基于深度迁移学习的烟雾识别方法

发布时间：2018-10-05 13:56

【摘要】：针对传统的基于传感器和图像特征的烟雾识别方法易被外部环境干扰且识别场景单一,从而造成烟雾识别精度较低,而基于深度学习的识别方法对数据量要求较高,对于烟雾数据缺失或数据来源受限的情况模型识别能力较弱的问题,提出一种基于深度迁移学习的烟雾识别方法。将ImageNet数据集作为源数据,利用VGG-16模型进行基于同构数据下的特征迁移。首先,将所有的图像数据进行预处理,对每张图像作随机变换(随机旋转、剪切、翻转等);其次,引入VGG-16网络,将其卷积层特征进行迁移,并连接预先使用烟雾数据在VGG-16网络中训练过的全连接层;进而构建出基于迁移学习的深度网络,从而训练得到烟雾识别模型。利用公开数据集以及真实场景烟雾图像进行实验验证,实验结果表明,和现有主流烟雾图像识别方法相比,所提方法有较高的烟雾识别率,实验精度达96%以上。
[Abstract]:The traditional smoke recognition method based on sensor and image features is easy to be interfered by the external environment and the recognition scene is single, which results in the low accuracy of smoke recognition, while the recognition method based on depth learning requires a high amount of data. A smoke recognition method based on deep migration learning is proposed to solve the problem of weak recognition ability of the model with missing smoke data or limited data source. The ImageNet data set is used as the source data and the VGG-16 model is used for feature migration based on isomorphic data. Firstly, all the image data are preprocessed, and each image is transformed randomly (random rotation, cut, flip, etc.). Secondly, the VGG-16 network is introduced to transfer the convolution layer features. The full connection layer trained in VGG-16 network with smoke data is connected, and then a deep network based on migration learning is constructed, and the smoke recognition model is obtained by training. The experimental results show that the proposed method has a higher smoke recognition rate than the existing mainstream smoke image recognition methods, and the experimental accuracy is over 96%.
【作者单位】：河南师范大学计算机与信息工程学院;河南省高校计算智能与数据挖掘工程技术中心;
【基金】：国家自然科学基金资助项目(U1204609) 河南省高校科技创新人才支持计划(15HASTIT022) 河南省高校青年骨干教师资助计划(2014GGJS-046) 河南师范大学优秀青年科学基金资助项目(14YQ007) 河南省高等学校重点科研项目计划(15A520078) 河南省科技攻关项目(172102210333)~~
【分类号】：TP181;TP391.41

【相似文献】