带特征监控的高维信息编解码端到端无标记人体姿态估计网络

针对点云空间三维信息非结构化和旋转易变性对预测结果的影响,提出一种带特征监控的三维信息编解码卷积神经网络,该网络可实现三维空间下单目深度图的端对端无标记人体姿态估计.所设计的网络由特征监控编解码组件串联而成,该组件第一部分使用三维卷积模块以类似沙漏结构的形式组合设计,实现对特征图的编码和解码;第二部分以不同参数残差块并联,实现对特征图的监控融合,第一部分与第二部分首尾连接构成组件.特征监控编解码组件能根据数据集大小,通过串联的方式搭建不同深度的网络,同时根据数据分辨率,设置组件参数,实现由粗到精的特征学习,最终获得最佳网络.通过ITOP数据库的实验表明,该网络实现了空间三维信息的端到端深度学习,显著提高了系统性能并具有更高的精度.     

基于多尺度特征融合和注意力机制的医学图像分割网络

针对传统编解码结构的医学图像分割网络存在特征信息利用率低、泛化能力不足等问题,该文提出了一种结合编解码模式的多尺度语义感知注意力网络(multi-scale semantic perceptual attention network,MSPA-Net) 。首先,该网络在解码路径加入双路径多信息域注意力模块(dual-channel multi-information domain attention module,DMDA) ,提高特征信息的提取能力;其次,网络在级联处加入空洞卷积模块(dense atrous convolution module,DAC) ,扩大卷积感受野;最后,借鉴特征融合思想,设计了可调节多尺度特征融合模块 (adjustable multi-scale feature fusion,AMFF) 和双路自学习循环连接模块(dual self-learning recycle connection module,DCM) ,提升网络的泛化性和鲁棒性。为验证网络的有效性,在CVC-ClinicDB、ETIS-LaribPolypDB、COVID-19 CHEST X-RAY、Kaggle_3m、ISIC2017和Fluorescent Neuronal Cells等数据 集上进行验证,实验结果表明,相似系数分别达到了94.96%、92.40%、99.02%、90.55%、92.32%和75.32%。因此,新的分割网络展现了良好的泛化能力,总体性能优于现有网络,能够较好实现通用医学图像的有效分割。     

基于注意力门控膨胀卷积网络的单通道语音增强

在有监督语音增强任务中,上下文信息对目标语音的估计产生重要影响,为了获取更加丰富的语音全局相关特征,该文以尽可能小的参数为前提,设计了一种新型卷积网络来进行语音增强。所提网络包含编码层、传输层与解码层3个部分:编解码部分提出一种2维非对称膨胀残差(2D-ADR)模块,其能明显减小训练参数并扩大感受野,提升网络对上下文信息的获取能力;传输层提出一种1维门控膨胀残差(1D-GDR)模块,该模块结合膨胀卷积、残差学习与门控机制,能够选择性传递特征并获取更多时序相关信息,同时采用密集跳跃连接的方式对8个1D-GDR模块进行堆叠,以增强层间信息流动并提供更多梯度传播方式;最后,对相应编解码层进行跳跃连接并引入注意力机制,以使解码过程获得更加鲁棒的底层特征。实验部分,使用了不同的参数设置以及对比方法来验证网络的有效性与鲁棒性,通过在28种噪声环境下训练及测试,相比于其他方法,该文方法以1.25×106的参数取得了更优的客观和主观指标,具备较强的增强效果与泛化能力。     

基于卷积循环网络与非局部模块的语音增强方法

现有的深度神经网络语音增强方法忽视了相位谱学习的重要性,从而造成增强语音质量不理想。针对这一问题,文中提出了一种基于卷积循环网络与非局部模块的语音增强方法。通过设计一种编解码网络,将语音信号的时域表示作为编码端的输入进行深层特征提取,从而充分利用语音信号的幅值信息以及相位信息。在编码端和解码端的卷积层中加入非局部模块,在提取语音序列关键特征的同时,抑制无用特征,并引入门控循环单元网络捕捉语音序列间的时序相关性信息。在ST-CMDS中文语音数据集上实验结果表明,与未处理的含噪语音相比,使用文中方法生成的增强语音质量和可懂度平均提升了61%和7.93%。     

