Spatial-temporal changes of wetlands and its driving factors in Guangdong-Hong Kong-Macao Greater Bay Area from 2010 to 2020EI北大核心CSCD

基于高分遥感影像提取2010年、2015年和2020年湿地数据,使用探索性空间数据分析和地理探测器,定量解析了粤港澳大湾区2010—2020年湿地时空变化及其驱动因素。结果表明:2010—2020年粤港澳大湾区湿地面积不断减少,其中,人工湿地减少更明显(305.33 km^(2)),自然湿地减少90.17 km^(2),湿地向建设用地、草地转移是湿地损失的主要形式;2010—2015年湿地损失的主要区域为大湾区中部核心区和沿海城镇带,2015—2020年湿地损失区域在新区、新建产业基地等局部区域聚集;人工湿地变化主要受GDP、渔业产量、第二产业产值等社会经济发展因素驱动,常住人口的影响力逐渐增强;气温、第一产业产值和道路密度对自然湿地变化起重要作用,其余因子影响较弱。     

主动迁移学习的海上任意方向船只目标检测

在基于深度学习的遥感图像目标检测任务中,船只目标通常呈现出任意方向排列的特性,而常见的水平框目标检测算法一般不能满足此类场景的应用需求。因此本文在单阶段Anchor-Free目标检测器CenterNet的基础上加入旋转角度预测分支,使其能输出旋转边界框,以用于海上船只目标的检测。同时针对海上船只遥感数据集仅有水平边界框标注,无法直接适用于旋转框目标检测,且人工手动标注旋转框标签成本较高的问题,提出一种主动迁移学习的旋转框标签生成方法。首先,提出一种水平框-旋转框约束筛选算法,通过水平真值边界框来对旋转预测框进行监督约束,筛选出检测精度较高的图像加入训练集,然后通过迭代这一过程筛选出更多的图像,最后通过标签类别匹配,完成对数据集的旋转框自动化标注工作。本文最终对海上船只遥感图像数据集BDCI中约65.59%的图片进行旋转框标注,并手动标注部分未标注的图片作为测试集,将本文方法标注的图片作为训练集进行验证,评估指标AP50达到90.41%,高于其他旋转框检测器,从而表明本文方法的有效性。     

基于HardSoftmax的并行选择核注意力

注意力被广泛地运用在卷积神经网络中,并有效地提升了卷积神经网络的性能。同时,注意力是非常轻量的,且几乎不需要改变卷积神经网络原来的架构。提出了基于HardSoftmax的并行选择核注意力。针对Softmax包含指数运算,对于较大的正输入很容易发生计算溢出的问题,提出了计算更安全的HardSoftmax来替换Softmax。不同于选择核注意力将全局特征的提取和转换放在特征融合之后,并行选择核注意力将全局特征的提取和转换单独放在一个分支,与具有不同核大小的多个分支构成并行结构。同时,并行选择核注意力的全局特征转换使用分组卷积,进一步减少参数量和计算量。并行选择核注意力通过HardSoftmax注意来关注不同核大小的多个分支。一系列的图像分类实验表明,只是简单地用HardSoftmax替换Softmax,也能保持或提升原注意力的性能。HardSoftmax的运行速度在实验中也比Softmax更快速。并行选择核注意力能够以更少的参数量和计算量追平或超越选择核注意力。     

Design of the low power wake-up receiver for internet of things communication

A wake-up receiver with high energy efficiency and low power consumption is proposed for solving the power consuming problems of wireless nodes communication in the Internet of Things. The proposed wake-up receiver based on the wake-up mechanism can effectively schedule the network nodes communication, and use the simple envelope detection structure to achieve frequency down-conversion, which can flexibly manage energy and reduce power consumption. Based on UMC 65nm CMOS process technology, the wake-up receiver is designed and simulated. The results show that it can achieve S₁₁ of -21dBm and a sensitivity of -75dBm at a data rate of 1Mb/s, when operating at the central frequency of 780MHz and input signal adopting an on-off keying (OOK) modulation, and the power consumption is 82μW at 1.2V voltage supply.     

