一种基于帧序列特征的三流网络人体行为识别方法

随着计算机科学和深度学习技术的发展,人体行为识别研究逐渐成为计算机视觉的一个重要课题。目前主流的双流网络模型无法做到在提取图像和运动特征的同时提取视频的帧间序列特征,当局部序列特征与长短时运动特征发生时空交互时,双流网络模型鲁棒性严重降低。针对于此,提出了一种基于视频序列特征的三流网络人体行为识别方法。通过预处理将视频的稠密光流帧输入时间网络,RGB帧输入空间网络和帧序列特征提取网络,同时对三个网络进行预训练。网络输出其对应的特征后使用权重相加的融合方法进行特征融合,最后采用多层感知机得到行为分类结果。将该方法分别在UCF11、UCF50和HMDB51数据集进行实验,得到行为分类准确率分别为99.17%、97.40%和96.88%。与传统的双流网络方法相比,该方法有效综合了行为的空间信息,时间信息和帧序列信息,识别准确率得到较大提升,具有更强的泛化能力。

基于多模态注意力网络的红外人体行为识别方法

深度学习网络对红外单一模态数据的学习表征能力具有一定的局限性,针对该问题,文中提出了基于多模态注意力网络的红外人体行为识别方法。由于深度学习网络模型无法直接对视频信息进行训练和分类,首先,通过预处理模块将得到的视频信息预处理成红外视图,再将得到的红外视图通过Sobel算子和基于L 1范数的全变分光流法分别提取红外视图的边缘信息和光流信息得到边缘视图和光流视图;其次,将红外视图、边缘视图、光流视图分别输入融合注意力机制模块的三流网络中进行特征学习;然后,对三流网络中每个网络提取的多模态特征进行融合;最后,将融合得到的特征向量输入随机森林进行训练和分类。在公开数据集NTU RGB+D和自建数据集上进行实验,结果表明了所提方法具有不错的识别效果。

基于三流卷积神经网络模型的图像分类方法

为了提高卷积神经网络在提取图像特征的充分性与有效性,提出了一种基于三流卷积神经网络模型的图像分类方法.第一个和第二个网络流的特征提取部分采用交叉“间隔”的方式训练提取图像的不同特征,第三个网络流的特征提取部分采用初始参数,以此来构建三流卷积神经网络模型,提取到更充分有效的图像特征.同时针对每个网络流训练一个分类器,然后运用分类器融合算法对每个网络流的分类器赋予不同权重,得到3个网络流的融合输出,实现最终的分类.在CIFAR-100、Stanford Dogs和UEC FOOD-100和数据集上的实验结果验证了该方法的有效性和鲁棒性.

注意力引导的三流卷积神经网络用于微表情识别

目的微表情识别在心理咨询、置信测谎和意图分析等多个领域都有着重要的应用价值。然而,由于微表情自身具有动作幅度小、持续时间短的特点,到目前为止,微表情的识别性能仍然有很大的提升空间。为了进一步推动微表情识别的发展,提出了一种注意力引导的三流卷积神经网络(attention-guided three-stream convolutional neural network,ATSCNN)用于微表情识别。方法首先,对所有微表情序列的起始帧和峰值帧进行预处理;然后,利用TV-L1(total variation-L1)能量泛函提取微表情两帧之间的光流;接下来,在特征提取阶段,为了克服有限样本量带来的过拟合问题,通过3个相同的浅层卷积神经网络分别提取输入3个光流值的特征,再引入卷积块注意力模块以聚焦重要信息并抑制不相关信息,提高微表情的识别性能;最后,将提取到的特征送入全连接层分类。此外,整个模型架构采用SELU(scaled exponential linear unit)激活函数以加快收敛速度。结果本文在微表情组合数据集上进行LOSO(leave-one-subject-out)交叉验证,未加权平均召回率(unweighted average recall,UAR)以及未加权F1-Score(unweighted F1-score,UF1)分别达到了0.7351和0.7205。与对比方法中性能最优的Dual-Inception模型相比,UAR和UF1分别提高了0.0607和0.0683。实验结果证实了本文方法的可行性。结论本文方法所提出的微表情识别网络,在有效缓解过拟合的同时,也能在小规模的微表情数据集上达到先进的识别效果。

Integrated logistics facilities network design for 3PLS under uncertainty

According to the operational characteristics of the logistics networks for the third party logistics supplier （3PLS）, the forward and reverse logistics networks together for 3PLS under the uncertain environment are designed. First, a fuzzy model is proposed by taking multiple customers, multiple commodities, capacitated facility location and integrated logistics facility layout into account. In the model, the fuzzy customer demands and transportation rates are illustrated by triangular fuzzy numbers. Secondly, the fuzzy model is converted into a crisp model by applying fuzzy chance constrained theory and possibility theory, and one hybrid genetic algorithm is designed for the crisp model. Finally, two different examples are designed to illustrate that the model and solution discussed are valid.