多线激光光条图像缺陷分割模型研究

受环境干扰以及反射光影响，室外采集的多线激光光条图像含有光斑和断裂缺陷.为了准确地分割图像缺陷，本文提出了一个轻量的UT(U-shape Target,U代表U型编解码网络结构，T代表靶形视野)分割模型，模型由3×3卷积和靶形卷积堆叠而成.靶形卷积是针对激光光条图像特点提出的多视野卷积模块，模块中四个卷积分支构成靶形卷积视野，能够提取激光光条图像几何结构特征、局部细节特征以及环绕纹理特征.实验表明，UT模型在多线激光光条图像上的缺陷分割精度高于主流分割模型，而且实现了分割精度和参数量的平衡.     

基于门控残差卷积编解码网络的单通道语音增强方法

针对卷积编解码网络（CED, Convolution encoder-and-decoder）对语音时序相关信息捕获困难的问题,本文提出了一种基于门控残差卷积编解码网络的语音增强方法。该方法在卷积编解码网络的基础上引入了门控机制、膨胀卷积与残差连接:门控机制能够很好地处理序列前后相关信息;膨胀卷积使得卷积过程获得更大的感受野,提取更加丰富的全局信息;残差连接能够防止梯度消失与梯度爆炸,提升网络精度。此外,采用频域损失函数与时域评价指标联合优化的策略对网络进行训练,以进一步提升网络增强效果。实验表明,在匹配噪声和不匹配噪声下,相比于基线CED与其他对比方法,本文方法取得了更高的PESQ、STOI与SI-SDR,对语音的清浊音都有较好恢复效果,且具有较强的泛化能力。      

基于语义导向的光场图像深度估计

光场图像的深度估计是3维重建、自动驾驶、对象跟踪等应用中的关键技术。然而,现有的深度学习方法忽略了光场图像的几何特性,在边缘、弱纹理等区域表现出较差的学习能力,导致深度图像细节的缺失。该文提出了一种基于语义导向的光场图像深度估计网络,利用上下文信息来解决复杂区域的不适应问题。设计了语义感知模块的编解码结构来重构空间信息以更好地捕捉物体边界,空间金字塔池化结构利用空洞卷积增大感受野,挖掘多尺度的上下文内容信息;通过无降维的自适应特征注意力模块局部跨通道交互,消除信息冗余的同时有效融合多路特征;最后引入堆叠沙漏串联多个沙漏模块,通过编解码结构得到更加丰富的上下文信息。在HCI 4D光场数据集上的实验结果表明,该方法表现出较高的准确性和泛化能力,优于所比较的深度估计的方法,且保留较好的边缘细节。     

基于异构卷积的轻量级图像分类网络

目前大多数大型神经网络都存在参数量大、计算难度高等问题,想要在移动端设备使用,则会受到计算资源的限制.虽然现有轻量级网络出现解决了一定的计算量的问题,但同时其网络中大量使用1×1点卷积,使得其成为了现在轻量级网络的计算瓶颈.针对点卷积造成的计算瓶颈的问题,首先提出使用GhostModel来代替其中一部分点卷积,然后结合异构卷积对残差结构进行改进并提出ResHetModel_A、B两个改进的模块,使用改进模块构成轻量级网络HSNet.最后对注意力特征图进行分析,在网络加入注意力机制来提高网络表达.在CAFIR10和CAFIR100数据集上的分类实验证明网络的有效性.最后在ImageNet大型数据集上实验表明HSNet具有一定的泛化性.     

