A Systematic Review of Computer Vision Techniques for Quality Control in End-of-Line Visual Inspection of Antenna Parts

The rapid evolution of wireless communication technologies has underscored the critical role of antennas in ensuring seamless connectivity.Antenna defects,ranging from manufacturing imperfections to environmental wear,pose significant challenges to the reliability and performance of communication systems.This review paper navigates the landscape of antenna defect detection,emphasizing the need for a nuanced understanding of various defect types and the associated challenges in visual detection.This review paper serves as a valuable resource for researchers,engineers,and practitioners engaged in the design and maintenance of communication systems.The insights presented here pave the way for enhanced reliability in antenna systems through targeted defect detection measures.In this study,a comprehensive literature analysis on computer vision algorithms that are employed in end-of-line visual inspection of antenna parts is presented.The PRISMA principles will be followed throughout the review,and its goals are to provide a summary of recent research,identify relevant computer vision techniques,and evaluate how effective these techniques are in discovering defects during inspections.It contains articles from scholarly journals as well as papers presented at conferences up until June 2023.This research utilized search phrases that were relevant,and papers were chosen based on whether or not they met certain inclusion and exclusion criteria.In this study,several different computer vision approaches,such as feature extraction and defect classification,are broken down and analyzed.Additionally,their applicability and performance are discussed.The review highlights the significance of utilizing a wide variety of datasets and measurement criteria.The findings of this study add to the existing body of knowledge and point researchers in the direction of promising new areas of investigation,such as real-time inspection systems and multispectral imaging.This review,on its whole,offers a complete study of computer vision approaches for quality control in antenna parts.It does so by providing helpful insights and drawing attention to areas that require additional exploration.

Origin and accumulation of high-maturity oil and gas in deep parts of the Baxian Depression, Bohai Bay Basin, China

Great quantities of light oil and gas are produced from deep buried hill reservoirs at depths of 5,641 m to 6,027 m and 190 ℃ to 201 ℃ in the Niudong-1 Well, representing the deepest and hottest commercial hydrocarbons discovered in the Bohai Bay Basin in eastern China. This discovery suggests favorable exploration prospects for the deep parts of the basin. However, the discovery raises questions regarding the genesis and accumulation of hydrocarbons in deep reservoirs. Based on the geochemical features of the hydrocarbons and characteristics of the source rocks as well as thermal simulation experiments of hydrocarbon generation, we conclude that the oil and gas were generated from the highly mature Sha-4 Member （Es4） source rocks instead of thermal cracking of crude oils in earlier accumulations. The source kitchen with abnormal pressures and karsted carbonate reservoirs control the formation of high-maturity hydrocarbon accumulations in the buried hills （i.e., Niudong-1） in conjunction with several structural-lithologic traps in the ES4 reservoirs since the deposition of the upper Minghuazhen Formation. This means the oil and gas exploration potential in the deep parts of the Baxian Depression is probably high.

Study on the crust-mantle structure in the central and southern parts of Shanxi

Another comparative interpretation was conducted with respect to the data from 5 DSS profiles in the central and southern parts of Shanxi, leading to the conclusion that in Linxian, Linfen and Xingtai earthquake regions, through which the five profiles pass, there exist anomalous crust mantle structure and abyssal crustal faults extending to Moho, all being regarded as the deep indications for earthquake occurrence.

Influences of material parameters on deep drawing of thin-walled hemispheric surface part

During deep drawing process,the material parameters of blank have a significant effect on the quality of the drawn part and the determination of process parameters. Here,a 3D finite element model is developed for the deep drawing process of a thin-walled hemispheric surface part. Then the influences of material parameters including hardening exponent n,yield stress σs and elastic modulus E on the process are investigated by simulation. The results show that the effects of n and σs on punch force,thickness variation and equivalent strain are more notable. The maximum equivalent plastic strain occurs outside the die corner. However,when the value of n is 0.03 or σs is smaller than 120 MPa,higher equivalent plastic strain occurs at ball top.

NUMERICAL SIMULATION AND EXPERIMENTAL RESEARCH ON HYDRAULIC COUNTER-PRESSURE DEEP DRAWING OF CONICAL PART

Hydraulic counter pressure deep drawing of truncated conical part is numerically simulated with MARK and the nature of increasing the forming limit in this process is searched.The effects of blank holding force and chamber pressure on forming results are investigated by experiments and,as a result,truncated conical parts with large drawing ratio are successfully formed in single step with this drawing method.

