基于Transformer改进的YOLOv5+DeepSORT的车辆跟踪算法

针对传统目标检测跟踪算法检测精度低、全局感知能力差、对遮挡和小目标物体的识别能力差等问题,提出了一种基于轻量化Transformer改进的YOLOv5和DeepSORT算法的车辆跟踪方法。首先,利用EfficientFormerV2模型改进YOLOv5算法模型,增强车辆的目标检测能力;然后,利用移位窗口(Swin)模型的优点改进DeepSORT多目标跟踪算法中的重识别(Re-Identification)模块,提高车辆的跟踪能力和精度;最后,通过数据集KITTI和VeRi开展对比试验和消融实验。结果表明,在复杂工况下,该方法的性能在车辆遮挡和小目标识别方面显著提高,平均准确度达到96.7%,目标跟踪准确度提高了9.547%,编号(ID)切换总次数减少了26.4%。

基于改进YOLOv5s的CNN-Swin Transformer森林野生动物图像目标检测算法

【目的】为提高野生动物在复杂森林环境中的检测精度,促进森林野生动物保护技术发展,提出一种基于YOLOv5s网络模型、针对陷阱相机所摄取森林野生动物图像的改进检测算法。【方法】以包含湖南壶瓶山国家级自然保护区几种典型森林野生动物在内的数据集为研究对象,首先,对真实标注框图像进行裁剪、归一化和缩放处理,随机将2~4张裁剪图像拼贴组成新的数据集元素,以丰富和增强数据集图像信息;其次,使用一种基于通道注意力思想的加权通道拼接方法,在通道拼接时引入权重改变通道数量,通过反向传播训练方法不断更新权重以增加重要特征信息的通道层数;接着,引入Swin Transformer模块与CNN网络相结合,为卷积神经网络特征提取加入自注意力机制,融合2种网络特征提取层的优势,提高特征提取的感受野;最后,选择更优的α-DIoU损失函数替代GIoU损失函数,针对边界框重叠面积和中心点距离造成的损失,引入新的几何因素惩罚项。【结果】在相同试验条件和数据集下,相比原YOLOv5s网络模型,改进算法极大提高检测的平均准确率和平均回归率,均值平均精度由74.1%提升至88.4%,获得14.3%的精度提升,同时也超过YOLOv3、YOLOXs、RetinaNet、Faster R-CNN等其他流行目标检测算法。【结论】针对陷阱相机所摄取森林野生动物图像背景与目标对比度低、遮挡重叠严重,致使检测误检率、漏检率高等问题,在检测算法中提出一系列改进措施,为我国森林野生动物的保护和数据获取提供一种新的可行性方案和思路。

Realization Methodology of a 5-axis Spline Interpolator in an Open CNC System

By making use of the advantages of non-uniform rational B-spline （NURBS） curves to represent spatial curves, an instruction format with double NURBS curves suitable for 5-axis coordinated real-time interpolation is presented to replace the current 5-axis coordinated linear interpolation method defective in low-speed, low-accuracy and enormous numerical control （NC） files in sculptured surface machining. A generation procedure of the NC files with the presented format is introduced and the method to realize the interpolation in an open computer numerical control （CNC） system is developed by ourselves. These illustrated the feasibility of the proposed method and its capability of avoiding all the shortages of 5-axis linear interpolation method.

基于Transformer改进YOLOv5的交通标志检测算法

交通标志检测作为自动驾驶的组成部分直接影响着行车安全。针对现有算法对图像中尺寸小、被遮挡的标志存在漏检、误检的问题,文中提出了基于改进YOLOv5的交通标志检测算法。首先对原模型注意力缺失的问题经过对比后构建了BiFormer-y,使模型可以更好获取长期依赖;接着针对层数较深造成的具有丢失特征的缺陷,利用残差结构重新设计检测层,从而更好地保留特征;最后对耦合头的空间错位问题引入解耦头并进行优化。CCTSDB2021的实验表明,精确率、召回率、mAP分别为97.0、95.9、97.9与先进工作相比具有明显优势。

Development of practical postprocessor for 5-axis machine tool with non-orthogonal rotary axes

