基于改进YOLOv7-tiny的高空作业人员安防装备检测算法

针对现有高空作业人员安防装备检测算法参数量较大且检测目标相对单一,难以适应复杂的高空作业场景等问题,提出一种基于改进YOLOv7-tiny的高空作业人员安防装备检测算法。首先,将主干网络重新设计为更轻量的YOLOv7-FasterNet,并调整空间金字塔池化结构,实现模型参数量的压缩;其次,在ELAN-L模块中扩展梯度传输路径的分支,解决了模型压缩造成的通道信息缺失问题,提升了特征信息的提取能力;最后,将网络中下采样部分替换为Involution模块,降低参数冗余,增强网络对全局的捕获能力。实验结果表明,改进的YOLOv7-tiny算法能够更好地适应复杂高空作业场景,在开源数据集上具备良好的性能。该算法的平均检测精度达到94.7%,较原模型提升1.5%,参数量较原模型下降11.6%,实验结果验证了算法改进措施的有效性。

Application of Isogeometric Analysis Method in Three-Dimensional Gear Contact Analysis

Gears are pivotal in mechanical drives,and gear contact analysis is a typically difficult problem to solve.Emerging isogeometric analysis(IGA)methods have developed new ideas to solve this problem.In this paper,a threedimensional body parametric gear model of IGA is established,and a theoretical formula is derived to realize single-tooth contact analysis.Results were benchmarked against those obtained from commercial software utilizing the finite element analysis(FEA)method to validate the accuracy of our approach.Our findings indicate that the IGA-based contact algorithmsuccessfullymet theHertz contact test.When juxtaposed with the FEA approach,the IGAmethod demonstrated fewer node degrees of freedomand reduced computational units,all whilemaintaining comparable accuracy.Notably,the IGA method appeared to exhibit consistency in analysis accuracy irrespective of computational unit density,and also significantlymitigated non-physical oscillations in contact stress across the tooth width.This underscores the prowess of IGA in contact analysis.In conclusion,IGA emerges as a potent tool for addressing contact analysis challenges and holds significant promise for 3D gear modeling,simulation,and optimization of various mechanical components.

Lateral tarsal strip procedure for involutional ectropion: A retrospective analysis of 85 cases and a comprehensive literature review

Purpose:To evaluate the postoperative effect of the lateral tarsal strip(LTS)procedure in treating lower eyelid involutional ectropion.Methods:A retrospective study was performed on 85 eyelids in 67 patients with involutional ectropion who underwent LTS procedure.Pre-and postoperative lower eyelid laxity and ocular symptoms as well as perioperative complications were evaluated.Snap back test was performed to evaluate the elasticity of lateral canthal tendon loosening and,a lower lid distraction test was performed to evaluate the degree of severity before surgery.Ocular surface diseases were evaluated by the Ocular Surface Disease Index(OSDI),and symptoms including conjunctivitis,corneal ulcer,dry eye syndrome,and ocular pain were recorded.All patients were evaluated within one week and during the follow-up period of 4.2
8.3 months.Furthermore,we reviewed the studies that also investigated the surgical effect of the LTS procedure in the literature from 1979 to 2019.Results:The success rate was 95%.Only four eyelids required a second surgical intervention.Seventy-three(86%)eyelids had an excellent position after surgery,9(11%)only little improvement,and three had no improvement.No significant difference was found in the postoperative effects between different degrees of ectropion(p>0.05).No statistical correlation was found between surgical improvements and the ectropion severity(P>0.05).Fiftytwo out of 85 eyes had no discomfort after the surgery.Mild complications included epiphora in 13 eyes(three cases caused by lacrimal punctum eversion),ocular pain in 12 eyes,wound hemorrhage in 12 eyes,and edema in 9 eyelids immediately after surgery,in which 91.2%(n=21)disappeared within one week and did not need any further treatment.Conclusions:The lateral tarsal strip procedure can provide an aesthetically pleasing result for correcting the mild to moderate lower eyelid ectropion while maintaining decent eyelid function.

