【作者】 康文；

【导师】 汪洪源；

【作者基本信息】 哈尔滨工业大学 ， 光学工程（专业学位）， 2019， 硕士

【摘要】 随着互联网技术的不断发展和大型存储设备不断升级,网络上的视频数量急剧增多,对视频处理技术的质量要求不断提高,如何利用计算机高速处理视频数据、识别视频信息,已成为一个亟待解决的问题。对视频信息的识别,主要是对视频内容的处理,由于以人体行为为主要内容的视频数量在互联网视频总量中占比很高,所以研究人体行为识别技术尤为重要。传统的人体行为识别方法主要依赖模式识别与匹配技术,不仅识别灵活性不高,而且需要耗费大量的人工成本。随着人工智能技术的兴起,基于人工智能技术的人体行为识别成为研究热点,本文为提高基于人工智能技术的人体行为识别方法上的识别准确率,降低人体行为识别算法的硬件要求和训练成本,主要开展以下研究内容:(1)通过迁移学习构建人体行为识别神经网络,基于UCF-101数据集对神经网络进行训练,利用非共性特征层的微调优化输出结果,得到人体行为识别模型。(2)为在不提高批处理步长的前提下,解决局部最优解的问题,引入基于FSRCNN的图像超分辨率重建方法,通过对原始图像的超分辨率重建,主动增加图像内包含的信息,使损失函数能够在低处理步长下迅速收敛,提高识别准确率的同时降低神经网络对硬件配置的依赖程度。(3)为解决人体行为识别模型训练速度缓慢的问题,本文引入FrFT视觉显著性检测方法,通过提前检测图像中的重点目标,将对全图像的识别转换为对重点目标的识别,抑制复杂背景对识别的影响,减少识别算法需要处理的信息,降低神经网络的训练时长。实验结果证明,本文使用的方法极大程度上提高了相同批处理步长下人体行为识别的准确率,同时降低了神经网络对硬件配置的要求,降低了对原始视频清晰度的要求以及缩减了神经网络的训练时间,为人体行为识别技术的发展提供了支撑。更多还原

【Abstract】 With the development of Internet technology and the upgrading of large storage devices,the number of video on the network have a explosive increase,meanwhile,the requirements of video processing technology are constantly improving.How to process video data and identify video information with computer efficiently has become an urgent problem needed to be solved.Because of most of the videos are based on human,it is particularly important to study human behavior recognition technology.Traditional human behavior recognition methods rely on pattern recognition and matching technology mainly,which not only has low recognition flexibility,but also requires a lot of labor costs.With the rise of artificial intelligence technology,human behavior recognition which based on artificial intelligence technology has become a research hotspot.In order to improve the recognition accuracy of human behavior recognition method which based on artificial intelligence technology and reduce the hardware requirements and training costs of human behavior recognition algorithm,this paper mainly carries out the following research contents:(1)Constructing human behavior recognition neural network by transfer learning.Training the neural network based on UCF-101 data set,and optimizing the output results by fine-tuning the non-common feature layer.Finally,obtaining the human behavior recognition model.(2)In order to solve the problem of local optimal solution without increasing the batch processing step size,a super-resolution image reconstruction method based on FSRCNN is introduced.Through super-resolution reconstruction of the original images,the information contained in the original images is increased,so that the loss function can converge rapidly with low processing step size,by what improving the recognition accuracy and reduce the dependence of network on hardware configuration.(3)In order to solve the problem of slow training speed of human behavior recognition model,FrFT visual saliency detection method is introduced in this paper.By detecting the key targets in the images in advance,the recognition of the whole images will be converted to the recognition of the key targets.In this method,the influence of complex background on the recognition will be restrained,and the information that the recognition algorithm needs to process will be reduced,Finally,reducing the training time of the neural network.The experimental results shown that these methods greatly improved the accuracy of human behavior recognition under the same batch processing step,and reduced the requirement of hardware configuration,the requirement of original videos’ clarity and the training time of the neural network,which provides support for the development of human behavior recognition technology.更多还原