In order to develop a practical postprocessor for 5-axis machine tool,the general equations of numerically controlled(NC) data for 5-axis configurations with non-orthogonal rotary axes were exactly expressed by the inverse kinematics,and a windows-based postprocessor written with Visual Basic was developed according to the proposed algorithm.The developed postprocessor is a general system suitable for all kinds of 5-axis machines with orthogonal and non-orthogonal rotary axes.Through implementation of the developed postprocessor and verification by a cutting simulation and machining experiment,the effectiveness of the proposed algorithm is confirmed.Compatibility is improved by allowing exchange of data formats such as rotational total center position(RTCP) controlled NC data,vector post NC data,and program object file(POF) cutter location(CL) data,and convenience is increased by adding the function of work-piece origin offset.Consequently,a practical post-processor for 5-axis machining is developed.

Circular Interpolation Algorithms of 5-Axis Simultaneous CNC System

Spatial circular arc can be machined conveniently by a 5-axis NC machine tool.Based on the data sampling method,circular interpolation in two-dimensional plane is discussed briefly.The key is to solve the problem of circular center ex- pressed in the workpiece coordinate system by means of the transformation matrix.Circular interpolation in three-dimensional space is analyzed in detail.The method of undetermined coefficient is used to solve the center of the spatial circle and the method of coor- dinate transformation is used to transform the spatial circle into the XY-plane.Circular arc in three-dimensional space can be ma- chined by the positional 5-axis machining and the conical surface can be machined by the continuous 5-axis machining.The velocity control is presented to avoid the feedrate fluctuation.The interpolation algorithms are tested by a simulation example and the inter- polation algorithms are proved feasible.The algorithms are applied to the 5-axis CNC system software.

Development of tool-path generation module for 5-axis control machining of impellers

An impeller is difficult to machine because of severe collision due to the complex shape,overlapping and twisted shape of the impeller blades.So,most computer aided manufacturing(CAM)software companies have developed CAM module for manufacturing impeller according to their CAM software.But these dedicated modules are difficult to use for inexperienced users.The purpose of this work is to develop a tool-path generation module for impellers.For this purpose,it is based on Visual Basic language and used CATIA graphical environment.The result of simulation for generated tool-path by the module is satisfactory.And it has slow processing speed compared to other commercial modules,but it is easy to use.

A Real-time NURBS Surface Interpolator for 5-axis Surface Machining

A real-time non-uniform rational B-spline （NURBS） surface interpolator is proposed and 5-axis machining method with a flat-end cutter is discussed. With the Taylor expansion and the coordinate transformation, the algorithms of NURBS interpolation, cutter effective machining radius, cutter offsetting and.inverse kinematics are deduced and implemented, respectively. Different from the conventional free-form surface machining, the proposed interpolator can real-time generate the motion commands of computer numerical control （CNC） machines with CC feedrate, rather than that of CL. An example part surface is demonstrated and the results of simulation show that the proposed method can be applied in actual 5-axis surface machining.

MOTION ERROR ESTIMATION OF 5-AXIS MACHINING CENTER USING DBB METHOD

In order to estimate the motion errors of 5-axis machine center, the double ball bar （DBB） method is adopted to realize the diagnosis procedure. The motion error sources of rotary axes in 5-axis machining center comprise of the alignment error of rotary axes and the angular error due to various factors, e.g. the inclination of rotary axes. From sensitive viewpoints, each motion error is possible to have a particular sensitive direction in which deviation of DBB error trace arises from only some specific error sources. The model of the DBB error trace is established according to the spatial geometry theory. Accordingly, the sensitive direction of each motion error source is made clear through numerical simulation, which is used as the reference patterns for rotational error estimation. The estimation method is proposed to easily estimate the motion error sources of rotary axes in quantitative manner. To verify the proposed DBB method for rotational error estimation, the experimental tests are carried out on a 5-axis machining center M-400 （MORISEIKI）. The effect of the mismatch of the DBB is also studied to guarantee the estimation accuracy. From the experimental data, it is noted that the proposed estimation method for 5-axis machining center is feasible and effective.