基于改进SSD的钢材表面缺陷检测

工业生产过程中,钢材表面缺陷的检测对于钢材的质量控制发挥着十分重要的作用,针对钢材表面缺陷检测中存在的检测精度低、检测速度慢等问题,提出了一种钢材表面缺陷检测的改进SSD算法。在所提算法中,采用Transformer多头注意力机制模块代替原SSD结构中的Conv5_1层,以提高小目标检测的能力;原SSD结构中的Conv7操作替换为Involution算子操作,以减少运算的参数量;对网络结构进行特征融合处理,以更全面地检测特征图中所包含的信息。利用NEU-DET数据集进行实验,实验结果表明改进后的SSD算法是有效的,可以高效检测到钢材表面的小目标缺陷,相比改进前平均检测精度提高了4.5%,检测速度提高了13.6%。

基于弱监督数据集的猪只图像实例分割

在智慧养殖研究中,基于深度学习的猪只图像实例分割方法,是猪只个体识别、体重估测、行为识别等下游任务的关键。为解决模型训练需要大量的逐像素标注图像,以及大量的人力和时间成本的问题,采用弱监督猪只分割策略,制作弱监督数据集,提出一种新的特征提取骨干网络RdsiNet:首先在ResNet-50残差模块基础上引入第2代可变形卷积,扩大网络感受野;其次,使用空间注意力机制,强化网络对重要特征的权重值;最后引入involution算子,借助其空间特异性和通道共享性,实现加强深层空间信息、将特征映射同语义信息连接的功能。通过消融实验和对比实验证明了RdsiNet对于弱监督数据集的有效性,实验结果表明其在Mask R-CNN模型下分割的mAP_(Semg)达到88.6%,高于ResNet-50、GCNet等一系列骨干网络;在BoxInst模型下mAP_(Semg)达到95.2%,同样高于ResNet-50骨干网络的76.7%。而在分割图像对比中,使用RdsiNet骨干网络的分割模型同样具有更好的分割效果:在图像中猪只堆叠情况下,能更好地分辨猪只个体;使用BoxInst训练的模型,测试图像中掩码具有更高的精细度,这更有利于开展下游分析。

一种基于Involution卷积的三维人体重建方法

针对从视频中恢复三维人体模型运动序列时,由于图像特征提取能力有限而导致三维人体模型运动序列重建效果不佳的问题,提出了一种基于Involution卷积的三维人体重建方法。首先为了引入自注意力机制,在ResNet50网络结构中加入Involution算子,获取视频图像帧的特征向量,然后使用姿态估计网络和形状估计网络获取人体姿势以及形状参数,最后使用蒙皮多人线性模型(skinned multi-person linear model, SMPL)生成三维人体模型的运动序列。在三维姿态户外数据集(3D pose in the wild, 3DPW)上与视频人体姿态形状估计推理(video inference for body pose and shape estimation, VIBE)方法以及时间一致性网格恢复(temporally consistent mesh recovery, TCMR)方法进行对比实验,平均精度相比于VIBE、TCMR分别提升了3.1%、0.7%,能够为运动捕捉、三维人体动画制作等工作提供更为准确的三维人体模型。

轻量化篮球裁判手势识别算法

针对一般手势识别算法的参数量、计算量和精度难以平衡的问题,提出一种轻量化篮球裁判手势识别算法。该算法在YOLOV5s(You Only Look Once Version 5s)算法的基础上进行重构:首先,用Involution算子替代CSP1_1的卷积算子,以扩大上下文信息捕获范围并减少核冗余;其次,在C3模块后加入协同注意力(CA)机制,以得到更强的手势特征提取能力;然后,用轻量化内容感知上采样算子改进原始上采样模块,并将采样点集中在目标区域而忽略背景部分;最后,利用以SiLU作为激活函数的Ghost-Net进行轻量化剪枝。在自制的篮球裁判手势数据集上的实验结果表明,该轻量化篮球裁判手势识别算法的计算量、参数量和模型大小分别为3.3 GFLOPs、4.0×10^(6)和8.5 MB,与YOLOV5s算法相比,分别减少了79%、44%和40%,mAP@0.5为91.7%,在分辨率为1920×1280的比赛视频上的检测帧率达到89.3 frame/s,证明该算法能满足低误差、高帧率和轻量化的要求。

