基于改进SSD的工件表面缺陷检测

工件的表面缺陷不仅影响外观而且直接影响产品的质量、寿命和性能,因此对工件进行实时表面缺陷检测很有必要。针对当前SSD算法不利于小目标检测易导致误检的情况,提出了一种基于单阶段多层检测器的改进SSD自动检测方法。采用了以ResNet替换SSD中原始的VGGNet的方法,研究了小目标检测的问题;采用了对深层特征进行反卷积且将深层特征与浅层特征融合的方法,研究了语义信息不足易误检的问题。结果表明,该方法较原SSD模型在工件的表面缺陷检测上m AP值提高了约4.6%,从而认为本方法可用于工件表面缺陷的实时自动检测。

基于深度学习SSD算法的高密度电法智能解译方法技术研究

高密度电法在探测灰岩区地下溶洞病害体方面得到广泛应用,但高密度电法反演结果依赖于初始模型,存在多解性,地质解译容易受专业人员主观因素影响。为此,本文从具有唯一性的视电阻率数据出发,研究了基于深度学习的SSD(Single Shot Multi-box Detector)目标检测算法的视电阻率异常智能解译方法技术。针对岩溶地质病害,设计了不同填充类型、形状、规模、数量的溶洞电性异常模型,利用Res2dmod软件进行视电阻率正演计算,构建了包含1400个样本的高密度电法视电阻率智能解译学习样本库(样本和标签)。基于TensorFlow框架,建立了基于深度学习SSD算法的高密度电法视电阻率异常智能解译方法技术,使用学习样本库训练网络权值,训练结束后对高密电法温纳装置视电阻率异常进行智能解译,单个视电阻率剖面异常智能解译耗时不到1 s,各类目标(填充型溶洞、未填充型溶洞)平均准确率为90.68%。研究结果表明:基于SSD算法的高密度电法视电阻率异常智能解译技术可显著提高高密度电法视电阻率解译效率,避免专业人员主观因素影响。

改进SSD算法在工件表面缺陷检测中的应用

为能够更为高效地检测工件表面加工缺陷,提出了一种MobileNetV3-SSD目标检测算法,它能够有效实现工件表面缺陷检测。此算法通过改进SSD的骨架网络为MobileNetV3-Large,有效减少网络参数量和计算量,通过结合自底向上的特征金字塔网络的改进算法,使网络能够捕捉不同尺度和形状的目标信息,将语义分割辅助任务与多层特征融合相结合,进一步提升检测性能。使用PASCAL VOC数据集和Kolektor工业缺陷数据集验证该算法,在PASCAL VOC数据集上的mAP达到77.3%,较SSD算法提高0.4%,在Kolektor数据集上的FPS达到了105,较SSD算法提高112%。验证了MobileNetV3-SSD算法相较于传统SSD网络,具有更好的检测精度和检测速度。

基于对比学习的改进SSD目标检测算法

现有基于深度学习的目标检测算法在图像的目标检测过程中存在物体视角的多样性、目标本身形变、检测物体受遮挡、光照性以及小目标检测等问题。为了解决这些问题,本文将对比学习思想引入到SSD(Single Shot MutiBox Detectior)目标检测网络中,对原有的SSD算法进行改进。首先,通过采用图像截块的方式随机截取样本图片中的目标图片与背景图片,将目标图像块与背景图像块输入到对比学习网络中提取图片特征进行对比损失计算。随后,使用监督学习的方法对SSD网络进行训练,将对比损失传入到SSD网络中与SSD损失值加权求和反馈给SSD网络,进行网络参数的优化。由于在目标检测网络中加入了对比学习的思想,提高了背景和目标在特征空间中的区分度。因此所提出的算法能显著提高SSD网络对于目标检测的精度,并在可见光和热红外图像中均取得了令人满意的检测效果。在PASCAL VOC2012数据集实验中,AP50值提升了0.3%,在LLVIP数据集实验中,AP50值提升了0.2%。

基于轻量化SSD的交通标志检测算法

实时精确的交通标志检测是自动驾驶和智能交通的关键技术。针对现有智能检测算法检测复杂真实道路场景下的交通标志速度慢、无法较好地适用于嵌入式终端设备的问题,提出了一种基于轻量化SSD的交通标识检测算法。该算法采用MobileNetV3_large网络替代VGG16网络,可减少模型参数,提高检测实时性;利用添加SE模块的逆残差结构B-neck替换对应的标准卷积增强低层特征层的语义信息;设计改进RFB网络提升小交通标志的检测能力,重新设置预设先验框的尺寸,提升模型对特定数据集的检测能力。实验结果表明,改进SSD算法在中国交通标志检测数据集上的mAP值可达89.04%,比MobileNet-SSD算法提高了5.26%;帧率可达60 frames/s,比SSD算法提高了23 frames/s。所提算法具有较高的实时性和检测精度,对复杂交通环境具有更好的鲁棒性。

