为了深度卷积神经网络能够更好地应用于边缘设备,减少深度神经网络的模型参数量,降低网络复杂度,对于轻量化卷积神经网络的研究日益增多.将常用于多尺度融合的卷积——空洞卷积首次应用于神经网络轻量化领域研究.利用空洞卷积采样与扩大感受野的结构特性,实现了模型参数量的减少与计算复杂度的降低.将空洞卷积单纯作用于残差网络结构,达到轻量化目的,并进一步与逐点卷积结合,提高轻量化效果,形成改进型空洞卷积轻量化方法.为减少准确率衰减,将改进型空洞卷积与普通卷积相融合,提出一种融合型空洞卷积轻量化方法.实验结果表明,改进型空洞卷积轻量化方法具有最显著的轻量化效果,融合型空洞卷积轻量化方法使模型参数量减少同时具有最佳速度与精度的权衡.

基于骨架数据的动作识别方法由于其对动态环境和复杂背景的强适应性而受到广泛的关注和研究,应用图卷积网络描述人体骨架实现人体动作识别可以取得很好的识别效果,但实现过程中图的拓扑结构通常是手动设置的,且在所有层和输入样本上的结构固定,只能捕获关节之间的局部物理关系,会遗漏非物理连接的关节相关性.提出了一种新的基于区域关联自适应图卷积网络的骨架动作识别,通过自适应图卷积使参数化的全局图和单个数据图的结构与模型卷积参数在不同的层中分别进行训练和更新,增加了模型中图形构造的灵活性与模型对于各种数据样本的通用性.同时引入区域关联图卷积,通过在关节特征与连接特征之间交替信息传递来捕获数据帧间各关节的非物理连接相关性.并加入骨骼的二阶数据对原有关节数据进行信息补充,融合两者构成双流网络提升识别网络的性能.在NTU-RGBD大规模数据集上的实验表明,该模型在动作识别的准确率上有了一定的提升.

针对单幅深度图像三维手姿估计中由于手部复杂结构捕捉困难导致的精度低和鲁棒性较差的问题,提出一种基于改进PointNet网络的三维手姿估计方法.该方法首先采用边界框定位网络预测三维边界框,从而准确裁剪手部区域.然后将手部深度图像表示为点云,模拟手部可见表面,有效地利用深度图像中的三维信息.最后将手部点云数据输入改进的PointNet网络,准确地进行三维手姿估计.改进的PointNet网络通过引入跳跃连接,充分利用不同层次的特征,更好地捕捉手部的复杂结构.在NYU手姿数据集上进行验证,实验结果表明,提出的方法优于现有的大部分方法,并且网络结构简单、易于训练,运行速度快.

面部微表情具有持续时间短、强度低等特点,因此微表情识别的准确率较低,且当前提出的基于深度学习的微表情识别方法难以理解网络模型的决策原因,以至于难以应用于实际.针对该问题提出一种基于残差单元的可解释时空卷积网络用于微表情识别的方法,该方法通过使用时空卷积网络,将微表情视频帧序列作为输入,并解释网络模型参数与特征,在实现模型可解释性的同时,提高了微表情识别的准确率.在CASME2、SMIC和SAMM数据集上对提出方法进行验证,实验结果表明,本文方法优于目前大部分基于深度学习的微表情识别方法.

To overcome the inaccuracy of depth estimation value caused by fuzzy contour information in single image depth estima-tion,this paper proposes a single image depth estimation method using Residual Dense Network. This method proposes a new neural-network model by introducing the Residual Dense Model into the encoder decoder structure with Skip connection. Unsupervised train-ing of neural networks is achieved by using a series of stereo image pairs from binocular cameras. The disparity map is obtained by in-putting the predicted image into the network model,and according to the geometric relationship between the disparity map and the depth map,the depth map of the image is obtained. The network model proposed in this paper is trained on the KITTI driving dataset, and the results show that the method obtains lower error value,higher accuracy and clearer edge contour information of the object than most of the existing methods on the test set. It proves the effectiveness and superiority of this paper in single image depth estimation.

