卷积神经网络诊断阿尔兹海默症的方法

针对阿尔兹海默症（AD）通常会导致海马体区域萎缩的现象,提出一种使用卷积神经网络（CNN）对脑部磁共振成像（MRI）的海马体区域进行AD识别的方法。测试数据来自ADNI数据库提供的188位患者和229位正常人的脑部MRI图像。首先,将所有脑图像进行颅骨剥离,并配准到标准模板;其次,使用线性回归进行脑部萎缩的年龄矫正;然后,经过预处理后,从每个对象的3D脑图像的海马体区域提取出多幅2.5D的图像;最后,使用CNN对这些图像进行训练和识别,将同一个对象的图像识别结果用于对该对象的联合诊断。通过多次十折交叉验证方式进行实验,实验结果表明所提方法的平均识别准确率达到88.02%。与堆叠自动编码器（SAE）方法进行比较,比较结果表明,所提方法在仅使用MRI进行诊断的情况下效果比SAE方法有较大提高。     

一种新的LOG算子对SAR图像的边缘检测能力

在对SAR图像的分析和识别中,边缘的检测信息十分重要。针对传统的LOG算子的缺陷,本文提出了一种基于自适应中值滤波的改进LOG算子对SAR图像的边缘检测方法,在平滑图像内部区域的同时保留了更多的细节,提高了LOG算子对SAR图像的弱边缘检测能力,是一种有效的方法。     

机载合成孔径雷达（SAR）距离高分辨的实现

本文应用一对共轭匹配的LiNbO_3基底SAW器件,在发射机中产生相干Chirp脉冲串信号,在接收机中频完成脉冲压缩,同时保存了Doppler相位信息。它较之光学、CCD及数字处理有一些不可替代的优点,并展现了机载SAR的某些特色。文中详细阐述了设计概念和LFM脉冲压缩系统,给出了实验结果和测绘飞行获得的一张雷达图像。它证明了方案和技术实施的有效性。     

Binder removal from ceramic stereolithography green bodies: A neutron imaging and thermal analysis study

Stereolithography of ceramics remains one of the most powerful additive manufacturing routes for the creation of intricate ceramic parts. Despite its utility as a forming tool, ceramic stereolithography requires a challenging debinding stage due to the requisite high polymeric loading. Earlier research has identified both the polymeric resin composition and debinding atmosphere to be crucial factors in improving debinding performance. Here, we use a combination of thermogravimetric analysis and neutron imaging to examine samples of different compositions printed using the same processing and exposure parameters. We quantify the influence of both polyethylene glycol addition and the use of different debinding atmospheres (argon and vacuum) on the debinding behavior of ceramic pellets. Specifically, we demonstrate a method for examining the concentration gradients that develop during thermal debinding with the aid of neutron tomography. We find that at a constant heating rate of 1°C/min up to 500°C, vacuum atmosphere appears to result in a greater number of cracks as compared to the use of argon. The vacuum atmosphere led to the development of lower concentration gradients in the samples on average. The greatest improvement resulted with the addition of polyethylene glycol to the samples. This addition led to significantly less cracking and much lower concentration gradients in samples during debinding. These results prompt us to conclude that while keeping printing and exposure parameters constant, composition modification has a more significant effect on the debinding improvement than heating atmosphere.     

GaAs阴极激活铯源质量对成像器件性能的影响

为了解决三代微光器件研究中出现的阴极灵敏度低、光电发射不均匀、稳定性差等问题,重点开展了铯源质量对阴极性能产生影响的试验研究。结果表明,激活铯源纯度、装配结构、铯源与阴极距离对阴极光电发射灵敏度、均匀性影响最大。铯源出口孔径多少大小及射出方向是阴极产生暗班的主要原因。在铯源材料成分结构与阴极距离确定之后,阴极灵敏度的高低主要决定于对铯源的除气工艺,通过对铯源除气工艺进行优化,制备出了阴极灵敏度大于1700μA/lm三代微光器件。     

双波段成象及融合技术研究

概括地分析了双波段成象技术的发展概况，分析了成象技术的基本原理，提出了红外双波段基本方案和涉及的关键技术，分析和计算了系统的基本性能。最后，讨论了双波段及融合技术，认为在双波段实现融合应以自主式结构为主。     

Development of an energy-binned photon-counting detector for X-ray and gamma-ray imaging

