Aiming at the problems of image contrast reduction and serious color cast in turbid water,we constructed a dataset of underwater image for experimental turbid water,and proposed an image enhancement method based on improved Shallow-UWnet network.Firstly,we employed the algorithm of gray scale for global color correction to original images.And then we utilized the improved Shallow-UWnet network,which learned the mapping relationship between the distorted and the normal images,to achieve underwater image enhancement.Finally,we improved the contrast of images to obtain final results,by employing contrast limited adaptive histogram equalization(CLAHE).The experimental results show that our method is superior to other 5 ones not only in subjective and objective evaluation indexes but also in key points matching.And it is effectively in correcting the color cast in different turbid water and improving the contrast and clarity.This method can be applied to underwater in-situ environment with turbidity,and is an available solution for improving underwater visualization.It has wide prospect in underwater detection,underwater salvation,underwater exploration and so on.
To address the problem of unsatisfactory classification results due to the limited number of labeled samples and insufficient extraction of diverse features in hyperspectral image classification tasks,this paper proposes a hyperspectral image classification method based on three-dimensional dilated convolution and graph convolution.Firstly,we introduce different scales of dilated convolution(DC)to build a three-dimensional dilated convolution network model to extract multi-scale deep spatial-spectral features.Secondly,we build a graph convolution neural network model by aggregating the neighborhood feature information of graph nodes to obtain the contextual features with spatial structure.Finally,to improve the representation capability of diverse features,we fuse deep spatial-spectral features with spatial contextual features and use Softmax to achieve classification.The proposed method can make full use of the diverse features of hyperspectral images and has a strong feature learning capability,which can effectively improve the classification accuracy.The proposed method is experimentally compared with seven related methods on the hyperspectral datasets of Indian Pines and Pavia University,and the results show that the proposed method could obtain optimal results with an overall classification accuracy of 99.33％and 99.41％.
We propose a second-order Raman fiber amplifier gain and noise co-prediction model combining convolutional neural network and long short-term memory network to study the influence of different LSTM layers on the performance of the prediction model, and the optimal parameter configuration model was obtained by optimizing with the Seahorse Algorithm. The model can accurately reflect the mapping relationship between pumping parameters, fiber length, target gain, and noise distribution, and effectively improves the design efficiency and performance of Raman fiber amplifiers. The experimental results show that the root mean square error of the finally established SHO-CNN-LSTM model in terms of gain and noise prediction is only 0.0431 and 0.0224dB, the error between the predicted value and the target value does not exceed 0.26dB, and the average design time does not exceed 0.0002s. This design scheme provides the best design methods and ideas for the flexible and fast design of future Raman fiber amplifiers.
In order to explore new fiber materials that can be used for discrete Raman amplification,first-and second-order Raman amplifiers are designed based on TiO2-doped fiber.A pump parameter configu-ration scheme with high power conversion efficiency and flat gain is given.The gain characteristics of first-order and second-order Raman amplifiers based on TiO2-doped fiber and second-order Raman am-plifiers based on GeO2-doped fiber are compared with the same total pump power.The simulation result shows that in the 60 nm bandwidth range of L-band,32 dBm pump light is injected forward into 6 km long TiO2-doped fiber to amplify 3 dBm signal light.Its power conversion efficiency can reach 41.57％,and the gain flatness is only 1.14 dB.Compared with the second-order Raman amplifier doped with GeO2 fiber,it has a more stable output gain.
Wavelength-tuning phase-shifting interferometry is a modern optical measurement method based on the propagation characteristics of light, which can deteunine sample topography by wavelength-level non-contact method with high-precision. The accuracy of this method can be up to nanometer and sub-nanometer precision. The measurement of transparent parallel plates is important and significant in optical measurement. Based on the design of window and sampling functions, the introduced 36 steps phase-shifting algorithm can effectively extract the target infounation. The influences of constant and first-order phase-shifting errors are analyzed. The least square separation method based on 19 steps iteration can deteunine the surface phase infounation accurately and separate surfaces including front, rear, thickness infounation, even the parasitic teen. The influence of different iteration steps for the PV and RMS values is considered, which provides the basis and reference for the design of the separation algorithm.
