies
基金
国家自然科学基金面上资助项目(61573130)
河南省高等学校重点科研资助项目(16A413009)
河南省科技创新人才杰出青年资助项目(164100510004)。
关键词
森林火灾识别
参数拟合
卷积神经网络
主成分分析
卷积核
forest fire recognition
parameter fitting
convolutional neural network(CNN)
principal component analysis(PCA)
convolution kernel