Interface trap effect on the transient characteristics of a photovoltaic InSb infrared detector

The transient characteristics can be significantly affected by the interface trap. In order to reveal the intrinsic mechanism and provide some references for the design of detectors, the interface trap effect on the transient characteristics of a back-illuminated photovoltaic InSb infrared detector is studied with two-dimensional simulations. The relationships between interface traps and some key physical parameters, such as the hole concentration, the recombination rate and the electric field, are analyzed to reveal the intrinsic mechanisms of the influence of the interface traps on the transient characteristics. Studies show that with the increase of the interface trap density, the recombination rate and the electric field near the interface increase, the hole concentration decreases, and the “black holes” in current density distribution appear and move forward to the pn junction. And the transient photoresponse decreases with the increase of the interface trap density.     

面向车辆识别的样本自反馈式级联检测方法

以解决机器学习的样本提取问题、提高车辆识别的精度为目的,以Adaboost算法、级联检测方法、样本自反馈方法为基础,提出了一种面向车辆识别的样本自反馈式级联检测方法。方法兼具Adaboost级联检测器的架构优势和样本自反馈方法的样本扩充能力。通过前车图像的识别实例阐述了样本自反馈式级联检测方法的识别过程,进而使用常规Adaboost检测器和样本自反馈式级联检测方法对图像数据集进行了识别,将生产者精度、用户精度、总体识别精度、k值作为评价指标对识别效果进行了比较,证明了样本自反馈式级联检测方法具有较高的识别精度。分析了样本自反馈式级联检测方法个别评价指标较低的原因,提出了相应问题的解决思路。     

基于四象限探测器的红外导引头目标捕获和跟踪系统

设计的红外导引头主要为激光半主动制导半实物仿真系统的目标捕获装置,它的性能直接影响整个半实物仿真系统的定位精度。首先在详细分析了导引头的性能指标基础上设计一套红外导引头的光学系统,并用ZEMAX验证了光学系统设计的可行性;同时设计了四象限探测器的处理及采集电路,实现了目标方位信息的捕获;最后结合三轴运动转台的控制实现了目标的跟踪。本系统经实验验证能够在实验室环境下完成对激光目标模拟器的捕获和跟踪。     

Near-zero-shrinkage Al2O3 ceramic foams with coral-like and hollow-sphere structures via selective laser sintering and reaction bonding

The large shrinkage that ceramics undergo during sintering is a severe challenge for high-performance porous ceramics. In this study, we report a powder-based selective laser sintering (SLS) approach to prepare Al₂O₃ ceramic foams with near-zero shrinkage, high porosity, and outstanding strength. The ceramic foams consist of specific coral-like and hollow-sphere structures derived from the raw Al₂O₃/Al composite powders via reaction bonding (RB). A near-zero shrinkage of 0.91 ± 0.15 % and a high porosity of 73.7 ± 0.2 % can be achieved based on the Kirkendall effect during the oxidation of Al particles. Meanwhile, the reinforced sintering necks and robust bond-bridge connections between hollow-sphere and coral-like structures result in a remarkable bending strength of 7.37 ± 0.37 MPa. This measured strength is more than six times higher than other fabricated samples from spherical Al₂O₃ powders, and the comprehensive performance of ceramic foams prepared by this novel SLS/RB strategy is exceptionally remarkable versus that via conventional forming methods.     

航空维修用手持式智能超声探伤仪的研制

针对海军飞机原位探伤需要,应用超声检测和计算机技术,采用嵌入式系统,结合C语言编程,研制了手持式智能超声探伤仪,并对某型直升机尾桨叶大梁裂纹试件进行了探伤试验.结果表明仪器达到了设计要求,可用于海军各型飞机的超声探伤.     

3D FE simulation for temperature evolution in the selective laser sintering process

A thermal model of selective laser sintering (SLS) has been developed. The model allows for the non-linear behavior of thermal conductivity and of specific heat due to temperature changes and phase transformations. The temperature evolution and the formation of the sintered part are simulated by a 3D finite element analysis based on continuous media theory. It is shown that the effect of sintering has a strong influence on thermal evolution through changing the thermal properties of the material. The results of the model were experimentally tested and confirmed by temperature measurements.