A Survey on Deep Learning-Based 2D Human Pose Estimation Models

In this article,a comprehensive survey of deep learning-based(DLbased)human pose estimation(HPE)that can help researchers in the domain of computer vision is presented.HPE is among the fastest-growing research domains of computer vision and is used in solving several problems for human endeavours.After the detailed introduction,three different human body modes followed by the main stages of HPE and two pipelines of twodimensional(2D)HPE are presented.The details of the four components of HPE are also presented.The keypoints output format of two popular 2D HPE datasets and the most cited DL-based HPE articles from the year of breakthrough are both shown in tabular form.This study intends to highlight the limitations of published reviews and surveys respecting presenting a systematic review of the current DL-based solution to the 2D HPE model.Furthermore,a detailed and meaningful survey that will guide new and existing researchers on DL-based 2D HPE models is achieved.Finally,some future research directions in the field of HPE,such as limited data on disabled persons and multi-training DL-based models,are revealed to encourage researchers and promote the growth of HPE research.

基于深度学习的设施甜樱桃水肥一体化智能控制系统

水肥一体化技术在现代农业种植中的应用越来越广泛,但目前甜樱桃水肥一体化技术存在缺乏智能控制系统、实时施肥策略指导等问题,为此设计了一种甜樱桃水肥一体化智能控制系统。首先通过词性标注完成了对甜樱桃水肥领域知识实体的信息提取,结合领域知识定义相关概念的类和属性,构建甜樱桃水肥本体模型。引入了一种基于深度学习的甜樱桃水肥领域文本相似度计算模型,在基于深度学习的相似度计算模型的基础上,提出了基于甜樱桃水肥领域综合属性的相似度计算、基于甜樱桃本体语义的相似度计算和甜樱桃水肥策略实例的相似度计算。综合甜樱桃水肥策略实例的地域、果树品种、目标产量以及果树生长时期等属性因素建立了甜樱桃水肥策略库,以Drools规则引擎结合基于深度学习的相似度计算作为推理机制,最终搭建出甜樱桃水肥一体化智能控制系统。

基于注意时序网络的中文词性标注方法

针对传统的基于统计与规则的词性标注模型存在的人工特征依赖、字向量表征单一、特征提取不全面等问题,提出一种有效的基于注意时序网络的中文词性标注模型。对原始的TCN模型结构进行三点改进,并提出将注意时序网络与BiLSTM模型融合到词性标注方法中。上述模型首先通过XLNet模型获取字级别的上下文表示,利用注意时序网络的因果卷积结构获取更高层次的文本序列特征并通过注意力机制优化特征,最后通过BiLSTM进一步学习序列上下文特征,提高词性标注的准确度。实验表明,上述模型性能相较于其它模型有明显提升。

融合细粒度特征编码的点云分类分割网络

有效获取点云特征是分析和处理三维点云场景的关键。针对目前深度学习方法特征信息提取不充分,难以捕捉深层次语义信息的问题,提出了一种融合细粒度特征编码的网络来提高点云分类与分割任务的准确率。首先,特征提取模块包含2个子模块:一个是扩张图卷积模块,相比图卷积能够提取更丰富的几何信息;另一个是细粒度特征编码模块,能够获取局部区域的细节特征。其次,通过可学习参数将二者动态融合,有效地学习每个点的上下文信息。最后,将提取的所有特征相加,通过通道亲和注意力模块来强调不同通道,协助特征图来避免可能的冗余。在ModelNet40及ScanObjectNN数据集上进行点云分类实验,总体分类精度分别为93.3%和80.0%。在ShapeNet Part数据集上进行点云部件分割实验,平均交并比为85.6%。实验结果表明,与目前主流方法相比,该网络具有较优的性能。

