摘要
面向海量视频数据的实时识别需求,提出了基于循环神经网络的分布式自适应视频识别框架(AVRNN),该框架能自适应地选择相关帧以进行实时、快速视频识别。构建了增强记忆的长短期记忆网络,结合全局上下文信息,实现视频帧的实时选择。提出基于策略梯度的训练方法,根据生成的预测,确定接下来要观察的视频帧,并设计了奖励函数来计算预期的效用。提出了基于效用的自适应前瞻推理策略,以降低框架的计算成本。最后使用真实数据集对该框架进行评估,实验结果表明,与现有的方法相比,该框架能有效提高识别精度,降低计算成本。
To meet the real-time recognition requirements of massive video data,a distributed adaptive video recognition framework based on recurrent neural network(AVRNN)is proposed,which can adaptively select relevant frames for fast real-time video recognition.The framework creates a long short-term memory network with enhanced memory to achieve real-time selection of video frames by combining global context.A training method based on policy gradient is proposed,the video frame to be observed next is determined according to the generated prediction,and a reward function is designed to calculate the expected utility.A utility-based adaptive forward-looking reasoning strategy is proposed to reduce the computational cost of the framework.Finally,the real data set is used to evaluate the framework.The experimental results show that compared with the existing methods,the framework can effectively improve the recognition accuracy and reduce the computational cost.
作者
胡六四
HU Liu-si(Software College,Anhui Vocational College of Electronics&Information Technology,Bengbu Anhui 233000,China)
出处
《萍乡学院学报》
2022年第6期41-45,共5页
Journal of Pingxiang University
基金
2020年安徽省高校优秀青年人才支持计划(gxyq2020142)
2021年度安徽高校自然科学研究重大项目(KJ2021DZ0172)
2021年安徽省质量工程项目(2021jxtd024)。
关键词
神经网络
视频识别
自适应策略
neural network
video recognition
adaptive strategy