改进人脸特征矫正网络的遮挡人脸识别方法

现有人脸识别模型受口罩等遮挡因素影响导致准确率无法提升。当前主流研究方法将有无遮挡场景分开训练后,整合应用于多场景。针对遮挡人脸识别模型的局限性,提出一种改进人脸特征矫正网络(FFR-Net)模型。该模型可同时用于有无遮挡人脸识别并应用于口罩与眼镜遮挡两种识别场景中。人脸特征矫正网络模型提出了一种人脸特征矫正模块,为保证充分利用无遮挡区域特征信息,在该模块中的空间分支引入involution算子扩大图像信息交互区域,增强在空间范围内面部特征信息;在通道分支引入坐标注意力机制,捕获跨通道信息以增强特征表示,利于模型准确地定位识别目标区域;将Meta-ACON作为该模块新的动态激活函数,通过动态调整线性或非线性程度以提高模型泛化能力和计算准确度。最后,利用改进的人脸特征矫正网络模型在CASIA-Webface经处理的有无口罩遮挡人脸数据集上进行训练,其在LFW经处理的有无口罩遮挡数据集、Meglass数据集上的测试结果准确率分别达到了82.50%和89.75%,优于现有算法,验证了所提方法的有效性。

双线性内卷神经网络用于眼底疾病图像分类

由于眼底图像具有复杂程度高、个体差异弱、类间距离短等特点,纯卷积神经网络(CNN)和基于注意力的网络并不能在眼底疾病图像分类任务上达到令人满意的精度。因此,采用involution算子实现了注意力双线性内卷神经网络(ABINN)模型用于眼底疾病图像分类。ABINN模型的参数量仅是传统双线性卷积神经网络(BCNN)模型的11%,并提取了眼底图像的底层语义信息和空间结构信息进行二阶特征融合,是CNN和注意力方法的有效并联。此外,提出了两种基于involution算子实现注意力计算的实例化方法:基于图块的注意力子网络(AST)和基于像素的注意力子网络(ASX),这两种方法可以在CNN的基础结构内完成注意力的计算,从而使双线性子网络能在同一个架构下训练并进行特征融合。在公开眼底图像数据集OIA-ODIR上进行实验,结果显示ABINN模型的精度为85%,比通用BCNN模型提高了15.8个百分点,比TransEye模型提高了0.9个百分点。

GRU和Involution改进的深度伪造视频检测方法

近年来深度伪造视频在网络上广泛传播,造成了负面影响。针对现有检测模型准确率低和信息提取不够充分和全面的问题,提出了一种GRU(gated recurrent unit)和Involution改进的深度伪造视频检测方法。首先使用VGG19作为主干网络提取空间特征,并将Involution算子嵌入主干网络,从空间和通道信息两方面加强了人脸图像的空间建模能力。然后通过主胶囊层关注特征的位置信息和使用GRU提取帧间的时序特征。最后在训练模型阶段使用focalloss作为损失函数来平衡样本。在Deepfakes、FaceSwap和Celeb-DF数据集中进行测试,实验结果表明该方法优于主流检测方法,改进对比实验进一步证明了检测方法的有效性。

注意力机制与Involution算子改进的人脸表情识别

针对复杂人脸表情识别面临背景干扰、空间信息分布不均匀等问题,提出一种注意力机制和Involution算子改进的人脸表情识别方法,该方法以VGG19为基线网络,前端引入注意力机制提取表情强相关特征,抑制背景干扰,并利用联合正则化策略平衡和改善特征数据分布,提高模型训练质量;后端采用密集连接加强有效特征复用,提取高层语义信息。所提方法在CK+、FER2013、RAF-DB等3个公开数据集上进行了验证,准确率均取得显著提高,且优于当前诸多先进方法。此外,为提高网络处理复杂条件下的数据集,在其后端引入Involution算子替代部分卷积层,提高了空间多样性信息学习能力。实验结果表明,所提模型可有效提高RAF-DB等复杂数据集的人脸表情识别准确率。