A MOF derived hierarchically porous 3D N‐CoP_(x)/Ni_(2)P electrode for accelerating hydrogen evolution at high current densities

Hydrogen evolution reaction is a critical reaction in water splitting for hydrogen production.However,developing effective and stable non‐noble‐metal electrocatalysts which work well at high current densities demanded by industry still remain great challenge.Herein,taking advantage of the highly tunable metal‐organic framework(MOF)templates,nitrogen doped binary transition metal phosphides electrocatalysts(N‐CoP_(x)/Ni_(2)P)with three‐dimensional(3D)conductive network structure were successfully synthesized.The 3D open porous channels could expose more catalytically active sites;nitrogen doping and the synergistic effect between CoP and Ni_(2)P can increase the electron density of Co atoms at active sites,further optimizing the Gibbs free energy of hydrogen(ΔGH*)and water(ΔG_(H_(2)O*)).As a result,the obtained N‐CoP_(x)/Ni_(2)P catalyst exhibits extraordinary electrocatalytic activity in a wide pH range.Especially,it requires an extremely low overpotential of 152 mV to deliver a high current density of 650 mA cm^(–2) in alkaline media.This work may shed some light on the rational design of cheap electrocatalysts and electrode materials that work well at high current densities.

Structure of Chinese City Network as Driven by Technological Knowledge Flows

Based on patent cooperation data,this study used a range of city network analysis approaches in order to explore the structure of the Chinese city network which is driven by technological knowledge flows.The results revealed the spatial structure,composition structure,hierarchical structure,group structure,and control structure of Chinese city network,as well as its dynamic factors.The major findings are:1) the spatial pattern presents a diamond structure,in which Wuhan is the central city;2) although the invention patent knowledge network is the main part of the broader inter-city innovative cooperation network,it is weaker than the utility model patent;3) as the senior level cities,Beijing,Shanghai and the cities in the Zhujiang(Pearl) River Delta Region show a strong capability of both spreading and controlling technological knowledge;4) whilst a national technology alliance has preliminarily formed,regional alliances have not been adequately established;5) even though the cooperation level amongst weak connection cities is not high,such cities still play an important role in the network as a result of their location within ′structural holes′ in the network;and 6) the major driving forces facilitating inter-city technological cooperation are geographical proximity,hierarchical proximity and technological proximity.

A General Study on Equivalent Resistance of Three-Terminal Ladder Network

In view of the application importance of resistance network in modern science and technology, this paper presents the basic structure of a three terminals ladder shaped resistance network, for which, to study in- depth the equivalent resistance, carry out network analysis by applying virtual current method and construct a model of two elements three orders differential equation. Based on different marginal conditions, two general adaptive rules for the three-terminal ladder shaped inlet resistance, as well as two ultimate rules for the equiva- lent resistance of three-terminal infinite ladder shaped were given.

One-pot ball-milling preparation of graphene/carbon black aqueous inks for highly conductive and flexible printed electronics

Stable aqueous carbon inks,with graphene sheets(GSs)and carbon black(CB)as conductive fillers,are prepared by a simple one-pot ball-milling method.The asprepared composite ink with 10 wt%GSs shows optimized rheological properties(viscosity and thixotropy)for screen printing.The as-printed coatings based on the above ink are uniform and dense on a polyimide substrate,and exhibit a sandwich-type conductive three dimensional network at the microscale.The resistivity of the typical composite coating is as low as 0.23±0.01Ωcm(92±4Ωsq^-1,25μm),which is 30%as that of a pure CB coating(0.77±0.01Ωcm).It is noteworthy that the resistivity decreases to 0.18±0.01Ωcm(72±4Ωsq^-1,25μm)after a further rolling compression.The coating exhibits good mechanical flexibility,and the resistance slightly increases by 12%after 3000 bending cycles.With the CB/GSs composite coatings as a flexible conductor,fascinating luminescent bookmarks and membrane switches were fabricated,demonstrating the tremendous potential of these coatings in the commercial production of flexible electronics and devices.