基于轻量级MobileNet-SSD模型的人流量检测

利用深度神经网络模型识别行人目标并检测具有十分高的价值。现实高密度行人检测场景中由于硬件基础以及网络性能消耗等问题的影响,常常需要选取处理速度高、硬件条件要求低的网络且同时兼顾视频监控的连续特性,因此选取了轻量级MobileNet-SSD网络来高效处理人头目标并引入帧间差分的方式,来有效对人头的椭圆特征目标进行跟踪;结合相关数理方法实现行人跨线计数的高性能人流量检测解决方案。在不同数据集上同现今一流的检测模型作比较,该方法皆表现出优良的检测性能。

基于RSSD的遥感图像目标检测算法

针对SSD算法检测遥感图像目标时存在容易漏检且检测精度低的问题,提出基于残差SSD网络的遥感图像目标检测算法。该算法在SSD网络结构的基础上,将基准网络模型VGG替换为残差网络模型ResNet-50,通过增加网络深度,充分提取遥感图像小目标数据集的底层特征,引入注意力模块,使感受野更关注目标特征,增强低层网络的信息表征能力,采用特征金字塔融合方法融合网络结构的高层语义特征和低层视觉特征,增强检测目标的定位能力。实验结果表明,该算法增强了复杂背景的干扰抑制性,提高了小目标的检测精度,比传统的SSD算法具有更强的检测性能。

基于深度SSD改进模型的传动设备状态在线监测研究

针对现有传动设备在线监测算法存在的检测精度低、效率差等问题,提出一种基于改进SSD网络模型的在线检测算法;先对故障集进行预处理,通过滤波调制、共振解调等环节滤除原始故障集的噪声干扰;以VGG-16为基础设计了SSD网络结构,同时增加了辅助卷积层和预测层;对SSD网络模型进行改进,引入了注意力机制模块和特征增强模块,改善模型各层的数据共享性能同时提高了模型的数据训练效率;基于通道拼合方式对故障数据进行多尺度特征融合,并优化SSD模型的各层金字塔结构,以更好的匹配先验框及选择最佳的损失函数;实验结果显示,提出算法的传动设备故障检测率达到98.8%,参与对比的3种传统算法故障检测率分别为94.2%、93.6%和93.7%,同时提出算法的检测效率也优于传统算法。

改进SSD的变电站电力设备识别方法

基于机器视觉的变电站设备识别是电力智能巡检的关键环节,在巡检图像上不同设备尺寸差异显著,极大影响了现有算法的识别性能。针对尺寸差异显著的识别对象,提出一种改进的SSD算法用于变电站设备识别任务。该方法主要改进SSD算法的两个重要部分:其一,保证多尺度特征提取,加强大尺度特征图的有效性。其二,优化默认框的生成策略,提出符合尺寸差异显著对象的多种默认框。此外,为增强网络的鲁棒性,采用迁移学习技术和数据增强技术训练电力设备识别网络模型。最后,构建变电站电力设备数据集,并进行实验测试;实验结果表明,论文方法相较于传统SSD算法,召回率提升了42.24%,解决了尺寸差异显著对象的识别问题。

基于超低延迟SSD的页交换机制关键技术

随着内存密集型应用的快速发展,应用对单机内存容量的需求日益增大.然而,受到颗粒密度的限制,内存容量的扩展度较低.页交换机制是进行内存扩展的经典技术,该机制通过将较少使用的内存页面暂存在存储设备,以达到扩展内存的目的.过去页交换机制由于慢速磁盘的读写速度限制,无法被广泛应用.近年来,得益于超低延迟固态硬盘(solid state drive,SSD)的快速发展,页交换机制可以利用其低延迟的读写特性,提升页交换效率.然而,在低I/O延迟的情况下,传统页交换机制的I/O栈存在巨大的软件开销.首先对使用超低延迟SSD的Linux页交换机制进行测试与分析,发现现有页交换机制的主要瓶颈在于发送请求时存在队头阻塞问题、I/O合并和调度开销,以及内核返回路径上的中断处理和直接内存回收开销.基于分析结果,提出基于超低延迟SSD的页交换机制Ultraswap.Ultraswap在Linux I/O栈的基础上增加对轮询请求的处理,并降低I/O合并与调度开销,实现轻量级的I/O栈.基于Ultraswap的I/O栈,对内核页交换机制的换入与换出路径进一步优化.通过优化对缺页、直接内存回收的处理,降低页交换机制关键路径上的时间开销.实验结果表明Ultraswap在应用测试场景下相比Linux页交换机制能够提升19%的平均性能;在可使用内存比例为20%的情况下,Ultraswap可达到33%的性能提升.