Point cloud registration is the key step of 3D non-contact precision measurement for complex mechanical parts based on machine vision. Due to extreme dependence on initial position, slow iterative convergence and many wrong corresponding point pairs, Iterative Closest Point (ICP) algorithm could not satisfy the requirements of point cloud registration efficiency and precision for measurement of large quantities of complex mechanical parts, so an improved ICP method of measurement point cloud registration for complex mechanical parts by combining ICP with Intrinsic Shape Signature-Fast Point Feature Histogram (ISS-FPFH) feature is proposed. In order to reduce the registration quantity of the point cloud and keep the original subtle features on surface of point cloud, the voxel filter based on the point close to center of gravity is proposed to preprocess for point cloud down sampling. It is difficult for traditional ICP algorithm to determine the appropriate initial position and will lead to the registration failure of multi-view measurement point cloud, the Sample Consensus Initial Registration (SAC-IA) algorithm based on ISS-FPFH is applied for coarse registration. In order to solve the problems of slow iterative convergence and many wrong corresponding point pairs of traditional ICP algorithm, the point-to-plane ICP algorithm by combining with normal vector angle constraint is proposed for fine registration. Through experimental comparison and analysis, the results show that the accuracy and efficiency of the method proposed in this paper can meet requirements of 3D non-contact precision measurement for large quantities of complex mechanical parts.
本文针对大功率垂直腔面发射激光器(vertical cavity surface emitting laser,VCSEL)阵列热阻大、出光不均匀的问题,研究p-GaAs层欧姆接触电阻值的作用机理,降低欧姆接触串联电阻的方法,以提高VCSEL阵列出射光功率的均匀性.基于3种常用欧姆接触金属Ti/Au、Ni/Au、Ti/Al/Ti/Au,研究各层金属厚度和金属组合对与p型欧姆接触电阻的作用规律;结合等离子体表面处理工艺,改变金属/p-GaAs界面态,研究界面态对欧姆接触电阻的影响规律.实验对比分析得到金属Ti/Au结构电极欧姆接触的比接触电阻率最低,为3.25×10-4Ω·cm2;基于金半接触势垒模型,通过表面等离子体处理,界面势垒可降低12.6％(0.269 2 eV降至0.235 3 eV),等离子体轰击功率可调控金半界面的势垒和态密度.
针对多无人机(unmanned aerial vehicle,UAV)作为空中基站辅助通信的吞吐量和公平性问题,提出了一种基于多智能体深度确定性策略梯度算法(multi-agent deep deterministic policy gradi-ent algorithms,MADDPG)的功率分配算法,该算法通过联合优化UAV基站的功率分配和用户接入以提高系统吞吐量和公平性.本文首先构建了 UAV基站为地面建立通信服务的三维场景,然后通过联合功率、用户关联和UAV位置约束,构建了吞吐量和公平性最大化的问题模型.考虑到该问题的复杂性,本文将所构建的优化问题建模为马尔科夫决策过程(Markov decision process,MDP),通过引入深度确定性策略梯度算法(deep deterministic policy gradient algorithm,DDPG)解决该问题.仿真结果表明,本文提出的基于MADDPG的UAV基站功率分配算法与其他算法相比,可以有效地提升系统的吞吐量和用户的公平性,提高通信的服务质量.
随着虚拟现实(virtual reality,VR)技术的飞速发展,各种应用层出不穷.然而目前大多应用仅限于定点的静态全景展示及游览,无法利用VR化不可能为可能的优越性.本文提出了一种融合VR技术与三维重建算法的类鸟飞行交互模拟系统.该系统搭建了结构稳定的硬件控制平台用以改变实际姿态,同时通过软件仿真了虚拟环境中鸟类的飞行,结合硬件驱动及软件模拟实现了深层次的飞行模拟.现有的虚拟环境大多通过人工建模、激光扫描仪或无人机(unmanned aeri-al vehicle,UAV)航拍构造,其中建模存在费时费力的缺点,激光扫描仪则无法适应大规模重建,而UAV成本高昂且需专业培训.本文引入了基于图像分簇的PMVS(patch based multi-view stereope-sis)算法,只需输入特定场景的图片即可利用计算机自动恢复出三维结构,不仅快速而且可对任意场景重构.使用户在足不出户的情况下即可在世界各地体验飞行,为旅游事业及VR产业提供了一种新的可能.
为了实现对环境温度的精确测量,提出了一种基于七芯光纤(seven-core fiber,SCF)的迈克尔逊干涉型温度传感器.该传感器由单模光纤(single mode fiber,SMF)和SCF熔锥构成,当光由SMF进入SCF时,由于光纤直径的急剧变小,在光纤细锥区域会激发出SCF中的高阶模,这些高阶模与纤芯基模经SCF端面反射后,再次回到细锥区域时发生干涉,并经由SMF输出.制作了不同长度SCF的传感器样品,并分别进行了温度传感实验研究.温度响应实验结果表明,在20-160℃温度范围内,长度为47 mm的传感器的温度灵敏度为0.127 4 nm/℃,拟合线性系数为0.9983,温度测量分辨率为0.0078℃,稳定性实验测得传感器的测量标准偏差为0.289 6℃.该温度传感器结构紧凑、易于制作、成本低廉、灵敏度高且测量范围大,在温度监测领域具有一定的应用潜力.