为解决复杂背景的手势分割问题,提出一种基于融合皮肤检测的卷积姿势机手势分割方法.通过两个CNN网络得到训练的手势分割部分和皮肤分割部分,通过最后一阶段的CNN网络输出最终的手势分割图像,皮肤分割的准确性对最终分割图像起辅助作用,其中核心部分即手势分割部分采用卷积姿势机网络,并运用中间监督的思想将皮肤信息融合.该网络将手势轮廓和经皮肤提取的手势细节结合,分别对轮廓、皮肤、融合3个子网络进行训练,结果对比提取手势的其它方法,验证了该方法的有效性.

Google Brain proposes a simple model for neural machine translation,the Transformer,based solely on attention mecha-nisms,dispensing with recurrent neural network entirely,which can solve the problem that RNN can′t be parallelized. This paper pro-poses a Transformer model improvement method by word-level weights,the weights are obtained through neural network self-learning according to the importance of different words in a sentence and combining the weights with the Transformer. This method improves the stability and accuracy of the Transformer and makes this model achieve a higher BLEU score on the IWSLT16 translation task.

最近几年在深度学习领域中,自动生成一副图像的自然语言描述引发了学界的广泛关注,原因是图像描述在实际应用中的重要性以及它连接了两个重要的人工智能领域:计算机视觉和自然语言处理.以往的模型大多采用基于模板或简单的编码-解码方式,生成的文本结构较为单一并且不能够根据图像中各个物体的相互关系表达出图像的深层意义.提出了一种基于注意力机制与多模态的图像描述方法,在LSTM( Long-Short Term Memory)的基础上改进了Attention机制,并在Attention结构后面添加了多模态层对图像的上下文特征信息以及LSTM的隐层状态进行融合处理.在两个公共数据集:MS COCO以及Flickr 30K上进行验证,实验结果证明所提方法有效且可以使生成的描述语句更加丰富.

在图像识别与计算机视觉领域,物体检测是研究热点,提出了一种融合多尺度特征的多物体检测方法,基于卷积神经网络在多尺度特征下提取物体的候选区域,然后将不同尺度下的特征进行融合,使多物体检测中出现的小物体被漏检的概率降低.最后采用基于中心点的非极大值抑制方法,计算检测窗口的中心点的欧式距离和iou来抑制冗余的窗口,从而提升多物体检测的精度.将提出的方法在PASCAL VOC数据集上进行验证,实验证明所提的方法能有效提高多物体检测的精确度.

深度学习作为人工神经网络的分支,在图像识别领域有广泛的应用,但其数据集的不足导致模型学习不够完善.通过对深度学习的数据规模要求进行分析,针对人体行为识别中的应用,发现人体数据集的采集工作是一个极具耗时耗力的工程,很难满足目前深度学习网络的需求.为了解决这一难题,提出了一种依靠原有的小规模数据集产生大量可靠数据集的半监督深度学习模型.通过将循环神经网络和生成式对抗网络相结合的方法使循环神经网络学习到数据的序列关系和特征,使生成式对抗网络产生合理数据进而扩展人体行为数据集.依靠该网络结构,可以很好地分析出采集数据的特征,并且依据这些特征可以生成大量的合理的数据,后经过数据处理等工作,形成可用于模型训练的可靠数据集,缓解了深度学习工作中数据集紧缺的问题.

人体行为识别是计算机视觉研究的热门领域之一,提出了一种基于隐藏语义的人体行为算法,采取人体骨骼点的三维数据进行处理后构成数据序列,将该数据序列作为分析人体行为的主语义的同时,分析人体行为细节信息在时空特性上的变化作为隐藏语义,然后将二者融合作为人体行为表示.最后使用改进的类均值核主成分分析算法对行为表示数据进行处理,并用支持向量机进行分类.将提出的方法在UTKinect、Florence和MSR Action 3D数据集上进行验证,实验结果证明了所提方法的有效性和普适性.