基于协调注意力机制的轻量级YOLOv4零件检测

针对零件自动检测任务在复杂工况下,如零件堆叠粘连、有杂物干扰等,存在实时性差、硬件资源占用大等问题,提出一种基于轻量级YOLOv4网络的零件检测方法。采用MobileNeXt代替CSPDarkNet53作为主干特征提取网络(backbone),并在每个卷积模块中添加协调注意力机制,用于增强特征层的语义表达能力;提出一种Fused-Sandglass模块插入到浅层的backbone中,提高网络的推理速度;网络训练方面引入渐进式训练方法和focal loss损失函数,提升训练速度,并且有效缓解正负样本失衡的问题。实验结果表明,该方法在15种零件的检测任务中能够保持和YOLOv4网络相近的准确率,但参数量大小仅为其20%,推理速度达到了43.7 fps,能够满足实际生产的需求。

基于深度迁移学习的再制造零/部件可靠性预测方法

可靠性是反映再制造零/部件质量及其稳定性的重要指标。相比于新产品零/部件,影响再制造零/部件的可靠性因素多、机理关系复杂,且毛坯剩余价值高,难以开展大量的可靠性试验,使得可靠性数据样本量少,导致再制造零/部件可靠性预测精度低。为此,提出一种基于深度迁移学习的再制造零/部件可靠性预测方法。首先,分析制造阶段个性化工艺过程对再制造零/部件可靠性的影响,结合服役阶段运行数据对可靠性进行预测,通过引入新产品与同一产品不同型号零/部件服役阶段运行数据、制造阶段工艺质量数据进行样本扩充,利用主成分分析法获取影响可靠性的运行时间、状态信息及加工精度等关键特征数据并构建源域数据集;其次,采用卷积神经网络构建源域关键特征数据与可靠性之间的深度学习模型,以再制造零/部件数据为目标域数据集,通过自适应梯度算法将模型迁移应用到再制造零/部件可靠性预测中,以提高预测精度;最后,以某数控机床需要再制造的零/部件为例,对所提出预测方法的有效性进行验证。

面向工业目标检测的密集控制阀零件数据集

工业生产中的自动化智能化离不开自动目标检测,而高准确性的自动目标检测则依赖于与实际场景相适应的数据集。本文针对工业实际场景,发布了一个密集控制阀零件数据集,命名为PD4CV(Part Detection for Control Valve)2023。该数据集的图像全部来源于控制阀生产车间,图像采集完成后,首先对数据集图片进行预处理操作,接着对数据集图片中的零件目标进行标注,然后再对数据集图片进行训练集、验证集以及测试集的划分。PD4CV2023数据集共涵盖9类零件,包括510张工盘图像和15 015个零件样本,平均每张图像含有约29个零件样本。与现有的目标检测数据集相比,该数据集具有零件摆放密集、遮挡,零件尺寸差异大,部分零件外形相似,零件样本数量不均衡等特点。最后,在不同类型数据集上的预训练对比实验表明,一般场景数据集、特定工业场景数据集只适用于一般和特定任务,而代表实际控制阀生产工况下的PD4CV2023数据集,可用于控制阀零件目标检测,其具有其特殊性和不可替代性;一系列目标检测算法在该数据集上的综合对比则验证了PD4CV2023数据集在一般性目标检测、多尺度目标检测、小规模、不均衡数据下目标检测中的有效性。PD4CV2023数据集可用于面向工业的目标检测的相关研究。

U型框类零件双向拉深成形研究

针对U型框类零件成形质量不稳定、修整手工量大等问题,通过对零件结构的分析研究,借助CATIA对其进行了工艺补充面设计,并借助数值模拟对成形过程进行了仿真,研究了双向拉深成形工艺,设计了一种特有传力方式的双向拉深模具结构;选取凸凹模间隙、摩擦因数、压边力及拉深速度4个因素为优化变量,通过正交试验得出最优工艺参数组合,即压边力大小为30 kN、摩擦因数为0.1、凸凹模间隙为1.6 mm、拉深速度为6000 mm·s^(-1);最后针对拉深后回弹现象,对校正模进行了分析和设计并修整。生产试验表明,此模具结构可以很好地改善零件成形质量,提高效率并满足生产及设计要求。