监控视频运动分割是视频浓缩、行为识别等视频智能处理的基础和前提,是计算机视觉领域的研究热点.现存运动分割方法大多步骤繁琐、计算量大,难以应用于计算能力有限的领域.为此,提出了一种联合二分思想和时空管道的监控视频运动分割方法.该方法首先使用嵌套椭圆时空管道模型计算初始累计时空流量来判断目标轨迹完整性(completeness of target trajectory,CTT);然后结合二分思想动态地调节椭圆采样线,自适应地捕捉采样区域的运动目标;最后提取采样线上的全部像素点形成自适应时空管道进行运动分割.实验结果表明,所提方法在保证精度的同时计算速度明显优于对比方法,且所提方法鲁棒性强,对运动情况多变的监控场景同样适用.
为提高激光诱导击穿光谱(laser induced breakdown spectroscopy,LIBS)技术定量分析的精度,开展了磁场约束下LIBS技术对土壤中重金属元素检测的研究,并采用多谱线强度归一化内标法进行数据处理.通过比较磁场强度分别为0T、0.3 T、0.8 T、1.25T时的光谱特性,得到光谱强度和信噪比(signal-to-noise ratio,SNR)随磁场强度增大而增大,在1.25 T磁场强度时,样品元素Cd和Cu的光谱强度和SNR要比无磁场作用分别增强了 34.77％、56.33％和40.83％、74.12％,构建了磁场强度为1.25 T时的Cd和Cu元素定量分析模型.结果显示,相对于传统内标法,采用多谱线强度归一化内标法的元素检测限分别从52.78 mg/kg和49.18mg/kg降低到23.87 mg/kg和18.06 mg/kg;相关系数分别从0.961 3和0.942 7提高到0.996 9和0.999 3.本实验研究改善了 LIBS的光谱特性,提高了定量分析的精度,采用多谱线强度归一化内标法降低了重金属的检出限和测量误差.
通过引入基于卷积神经网络(convolutional neural network,CNN)的分类算法,高光谱图像(hyperspectral image,HSI)分类任务的精度取得显著的提升,但目前主流CNN算法往往较为复杂且参数量大,从而导致网络难以训练以及容易产生过拟合问题.为在保证网络分类性能的前提下实现轻量化,本文提出一个轻量级架构的基于光谱-空间注意力交互机制的CNN网络用于HSI分类.为实现HSI的光谱-空间特征提取,构建了一个轻量化的双路径骨干网络用于两种特征的提取和融合.其次,为提高特征的表征能力,设计了两个注意力模块分别用于光谱和空间特征的权重再调整.同时,为加强双路径特征之间的关联以实现特征的更好融合,注意力交互机制被引入到网络中以进一步提升网络性能.在3个真实HSI数据集上的分类结果表明,本文所提网络可达到99.5％的分类准确度,并相比于其他网络至少减少50％的参数量.
超声成像因非侵入式、成本低且实时性好而被广泛应用.超声系统需要大量的采集通道数据和较高的采样率来提高图像重建质量,导致成像耗时,系统复杂.压缩感知(compressed sens-ing,CS)算法能够在欠采样的条件下用较少的测量值重构出原始信号.因此,针对系统面临的采样率高,数据量大的问题,本文将CS理论中的DWT-IRLS算法应用在超声成像中,通过离散小波变换基(discrete wavelet transformation,DWT)对超声数据进行稀疏转换,对高低频系数进行采样测量,并使用迭代重加权最小二乘法(iterative reweighted least squares,IRLS)进行测量系数重构,最后对变换域系数进行DWT逆转换得到重建图像.通过实验分析,以50％原始数据重建图像效果逐渐趋于稳定,在均方误差和峰值信噪比方面进行对比分析,DWT-IRLS算法相比较于DWT-OMP、DWT-CoSamp和DCT-IRLS等重构算法,成像质量更高,细节特征更为明显.
针对存在明显光照变化或遮挡物等室外复杂场景下,现有基于深度学习的视觉即时定位与地图构建(visual simultaneous localization and mapping,视觉SLAM)回环检测方法没有很好地利用图像的语义信息、场景细节且实时性差等问题,本文提出了一种YOLO-NKLT视觉SLAM回环检测方法.采用改进损失函数的YOLOv5网络模型获取具有语义信息的图像特征,构建训练集,对网络重训练,使提取的特征更加适用于复杂场景下的回环检测.为了进一步提高闭环检测的实时性,提出了一种基于非支配排序的KLT降维方法.通过在New College数据集和光照等变化更复杂的Nordland数据集上进行实验,结果表明:室外复杂场景下,相较于其他传统和基于深度学习的方法,所提方法具有更高的鲁棒性,可以取得更佳的准确率和实时性表现.