Establishment of a transformation system in close relatives of wheat under the assistance of TaWOX5

Species closely related to wheat are important genetic resources for agricultural production,functional genomics studies and wheat improvement.In this study,a wheat gene related to regeneration,TaWOX5,was applied to establish the Agrobacterium-mediated transformation systems of Triticum monococcum,hexaploid triticale,and rye(Secale cereale L.)using their immature embryos.Transgenic plants were efficiently generated.During the transformation process,the Agrobacterium infection efficiency was assessed by histochemical staining forβ-glucuronidase(GUS).Finally,the transgenic nature of regenerated plants was verified by polymerase chain reaction(PCR)-based genotyping for the presence of the GUS and bialaphos resistance(bar)genes,histochemical staining for GUS protein,and the QuickStix strip assay for bar protein.The transformation efficiency of T.monococcum genotype PI428182 was 94.4%;the efficiencies of four hexaploid triticale genotypes Lin456,ZS3297,ZS1257,and ZS3224 were 52.1,41.2,19.4,and 16.0%,respectively;and the transformation efficiency of rye cultivar Lanzhou Heimai was 7.8%.Fluorescence in situ hybridization(FISH)and genomic in situ hybridization(GISH)analyses indicated that the GUS transgenes were integrated into the distal or near centromere(proximal)regions of the chromosomes in transgenic T.monococcum and hexaploid triticale plants.In the transgenic hexaploid triticale plants,the foreign DNA fragment was randomly integrated into the AABB and RR genomes.Furthermore,the transgene was almost stably inherited in the next generation by Mendel’s law.The findings in this study will promote the genetic improvement of the three plant species for grain or forage production and the improvement of cereal species including wheat for functional genomics studies.

基于YOLO V5和Transformer的实时手语识别和翻译方法研究

针对手语识别和翻译的需求,文中首先结合OpenCV和Mediapipe hands模型,制作了一个包含多种手势、不同角度、明暗区域、简单和复杂背景等特定环境下的手语数据集,该数据集共计3000张图像。其次,为了实现更准确和实时的手语识别和翻译,选用YOLO V5模型进行了手语识别检测。在测试集上,平均准确率mAP(0.5)达99.1%。然后,结合Transformer模型进行手语孤立词连接,实现手语句子的生成。这些技术的应用能为听障人士提供更智能、便捷的交流方式,将在未来得到更广泛的应用。

融合Swin Transformer的YOLOv5口罩检测算法

针对口罩佩戴检测算法未平衡模型规模与检测精度的问题,提出了一种口罩佩戴检测改进算法。该算法以YOLOv5网络为基础框架:首先,应用轻量级Mixup数据增强方式和Mish激活函数以提高模型泛化能力;其次,引入Swin Transformer结构和ECA注意力机制来增强复杂场景下口罩目标的提取效率;第三,使用SIoU损失函数以提高检测精度;最后,设计了新的Neck网络卷积模块来实现模型轻量化。实验结果表明:相比于原始的YOLOv5算法,mAP提升2.9%,参数量减少54.2%,模型体积减少52.1%。该算法很好地平衡了模型规模与检测精度,在口罩检测实际场景中更具优势。

基于S-YOLO V5和Vision Transformer的视频内容描述算法

视频内容描述的自动生成是结合计算机视觉和自然语言处理等相关技术提出的一种新型交叉学习任务。针对当前视频内容生成描述模型可读性不佳的问题,本研究提出一种基于S-YOLO V5和Vison Transformer(ViT)的视频内容描述算法。首先,基于神经网络模型KATNA提取关键帧,以最少帧数进行模型训练;其次,利用S-YOLO V5模型提取视频帧中的语义信息,并结合预训练ResNet101模型和预训练C3D模型提取视频静态视觉特征和动态视觉特征,并对两种模态特征进行融合;然后,基于ViT结构的强大长距离编码能力,构建模型编码器对融合特征进行长距离依赖编码;最后,将编码器的输出作为LSTM解码器的输入,依次输出预测词,生成最终的自然语言描述。通过在MSR-VTT数据集上进行测试,本研究模型的BLEU-4、METEOR、ROUGEL和CIDEr分别为42.9、28.8、62.4和51.4;在MSVD数据集上进行测试,本研究模型的BLEU-4、METEOR、ROUGEL和CIDEr分别为56.8、37.6、74.5以及98.5。与当前主流模型相比,本研究模型在多项评价指标上表现优异。