基于目标搜寻和细节增强的水下单像素成像方法

针对当前水下单像素成像方法侧重于从整体角度重构目标图像,难以理想地恢复目标细节的问题,提出了一种基于目标搜寻和细节增强的水下单像素成像方法。目标搜寻旨在从图像中判断出目标部分和背景部分,从而增强目标信号,降低背景噪声;细节增强旨在学习采集信息的细粒度特征、增强重构图像的细节。首先用传统单像素成像方法快速重构目标图像;其次通过判断各行、各列最大像素点的差值来区别目标和背景环境;最后用基于分块模型的神经网络学习目标的细粒度特征,提高目标图像的细节部分。为了验证提出方法的可靠性,重构了空间环境和水下环境中的目标图像,实验结果表明,在两种实验环境下,该方法都可以较好地保存目标的细节信息,获得高质量的目标图像。

基于三维点云的植株联合任务分割框架

传统的植株器官分割方法依赖经验选择阈值参数,而当前的深度学习浅层框架可能会导致植株重要的几何特征丢失,并难以有效整合植株的局部和全局特征。因此,提出了一个基于三维点云的植株器官分割网络(local global feature fusion segmentation network,LGF-SegNet)模型,通过引入双权重注意力机制模块和位置编码,更适合在植株点云数据中表达几何特征。在提出的框架的解码层引入特征聚合模块,融合植株点云的局部和全局特征,使得该框架能够关注植株的整体特征轮廓同时保留细节植物纹理(如茎和叶)。实验结果表明,提出的架构在语义分割的交并比、精确率和F1分数的平均值分别达到85.76%、93.18%、91.08%,在实例分割的平均精确率、平均实例覆盖率以及平均实例加权覆盖率达到85.27%、78.46%、79.63%,优于当前流行的植株点云分割任务中使用的深度学习网络架构,并适用于植株语义分割和实例分割的双重任务。这为后续的植株生长预测等研究奠定基础。

一种基于CP-U-Net的鸡部位分割方法

鸡部位分割是鸡肉分体的主要任务,论文以鸡部位分割为目的,提出了一种基于U-Net的鸡部位图像分割算法。根据鸡部位检测的需求,选择了合适类型的工业相机、摄像头、光源和PC机构成鸡部位图像采集系统。采集3000张鸡部位图像,然后进行数据增强扩充,构建了鸡部位数据集;构建了一种基于CP(Chicken parts)-U-Net鸡部位图像分割模型,提取鸡部位特征,为了实现鸡部位图像的端到端语义分割,采取了以下步骤:首先通过池化计算,能够获得鸡部位的深层特征和浅层。接着,通过多次的反卷积处理,能够得到特征的融合。最后生成了鸡部位区域的二值图像。为了评估分割性能,采用平均交并比(MIoU)、均像素精度(MPA)和精度(PA)这三种评判标准,将CP-U-Net计算网络与三种经典算法进行了比较。实验表明:CP-U-Net在进行鸡部位的语义分割时PA、MPA、MIou分别为93.65%,89.12%,85.37%。比较对比的三种其他图像语义分割方法分别高出11.23%,8.74%,6.68%,且处理单幅鸡部位图像时间比SegNet缩短44 ms。

基于YOLOv5的机械零件识别及表面损伤判识研究与应用

零件识别及表面损伤判识是机器视觉在工业领域中的重要应用。文章提出一种基于深度学习的图像识别方法来进行机械零件缺陷判别。首先建立机械零件缺陷样本库,使用YOLOv5算法进行模型搭建,并经过训练、调参后获得最优模型;在求解模型过程中,文章针对机械零件数据集进行测试,结果表明模型在验证集上的精度达到92%以上,最后经过验证,模型可以准确识别缺陷,为实际机械零件识别提供了有效的决策支持。

基于机器视觉的零件分选技术研究进展

介绍机器视觉检测系统工作流程,详细介绍图像处理关键技术和方法;重点分析机器视觉技术在零件分选中的主流方法、研究进展,对未来发展方向进行展望。

基于有限元法的某深孔件模锻成形工艺研究

某深孔件是船用低速柴油机中的关键构件,由于其特殊的工作环境,该工件的制造技术有着极高的要求。精密模锻是一种可以改进锻件整体质量及提高产品性能的工艺,因此可以通过模锻工艺来加工该深孔件。本文采用有限元模拟软件对该深孔件在压力机上模锻成形过程进行了数值模拟仿真,分析了锻件的温度场及应力场的分布特征,从而指导该深孔件的实际生产。