轻量化SSD的人脸检测

针对当前目标检测算法难以在资源有限的硬件平台上运行的问题,以Single Shot MultiBox Detector(SSD)为基础,在SSD中引入GhostNet网络来替换原本的VGG16主干网络,再引入轻量化注意力机制来补偿轻量化模型带来的检测精度损失,提高网络检测能力。在同一数据集上进行检测,改进后的GC-SSD模型大小由原SSD模型的92.11 M减小到10.90 M,mAP相比于Ghost-SSD模型提高了15.43%。实验结果表明,改进后的GC-SSD模型mAP达到92.06%,模型大小大量减小,更适用于移动端设备。

基于改进SSD算法的两相把手式开关识别

为协助操作人员准确判断设备开关状态,越来越多的配电站房引入智能巡检机器人进行设备状态监测。在智能巡检机器人参与采集图像数据的硬件基础上,提出了一种改进单激发多框探测器(SSD)图像识别算法,实现了对两相把手式开关的识别。改进SSD算法通过将基于卷积块的注意机制(CBAM)嵌入到SSD算法中,从通道和空间维度中选择有效的特征,抑制不相关的特征,提高模型的识别精度。实验结果表明,改进SSD算法对开关状态目标检测的平均精度(AP)在98.66%以上,平均精度均值(mAP)为99.10%。相较于传统的SSD算法,改进的SSD算法能更好地识别开关状态,提高变电站智能化水平。

基于改进型SSD算法的铁路货场异物侵限小目标检测研究

为解决铁路货场异物侵限场景中小目标检测难度大、准确率不高、检测效果不佳的问题,提出基于特征金字塔网络的改进型单阶多框检测(SSD)算法。通过研究现阶段铁路货场业务管理现状、异物侵限场景及对应的检测技术,对小目标检测现存问题进行归类总结;通过在SSD算法的检测网络部分增加不同特征层信息的金字塔网络结构,提高小目标检测效率。根据改进前后2种算法在铁路货场异物侵限场景的试验数据对比,得出改进型SSD算法推理阶段的检出精度更高,可有效提高小目标检测效率和准确率,为铁路货场智能化安全管控提供有力的技术支撑。

Novelty of Different Distance Approach for Multi-Criteria Decision-Making Challenges Using q-Rung Vague Sets

In this article,multiple attribute decision-making problems are solved using the vague normal set(VNS).It is possible to generalize the vague set(VS)and q-rung fuzzy set(FS)into the q-rung vague set(VS).A log q-rung normal vague weighted averaging(log q-rung NVWA),a log q-rung normal vague weighted geometric(log q-rung NVWG),a log generalized q-rung normal vague weighted averaging(log Gq-rung NVWA),and a log generalized q-rungnormal vagueweightedgeometric(logGq-rungNVWG)operator are discussed in this article.Adescription is provided of the scoring function,accuracy function and operational laws of the log q-rung VS.The algorithms underlying these functions are also described.A numerical example is provided to extend the Euclidean distance and the Humming distance.Additionally,idempotency,boundedness,commutativity,and monotonicity of the log q-rung VS are examined as they facilitate recognizing the optimal alternative more quickly and help clarify conceptualization.We chose five anemia patients with four types of symptoms including seizures,emotional shock or hysteria,brain cause,and high fever,who had either retrograde amnesia,anterograde amnesia,transient global amnesia,post-traumatic amnesia,or infantile amnesia.Natural numbers q are used to express the results of the models.To demonstrate the effectiveness and accuracy of the models we are investigating,we compare several existing models with those that have been developed.

基于改进SSD的自然场景小交通标志检测

为提高在复杂的自然交通场景下对小交通标志检测的准确度,改进了SSD模型。在SSD多个检测层加入并行多尺度特征融合,通过结合深、浅特征层的检测优势,改善了SSD模型在小目标检测方面的不足;在SSD模型的多个检测头分别加入注意力机制模块,增强对小交通标志的特征提取效果;加入focal loss损失函数减小背景对整体损失的贡献,防止背景过拟合。实验结果表明,在复杂自然场景下,改进的方法相比原始模型对小交通标志检测的mAP提升了4.9百分点。

Legendre-Jacobi’s Elliptic Integrals Shed Light on the Luminosity Distance in Cosmology

This article concerns the integral related to the transverse comoving distance and, in turn, to the luminosity distance both in the standard non-flat and flat cosmology. The purpose is to determine a straightforward mathematical formulation for the luminosity distance as function of the transverse comoving distance for all cosmology cases with a non-zero cosmological constant by adopting a different mindset. The applied method deals with incomplete elliptical integrals of the first kind associated with the polynomial roots admitted in the comoving distance integral according to the scientific literature. The outcome shows that the luminosity distance can be obtained by the combination of an analytical solution followed by a numerical integration in order to account for the redshift. This solution is solely compared to the current Gaussian quadrature method used as basic recognized algorithm in standard cosmology.