基于深度残差神经网络的红外图像超分辨率重构算法

提出了一种基于深度残差神经网络的红外灰度图像超分辨率重构算法。首先使用残差卷积模块增加网络深度提高了网络的学习能力,使得卷积层在学习过程中能够利用到更多的邻域信息对于复杂场景有更好的学习能力。然后使用跳跃连接方式增加高频信息传输以实现对于图像细节的增强。实验结果表明,该网络能够有效地丰富重构图像的细节,重构图像中的目标轮廓有明显改善。     

基于形变卷积神经网络的行为识别

人类行为识别作为视频分类中的重要问题,成为计算机视觉中的热门话题.由于卷积神经网络(CNN)的几何结构固定统一,这将会使得其几何变形建模受限,使得行为识别网络难以鲁棒性的识别行为类别.本文提出了一种融入可形变卷积的行为识别网络模型.首先,引入可形变卷积,构建了一种可协同学习空间外观和时间运动线索的模块,该模块分别学习视频数据3个正交视图特征进行融合;其次,在ResNet网络的基础上,用该模块将其网络中部分关键性卷积模块进行替换,产生一种新颖的改进版本的3D-ResNet网络,用于视频数据集的训练和测试;最后,在UCF101和HMDB51数据集训练和测试,得到识别精度优于现有的大多数先进方法.     

Resolution Enhancement Technology:Integrated Process Solutions using OPC

DUV lithography, using the 248 nm wavelength, is a viable manufacturing option for devices with features at 130 nm and less. Given the low kl value of the lithography, integrated process development is a necessary method for achieving acceptable process latitude. The application of assist features for rule based OPC requires the simultaneous optimization of the mask, illumination optics and the resist.Described in this paper are the details involved in optimizing each of these aspects for line and space imaging.A reference pitch is first chosen to determine how the optics will be set. The ideal sigma setting is determined by a simple geometrically derived expression. The inner and outer machine settings are determined, in turn,with the simulation of a figure of merit. The maximum value of the response surface of this FOM occurs at the optimal sigma settings. Experimental confirmation of this is shown in the paper.Assist features are used to modify the aerial image of the more isolated images on the mask. The effect that the diffraction of the scattering bars (SBs) has on the image intensity distribution is explained. Rules for determining the size and placement of SBs are also given.Resist is optimized for use with off-axis illumination and assist features. A general explanation of the material' s effect is discussed along with the affect on the through-pitch bias. The paper culminates with the showing of the lithographic results from the fully optimized system.     

Cost-effective test solutions for smart card and other integrated flash applications

From its emergence in the late 1980s as a lower cost alternative to early EEPROM technologies, flash memory has evolved to higher densities and speedsand rapidly growing acceptance in mobile applications.In the process, flash memory devices have placed increased test requirements on manufacturers. Today, as flash device test grows in importance in China, manufacturers face growing pressure for reduced cost-oftest, increased throughput and greater return on investment for test equipment. At the same time, the move to integrated flash packages for contactless smart card applications adds a significant further challenge to manufacturers seeking rapid, low-cost test.     

An improved parallel thinning algorithm with two subiterations

The parallel thinning algorithm with two subiterations is improved in this paper. By analyzing the notions of connected components and passes, a conclusion is drawn that the number of passes and the number of eight-connected components are equal. Then the expression of the number of eight-connected components is obtained which replaces the old one in the algorithm. And a reserving condition is proposed by experiments, which alleviates the excess deletion where a diagonal line and a beeline intersect. The experimental results demonstrate that the thinned curve is almost located in the middle of the original curve connectivelv with single pixel width and the processing speed is high.     

Theoretical analysis of the relationship between the Brillouin gain coefficient and the strain in the optical-fiber sensors

The relation between the power of the Brillouin signal and the strain is one of the bases of the distributed fiber sensors of temperature and strain. The coefficient of the Bfillouin gain can be changed by the temperature and the strain that will affect the power of the Brillouin scattering. The relation between the change of the Brillouin gain coefficient and the strain is thought to be linear by many researchers. However, it is not always linear based on the theoretical analysis and numerical simulation. Therefore, errors will be caused if the relation between the change of the Brillouin gain coefficient and the strain is regarded as to be linear approximately for measuring the temperature and the strain. For this reason, the influence of the parameters on the Brillouin gain coefficient is proposed through theoretical analysis and numerical simulation.     

Novel fiber-optical interferometer with miniaturized probe for in-hole measurements

Today, micro-system technology and the development of new MEMS （Micro-Electro-Mechanical Systems） are emerging rapidly. In order for this development to become a success in the long run, measurement systems have to ensure product quality. Most often, MEMS have to be tested by means of functionality or destructive tests. One reason for this is that there are no suitable systems or sensing probes available which can be used for the measurement of quasi inaccessible features like small holes or cavities. We present a measurement system that could be used for these kinds of measurements. The system combines a fiber optical, miniaturized sensing probe with low-coherence interferometry, so that absolute distance measurements with nanometer accuracy are possible.     