改进的YOLOv5s遥感影像机场场面飞机小目标识别

遥感影像具有地物拍摄模糊以及背景环境复杂等特点,导致大面积地面物体识别准确率较低。针对此问题,提出了一种基于YOLOv5s改进网络模型。该模型对YOLOv5s中的骨干提取网络和颈部多尺度特征融合网络进行优化,引入Swin Transformer以获得更多关于目标物体的特征信息;同时对骨干网络中的模块做了修剪;此外,模型中还添加了坐标注意力机制来提升特征提取和融合效果。对于遥感数据集进行了小目标检测识别测试实验,改进后的YOLOv5s网络的mAP值为0.8375,比YOLOv5s网络模型提高了0.0225。实验结果表明,改进后的YOLOv5s网络模型对比YOLO系列网络和EfficientDet模型有效地提高了识别准确率、召回率以及mAP值,并且在训练时间上也比YOLOv5s减少了1/12。

改进YOLOv5s的钢材表面缺陷检测算法

为提高钢材表面缺陷检测的准确率,提出一种改进YOLOv5s的钢材表面缺陷检测算法。首先,在特征提取网络中引入Swin Transformer结构,增强网络对特征的感知能力;其次,添加坐标注意力机制,加强对重要特征信息的关注;最后,针对钢材缺陷的特点增加检测层,提升多尺度目标检测能力,并使用SIOU损失函数评估检测效果。将所提出的算法在公开数据集NEU-DET上进行消融实验,结果表明:所提算法能有效提高钢材表面缺陷目标检测的准确率。

CXCL5通过诱导血管钙化参与颈动脉斑块的形成

背景:CXC基序趋化因子5(CXC-motif chemokine 5,CXCL5)为上皮细胞衍生的中性粒细胞激活肽,研究发现其可能参与动脉病变。然而,CXCL5在血管钙化中的作用未见报道。目的:探讨CXCL5在颈动脉粥样硬化的血管钙化中的作用。方法:①细胞实验:将小鼠血管平滑肌细胞分成以下各组:成骨培养基组,Vector组(空白质粒转染到细胞中),CXCL5组(CXCL5质粒转染到细胞中),si-NC组(CXCL5阴性对照siRNA转染到细胞中),si-CXCL5组(CXCL5 siRNA转染到细胞中),Vector+LY2157299组和CXCL5+LY2157299组(细胞转染24 h后,将转化生长因子β受体1激酶抑制剂LY2157299加入细胞中)。进行茜素红染色、碱性磷酸酶染色和钙含量测定以评估血管平滑肌细胞成骨分化水平。②动物实验:48只ApoE-/-小鼠随机分成4组:Con+si-NC组、Con+si-CXCL5组、CAS+si-NC组和CAS+si-CXCL5组,前2组不造模,尾静脉注射si-NC或si-CXCL5慢病毒;后2组制备颈动脉粥样硬化模型,尾静脉注射si-NC或si-CXCL5慢病毒。采用Von Kossa染色和免疫组织化学染色评估小鼠颈动脉血管钙化以及CXCL5、转化生长因子β受体1表达情况。结果与结论:①CXCL5组细胞Runt相关转录因子2蛋白水平上调、α-平滑肌肌动蛋白水平下调,si-CXCL5组中的发现与其相反;CXCL5过表达上调了转化生长因子β受体1水平,而CXCL5敲低抑制了转化生长因子β受体1水平。②与Vector组相比,CXCL5组细胞茜素红染色的强度、碱性磷酸酶活性和钙含量显著增加(P<0.05);与si-NC组相比,si-CXCL5组上述2项指标显著降低(P<0.05);当用LY2157299抑制转化生长因子β受体1表达时,CXCL5对平滑肌细胞的成骨转化作用减弱。③与Con+si-NC组相比,CAS+si-NC组大鼠颈动脉中CXCL5蛋白表达和血管钙化面积显著增加(P<0.05);与CAS+si-NC组相比,CAS+si-CXCL5组颈动脉中上述2项指标显著降低(P<0.05)。④与Con+si-NC组相比,CAS+si-NC组大鼠颈动脉中Runt相关转录因子2蛋白表达显著增加(P<0.05)和α-平滑肌肌动蛋白表达显著降低(P<0.05);与CAS+si-NC组相比,CAS+si-CXCL5组颈动脉中上述2项指标呈相反变化(P<0.05)。⑤结果说明,CXCL5通过激活转化生长因子β受体1通路诱导血管平滑肌细胞成骨样转化,抑制CXCL5表达对于改善颈动脉粥样硬化小鼠颈动脉血管钙化是有效的。