Effect of long reaction distance on gas composition from organic-rich shale pyrolysis under high-temperature steam environment

When high-temperature steam is used as a medium to pyrolyze organic-rich shale,water steam not only acts as heat transfer but also participates in the chemical reaction of organic matter pyrolysis,thus affecting the generation law and release characteristics of gas products.In this study,based on a long-distance reaction system of organic-rich shale pyrolysis via steam injection,the effects of steam temperature and reaction distance on gas product composition are analyzed in depth and compared with other pyrolysis processes.The advantages of organic-rich shale pyrolysis via steam injection are then evaluated.The volume concentration of hydrogen in the gas product obtained via the steam injection pyrolysis of organic-rich shale is the highest,which is more than 60%.The hydrogen content increases as the reaction distance is extended;however,the rate of increase changes gradually.Increasing the reaction distance from 800 to 4000 mm increases the hydrogen content from 34.91%to 69.68%and from 63.13%to 78.61%when the steam temperature is 500°C and 555°C,respectively.However,the higher the heat injection temperature,the smaller the reaction distance required to form a high concentration hydrogen pyrolysis environment(hydrogen concentration>60%).When the steam pyrolysis temperature is increased from 500°C to 555°C,the reaction distance required to form a high concentration of hydrogen is reduced from 3800 to 800 mm.Compared with the direct retorting process,the volume concentration of hydrogen obtained from high-temperature steam pyrolysis of organic-rich shale is 8.82 and 10.72 times that of the commonly used Fushun and Kivite furnaces,respectively.The pyrolysis of organic-rich shale via steam injection is a pyrolysis process in a hydrogen-rich environment.

Responses of soil stoichiometry and soil enzyme activities in the different distance around opencast coal mine of the Hulun Buir Grassland of China

The objectives of this study were to explore the changes in soil stoichiometry and enzyme activities at different distances from an opencast coal mine in the Hulun Buir Grassland of China. Four transects were established on north and east sides of the opencast coal mining area, and samples were collected at 50 m, 550 m, and 1550 m from the pit on each transect. Control samples were collected from a grassland station 8 km from the opencast coal mining area that was not disturbed by mining. Four replicate soil samples were collected at each point on the four transects. Soil physicochemical properties and enzyme activities were determined, and correlations between soil properties and stoichiometric ratios and enzyme activities were explored using redundancy analysis. The increase in distance from mining did not significantly affect soil properties, although soil urease activity was significantly lower than that of the control area. Soil properties 1550 m from the mine pit were similar to those at the grassland control. In addition, soil total nitrogen had the greatest effect on soil stoichiometry, and soil total potassium had the greatest effect on soil enzyme activities. Coal dust from opencast mining might be the main factor affecting soil stoichiometry and enzyme activities. The results of this study provide direction for the next step in studying the influence of mining areas on soil properties and processes.

基于SSD和切片双谱的煤机轴承诊断方法研究

针对煤机轴承振动信号受噪声、冲击载荷等因素影响,其故障特征难以提取的问题,提出了基于奇异谱分解(SSD)与切片双谱相融合的故障诊断方法。首先利用SSD将煤机轴承信号分解成若干个奇异谱分量,并计算每个分量的峭度值;然后根据峭度值筛选出的分量进行希尔伯特包络;最后对包络信号进行切片双谱分析得到准确的故障特征频率。现场故障案例表明,该方法几乎可以完全抑制噪声,有效地将强背景噪声下的煤机轴承故障特征频率提取出来,为煤机轴承故障诊断提供了一种新途径。

Robust free-space optical frequency transfer in time-varying link distances conditions

Future inter-satellite clock comparison on high orbit will require optical time and frequency transmission technology between moving objects.Here,we demonstrate robust optical frequency transmission under the condition of variable link distance.This variable link is accomplished by the relative motion of a single telescope fixed on the experimental platform to a corner-cube reflector(CCR)installed on a sliding guide.Two acousto–optic modulators with different frequencies are used to separate forward signal from backward signal.With active phase noise suppression,when the CCR moves back and forth at a constant velocity of 20 cm/s and an acceleration of 20 cm/s^(2),we achieve the best frequency stability of 1.9×10^(-16) at 1 s and 7.9×10^(-19) at 1000 s indoors.This work paves the way for future studying optical frequency transfer between ultra-high-orbit satellites.