摘要
针对传统的机器学习算法以及常规的深度学习模型对于大数据量的航运监控视频识别效果不佳的问题,设计一种组约束深度神经网络模型(GCDNN)对实时航运监控视频进行识别。模型主要由结合Inception结构的VGG-16组件和优化LSTM单元的深层双向循环神经网络DBO-LSTM组件构成,充分提取视频帧序列的时空特征,使用稀疏组套索正则化算法进行网络稀疏处理,使用随机森林算法输出分类结果。实验结果表明,所提模型可以较好提升大数据量下的视频识别准确率,对于受恶劣天气影响的数据具有较强的识别能力,验证了所提模型的有效性。
Aiming at the problem that the traditional machine learning algorithm and the conventional deep learning model have poor effects on large-volume shipping surveillance video recognition,a group-constrained deep neural network model(GCDNN)was designed to identify real-time shipping surveillance video.The model was mainly composed of the VGG-16 component combined with the Inception structure and the deep bidirectional cyclic neural network DBO-LSTM component of the optimized LSTM unit.After fully extracting the spatio-temporal features of the video frame sequence,the sparse group lasso regularization algorithm was used for network sparse processing.The classification result was outputted using a random forest algorithm.Experimental results show that the proposed model can improve the video recognition accuracy under large data volume and has strong recognition ability for the data affected by bad weather,which verifies the validity of the proposed model.
作者
明道睿
张鸿
MING Dao-rui;ZHANG Hong(College of Computer Science and Technology,Wuhan University of Science and Technology,Wuhan 430065,China;Hubei Province Key Laboratory of Intelligent Processing and Real-Time Industrial Systems,Wuhan University of Science and Technology,Wuhan 430065,China)
出处
《计算机工程与设计》
北大核心
2020年第10期2949-2957,共9页
Computer Engineering and Design
基金
国家自然科学基金项目(61373109)。
关键词
航运监控
视频识别
神经网络
稀疏组套索
随机森林
shipping monitoring
video recognition
neural network
sparse group lasso
random forest