Study of waveguide coupling efficiency in hybrid silicon lasers

This paper presents a new method to increase the waveguide coupling efficiency in hybrid silicon lasers. We find that the propagation constant of the InGaAsP emitting layer can be equal to that of the Si resonant layer through improving the design size of the InP waveguide. The coupling power achieves 42% of the total power in the hybrid lasers when the thickness of the bonding layer is 100 nm. Our result is very close to 50% of the total power reported by Intel when the thickness of the thin bonding layer is less than 5 nm. Therefore, our invariable coupling power technique is simpler than Intel＇s.     

Curved crystal spectrometer for the measurement of X-ray lines from laser-produced plasmas

In order to diagnose the laser-produced plasmas, a focusing curved crystal spectrometer has been developed for measuring the X-ray lines radiated from a laser-produced plasmas. The design is based on the fact that the ray emitted from a source located at one focus of an ellipse will converge on the other focus by the reflection of the elliptical surface. The focal length and the eccentricity of the ellipse are 1350 mm and 0.9586, respectively. The spectrometer can be used to measure the X- ray lines in the wavelength range of 0.2-0.37 nm, and a LiF crystal （200） （2d = 0.4027 nm） is used as dispersive element covering Bragg angle from 30° to 67.5°. The spectrometer was tested on Shengnang- Ⅱ which can deliver laser energy of 60-80 J/pulse and the laser wavelength is 0.35 μm. Photographs of spectra including the 1 s2p ^1P1-1s^2 ^1S0 resonance line（w）, the 1s2p ^3P2-1s^2 1S0 magnetic quadrupole line（x）, the 1s2p ^3P1-1 s^2 ^1S0 intercombination lines（y）, the 1 s2p ^3S~1-1 s^2 ^1S0 forbidden line（z） in helium-like Ti Ⅹ Ⅺ and the 1 s2s2p ^2P3/2-1 s622s ^2S1/2 line（q） in lithium-like Ti Ⅹ Ⅹhave been recorded with a X-ray CCD camera. The experimental result shows that the wavelength resolution（λ/△ 2） is above 1000 and the elliptical crystal spectrometer is suitable for X-ray spectroscopy.     

Preparation and Characterization of TaN ALD Precursors

High purity organic-tantalum precursors for thin film ALD TaN were synthesized and characterized.Vapor pressure and thermal stability of these precursors were studied.From the vapor pressure analysis,it was found that TBTEMT has a higher vapor pressure than any other published liquid TaN precursor,including TBTDET,TAITMATA,and IPTDET.Thermal stability of the alkyl groups on the precursors was investigated using a 1H NMR technique.The results indicated that the tertbutylimino group is the most stable group on TBTDET and TBTEMT as compared to the dialkylamido groups.Thermal stability of TaN precursors decreased in the following order:TBTDET ＞ PDMAT ＞ TBTEMT.In conclusion,precursor vapor pressure and thermal stability were tuned by making slight variations in the ligand sphere around the metal center.     

Electronic Structure of Semiconductor Nanocrystals

This paper reviews our recent development of the use of the large-scale pseudopotential method to calculate the electronic structure of semiconductor nanocrystals, such as quantum dots and wires, which often contain tens of thousands of atoms. The calculated size-dependent exciton energies and absorption spectra of quantum dots and wires are in good agreement with experiments. We show that the electronic structure of a nanocrystal can be tuned not only by its size,but also by its shape. Finally,we show that defect properties in quantum dots can be significantly different from those in bulk semiconductors.     

A rigorous proof of MIMO channel capacity’s increase with antenna number

It is well known that adding more antennas at the transmitter or at the receiver may offer larger channel capacity in the multiple-input multiple-output（MIMO） communication systems. In this letter, a simple proof is presented for the fact that the channel capacity increases with an increase in the number of receiving antennas. The proof is based on the famous capacity formula of Foschini and Gans with matrix theory.