The aim of this study is to develop an energy-binned photon-counting (EBPC) detector that enables us to provide energy information of x-rays with a reasonable count statistics. We used Al-pixel/CdTe/Pt semiconductor detectors, which had an active area of 8 mm×144 mm and consisted of 18 modules aligned linearly. The size of a CdTe detector module was 8 mm×8 mm and the thickness of the CdTe crystal was 1 mm. Each module consisted of 40×40 pixels and the pixel size was 200 μm×200 μm. We applied the bias voltage of −500 V to the Pt common electrode. The detector counted the number of x-ray photons with four different energy windows, and output four energy-binned images with pixel depths of 12, 12, 11 and 10 bits at a frame rate of 1200 Hz (300 Hz×4 energy bins). The basic performance of the detector was evaluated in several experiments. The results showed that the detector realized the photon counting rate of 0.4×10⁶ counts/sec/pixel (10⁷ counts/sec/mm²), energy resolution 4.4% FWHM at 122 keV. The integral uniformity of the detector was about 1% and the differential uniformity was about 1%. In addition, the image quality was examined with a resolution chart and step-wedge phantoms made of aluminum and polymethyl methacrylate. And we compared the quality of an acquired image with that acquired with an energy integration detector. The results of these experiments showed that the developed detector had desirable intrinsic characteristics for x-ray photon counting imaging.     

A nonintrusive flow measurement technique to validate the simulated laminar fluid flow in a packed container with vented walls

A nonintrusive flow measurement technique (particle image velocimetry) was used to determine the airflow field in a package with a container-to-product diameter ratio of less than 10. The complexity and uneven distribution of the measured flow field supported the requirement of a geometrical and mathematical model capable of describing the geometry and physics of flow within the package. Using novel computational fluid dynamics (CFD) codes, an accurate model of the packed structure was developed and the 3D Navier–Stokes equations were solved. A good agreement was obtained between experimental and predicted velocities. The detailed insight on the airflow pattern provided by the CFD analysis makes this approach an ideal tool to analyze the effect of different vent designs in the airflow field distribution in complex packaging systems.     

基于fMRI功能网络和贝叶斯矩阵分解的脑电源成像方法

脑电(EEG)是一种重要的脑功能成像技术,根据头皮记录的EEG信号重构皮层脑活动称为EEG源成像。然而脑源活动位置和尺寸的准确重构依然是一个挑战。为充分利用EEG和功能磁共振(fMRI)信号在时空分辨率上的互补信息,该文提出一个新的源成像方法——基于fMRI脑网络和时空约束的EEG源重构算法(FN-STCSI)。该方法在参数贝叶斯框架下,基于矩阵分解思想将源信号分解为若干时间基函数的线性组合。此外,为融合fMRI的高空间分辨率信息,FN-STCSI利用独立成分分析提取fMRI信号的功能网络,构建EEG源成像的空间协方差基,通过变分贝叶斯推断技术确定每个空间协方差基的相对贡献,实现EEG-fMRI融合。通过蒙特卡罗数值仿真和实验数据分析比较了FN-STCSI与现有算法在不同信噪比和不同先验条件下的性能,结果表明FN-STCSI能有效融合EEG-fMRI在时空上的互补信息,提高EEG弥散源成像的性能。     

基于轻量级U-Net深度学习的人体安检隐匿违禁物的实时检测

在高端智能安检系统研发中,如何使受检者在无接触正常行进过程中,对其实施人体是否携带隐匿违禁物的快速高效检测是具有挑战性的关键性技术。被动毫米波成像以其安全无害、穿透性强等突出优势而成为安检成像的热门选项。该文利用被动毫米波成像和可见光成像的优势互补,通过轻量级U-Net的深度学习,研究提出人体安检隐匿违禁物的高性能实时检测算法。首先构建和训练轻量级U-Net分割网络,进行被动毫米波图像(PMMWI)和可见光图像(VI)中人体轮廓的快速分割,实现人体与背景的有效分离,以获取疑似隐匿违禁物的轮廓信息。进而,以轻量级U-Net为工具,通过基于相似性测度的无监督学习方法进行被动毫米波人体轮廓图像与可见光人体轮廓图像的配准,以滤除虚警目标,并在可见光图像中进行疑似目标定位,得到单帧图像的检测结果。最后,通过序列多帧图像之检测结果的综合集成与推断,给出最终检测结果。通过在专门构建的数据集上的实验结果表明,该文所提方法的F1指标达到92.3%,展现出良好的性能优势。