图像重采样检测是图像取证领域的重要任务,其目的是检测图像是否经过重采样操作.现有的基于深度学习的重采样检测方法大多只针对特定的重采样因子进行研究,而较少考虑重采样因子完全随机的情况.本文根据重采样操作中所涉及的插值技术原理设计了一组高效互补的图像预处理结构以避免图像内容的干扰,并通过可变形卷积层和高效通道注意力机制(efficient channel attention,ECA)分别提取和筛选重采样特征,从而有效提高了卷积神经网络整合提取不同重采样因子的重采样特征的能力.实验结果表明,无论对于未压缩的重采样图像还是JPEG压缩后处理的重采样图像,本文方法都可以有效检测,且预测准确率相比现有方法均有较大提升.
像素探测器一直是高分辨率、高速率粒子跟踪的工作平台.本文以多晶硅/氧化硅(poly-Si/SiOx)钝化接触异质结结构设计了硅像素探测器,为了实现探测器的超快响应,采用Silvaco TCAD对异质结硅像素探测器进行器件仿真,一方面研究了不同衬底厚度对载流子输运和收集的影响,另一方面研究了器件结构设计对异质结像素探测器击穿电压的影响.仿真结果表明:在相同的偏置电压下,较薄的硅衬底可以获得更强的漂移电场,进而提高器件对信号电荷的输运与收集速率,有利于提高探测器的时间响应.较小的保护环-有源区间距有利于提高器件的击穿电压,而在保护环-有源区间距较大的情况下,在有源区边缘设计金属场板结构,也能够有效提高器件的击穿电压,使探测器可以工作在较高电压下,从而提升探测器的响应速率.
本文提出了一种基于光谱积分宽度法来测量发光二极管(light emitting diode,LED)结温的新方法,并进行了理论分析和实验研究.本方法主要分为光谱数据采集、定标函数的测定和结温测量三个过程.首先,为了测量成本的降低和精度的提高而采用在正常工作电流下采集LED光谱数据,并通过采用不同温度下的光谱积分宽度与选定的某一基准状态下的值逐差可得到线性度达0.99以上的定标函数,并通过此定标函数可实时测定任意状态下的结温.其次,为了比较本方法测量结温的精确性,分别对单色和白光LED采用本方法和业界主流的正向电压法,通过自行设计的基于积分宽度法结温测量系统和美国Mentor Graphics公司的T3Ster型仪器的测量结果进行比较,两种方法测出的结温最大偏差为2.1℃,在可接受的误差范围内.实验结果表明积分宽度法测结温具有高效便捷且低成本的的特点,具有一定的应用前景.
随着调谐激光吸收光谱(tunable laser absorption spectroscopy,TLAS)技术在气体检测中的应用越来越广泛,二次谐波信号的质量与检测参数紧密相关,因而分析检测参数的优化方法很有意义.本文根据谱线预处理中的滤波参数、系统采样时间、锁相放大器的时间常数对信号的影响以及参数间的联系,总结检测参数的选取规律.根据滤波原理和不同浓度下信号的均方根误差(root mean squared error,RMSE)值选择合适的滤波阶数和窗宽.选择信噪比(signal-to-noise rati-o,SNR)、RMSE、信号与噪声频域幅度之比(ratio of amplitude in frequency domain of signal to noise,fSNR)3种评价指标的曲线变化趋势进行时频分析,得到最佳采样周期数为30,结合实验系统具体参数可计算最佳采样时间.通过信号主频带与截止频率的关系和滤波效果选择合适的时间常数.综合分析3个检测参数并总结选取方法,可提高二次谐波信号的质量.本文提出的参数选取方法对提高二次谐波在实际应用中的准确度有重要意义.
折射率(refractive index,RI)是眼镜镜片的重要参数,影响眼镜镜片的厚度和舒适程度.RI的精确测量,能够为屈光不正患者选择合适的镜片提供指导.本文提出了一种利用频域光学相干层析(optical coherence tomography,OCT)系统测量成型眼镜镜片RI的方法.利用放置镜片前后两次干涉的图样和简单的数据分析计算即可得到待测成型镜片的RI.介绍了使用这种方法的工作原理以及装置,首先对一个K9玻璃材质透镜RI进行测量,结果为1.507(1 310 nm),相对误差不超过2％,以此验证了该方法的合理性、可行性.随后利用此方法对一只成型眼镜镜片进行了测量,得到其RI为1.554(1310 nm),测量结果的扩展不确定度为0.008 6(k=2).该方法属于非接触测量,不会对成型镜片造成损坏,操作简便,测量相对误差较低,可基本满足对于成型眼镜镜片RI的测定精度要求.