基于改进YOLOv5s 的白酒瓶盖瑕疵检测

目的瓶装白酒生产过程中,瓶盖表面瑕疵会影响产品外观质量。针对白酒瓶盖表面瑕疵检测效率低和目标检测效果差的问题,提出一种基于YOLOv5s的改进算法DTS-YOLO。方法首先,在主干网络中引入可变形卷积,以提高模型对极端长宽比瑕疵的检测精度。其次,引入Transformer编码块,使网络聚焦于提取图像的全局信息。最后,在颈部网络构建C3SE-Lite模块,将C3模块嵌入SE注意力模块的同时引入Ghost卷积,减少参数量的同时,增强对瓶盖瑕疵的检测能力。结果实验结果表明,本文所提方法相较于基础网络,参数量减少了10%,平均精度均值达95%,平均检测速度达30帧/s。结论本文方法有效实现了白酒瓶盖表面瑕疵快速、准确地检测,可广泛应用于瓶装白酒生产过程中瓶盖表面检测。

基于改进YOLOv5s的储粮害虫检测方法研究

针对当前因检测背景复杂、粮虫个体较小等因素造成的储粮害虫检测准确率较低、漏检、误检的现状,提出一种改进YOLOv5s的检测方法。使用MobileNetv3作为YOLOv5s模型的主干特征提取网络,并将其SE注意力机制修改为ECA模块,减少计算量和参数量;同时,调整PANet网络为加权双向特征金字塔BiFPN结构,增强特征融合能力;颈部网络部分引入Swin Transformer,解决全局提取特征不足的问题,提升识别精度;采用EIOU损失函数,提高模型的收敛速度。结果表明:改进后的模型mAP达到97.8%,FPS达到133.3,与主流目标检测模型比较,表现出较强的鲁棒性和泛化能力;改进的YOLOv5s检测模型能克服复杂环境的影响,显著提高储粮害虫在密集分布及遮挡条件下的检测效果。研究结果为移动设备上实现储粮害虫实时检测提供技术参考。

融合Transformer和改进PANet的YOLOv5s交通标志检测

针对交通标志检测速度慢和目标大小与类别极度不平衡等问题,提出一种融合Transformer和改进PANet网络的YOLOv5s交通标志检测算法。首先在不增加模型复杂度的前提下,将主干网络末端与Transformer融合以提高网络特征提取能力;其次由于所采用交通标志数据集的目标尺度太小,导致网络32倍大尺度检测层检测效果不佳,故不采用相关网络层,同时采用K-means算法得出适合的预测候选框;然后改进损失函数以解决正负样本极度不平衡问题。最后将所提出的改进算法在Jetson AGX Xavier平台上部署验证。实验结果表明,所提算法检测性能更佳,其准确率和召回率在原网络的基础上分别提高了2.2%和0.7%,模型参数量和计算复杂度分别减少了25.8%和10.1%。在Xavier上的检测速度达到76FPS,满足实时交通标志检测的要求且易于在实际场景部署。

基于YOLOv5的零件识别轻量化算法

为了解决现有的基于深度学习的零件识别模型参数量过大、检测速度慢、检测精度低的问题,以YOLOv5模型为基础,提出了结合轻量级网络和Transformer的零件识别算法。首先,设计了一种轻量级主干特征提取网络,以减少网络的参数量和计算量,并提升推理速度;其次,将Transformer模块与C3模块融合构成C3TR模块,以增强小目标的检测能力;最后,引入噪音净化模块,通过过滤噪音来提高零件识别模型的准确率。模型的检测平均准确率和平均召回率分别达到了86.7%和85.5%,相较原模型分别提升了和24.2%和17.4%。实验结果表明,改进后的模型在实现模型轻量化的同时,具有更快的检测速度和更高的识别准确率。