创新创业教育是高水平大学人才培养的重要使命.针对高校创新创业培养中存在的问题,文章将创新创业教学与现有高校教学模式相结合,对通识教育模式下创新创业教育体系、课程体系、创新创业平台建设、师资队伍建设等方面进行了探索、实践,为地方性高校创新创业人才培养提供了借鉴.

This paper combinesthe internet technology with the experimental center teaching management and evaluation system.According to the characteristics of the open teaching of the university experimental center,the framework of the teaching management and evaluation system is constructed by using the browse/server and client/server hybrid mode.On this basis of this,through the addition of the course management,experiment preparation,experimental evaluation,test library editing and other content,an Internet-based experimental teaching management and evaluation system was established.It has been proved by practice that the system realizes the organic integration of experimental teaching and modern information technology,and integrates modern educational thoughts and concepts,and informational teaching content and methods into the practical teaching process,which satisfies the needs of into the practical teaching process,which satisfies the needs of experimental teaching.

为适应应用型专业转型的需要,促进电子信息类专业课程向应用型课程转化,文章以"电子系统设计"课程为例,从专业课程的课程理念、教学内容、教学模式、考核方式等方面对适应于应用型人才教育的专业课程改革模式进行了探索,给出了相应的改革内容,为其他应用型课程改革提供了参考.

To overcome the high complexity and the large computation of the recovery scenarios depth information algorithm using image high-level cues,a depth imformation extration algorithm for the single viewimage using profile sharpness of image low-level cues was proposed. In this paper,the blur information is considered as depth extration constraints. It establish an improved model of contour tracking with the edge contour sharpness information and consequently the coutour is extrated. A prior hypothesis of depth gradient is used to assign depth to extrate the scenarios depth information. Avoiding the interferences by image noise and the errors with the local optimum of the contour tracking,the depth map is optimized by the cross bilateral filter. Experimental results on a variety of images showthat a liable extration of the depth for the single viewimage can be acquireared availably with this simple but effective algorithm.

It is important to have a good professional cognitive teaching for undergraduates who are in electronic information specialty .It can help them to study professional skills and develop the ability of innovation and practice .This paper analyzes the professional cognitive teaching course of Introduction to Electrical and Systems Engineering in University of Pennsylvania ,USA .By studying the course content ,teaching means , experimental methods and assessment methods ,combining the development of cognitive teaching in China ,this paper explores the cultivation of the innovative ability and practical ability for the students .

Electronic design competition promotes the reform of experimental teaching for the undergraduate and the construction of the laboratory.According to the characteristics of electronic design competition and the current status of undergraduate experiment teaching,it proposes a experimental teaching model which integrates the development of the electronic design competition and the experimental teaching.The model is suitable for the cultivation of innovative talents.This paper demonstrates its applications on teaching philosophy of the experiment,experimental teaching system,experimental teaching content,experimental teaching methods and the experimental teaching environment.Through the practice,the applications get good results.

本文对利用VHDL语言设置有限状态机控制器的过程进行系统的论述.通过对控制器控制对象的时序分析,抽象出控制器的行为描述,并划分控制器的状态.在此基础上,针对ADC0809模数转换器的控制器进行设计,并通过了系统仿真和逻辑分析仪测试.从而给出了利用VHDL语言设计有限状态机控制器的一种方法.

通过对ELANTEC公司的EL5X20 CMOS Rail-to-Rail运算放大器的版图结构、电路原理进行分析,利用UMC公司的hspice leve149(sim3.3)0.6um N阱双多晶双金属高压工艺MODEL进行运算放大器参数仿真拟合,研究了国外的先进放大器设计方法,为高性能放大器研制奠定了基础.

根据FPGA系列芯片的特点,提出在基于FPGA的数字系统设计中应用微程序技术实现控制器的设计思想,介绍了微程序设计方法的相关原理和技巧.