基于多种卷积核特征提取自适应融合的滚动轴承故障诊断方法

传统卷积神经网络模型采用单一类型卷积核,面对复杂工况下的实际数据时存在特征提取不充分,故障识别率低等问题,因此提出了一种基于多种卷积核特征提取自适应融合的滚动轴承故障诊断方法(MCK-CNN)。首先,将轴承一维振动信号经过小波变换转换为二维时频图,将时频图经过一个共同的特征提取网络初步进行特征提取后并行经过一个常规的Convolution和Involution卷积网络;然后,经过2个由不同卷积核构成的网络进行不同方式的特征提取并通过CBAM注意力模块将两类特征自适应地融合;最后,将融合特征输入全连接层并通过Softmax函数输出分类结果。CWRU和实验室轴承数据集的试验结果表明,MCK-CNN模型的训练效率和故障识别率均较高。

基于Involution算子和交叉注意力机制的合成语音检测方法

随着科学技术的迅速发展,基于深度学习生成的合成语音给语音认证系统和网络空间安全带来了新的挑战。针对现有检测模型准确率较低和语音特征挖掘不够充分的问题,提出了一种基于Involution算子和交叉注意力机制改进的合成语音检测方法。前端将语音数据提取线性频率倒谱系数(LFCC)特征和恒定Q变换(CQT)谱图特征,两个特征分别输入到后端的双分支网络中。后端网络使用ResNet18作为主干网络先进行浅层的特征学习,并将Involution算子嵌入主干网络,扩大特征图像学习区域,增强在空间范围内学习到的频谱图像特征信息。同时在训练分支之后引入cross-attention交叉注意力机制,使LFCC特征和CQT谱图特征构建交互的全局信息,强化模型对特征的深层挖掘。所提模型在ASVspoof 2019 LA测试集上取得了0.84%的等错误率和0.026的最小归一化串联检测代价函数的实验结果,展现了优于主流的检测模型。结果表明,改进的模型能够有效融合不同的频谱特征,提高模型的特征学习能力,从而强化模型的检测能力。

Numerical Simulation of Dust Removal in the Cyclone Collector of a Straw Crusher Based on a Discrete Phase Model

The cyclone dust collector is an important subsystem of straw crushers used in agriculture.In the present study,a new type of dust collector with involute morphology is proposed to obtain better dust removal efficiency with respect to that of classical tangential and spiral dust collectors.A discrete phase model(DPM)method is used in synergy with a turbulence model,and the SIMPLE algorithm to simulate the flow field inside the dust collector and the related particle dynamics.It is shown that the internal flow field features a primary swirl,a secondary swirl and blockage effects.Moreover,for the involute dust collector,the tangential velocity in the initial stage and the pressure in the high-pressure area are larger than those obtained for the classical types.The dust removal efficiency is 37.11%,25.3%,and 16.37%for the involute type dust collector,the tangential type and the spiral type,respectively.

带有对合的素环的微分恒等式

Let R be a prime ring of characteristic different from two with the second involution∗andαan automorphism of R.An additive mapping F of R is called a generalized(α,α)-derivation on R if there exists an(α,α)-derivation d of R such that F(xy)=F(x)α(y)+α(x)d(y)holds for all x,y∈R.The paper deals with the study of some commutativity criteria for prime rings with involution.Precisely,we describe the structure of R admitting a generalized(α,α)-derivation F satisfying any one of the following properties:(i)F(xx∗)−α(xx∗)∈Z(R),(ii)F(xx∗)+α(xx∗)∈Z(R),(iii)F(x)F(x∗)−α(xx∗)∈Z(R),(iv)F(x)F(x∗)+α(xx∗)∈Z(R),(v)F(xx∗)−F(x)F(x∗)∈Z(R),(vi)F(xx∗)−F(x∗)F(x)=0 for all x∈R.Also,some examples are given to demonstrate that the restriction of the various results is not superfluous.In fact,our results unify and extend several well known theorems in literature.

基于改进SSD算法的光伏板缺陷检测研究

在光伏产业中,光伏板缺陷的检测至关重要,针对目标检测算法SSD在光伏板缺陷检测中精度不高、速度不快、存在漏检、误检等问题,本文提出了一种改进的SSD算法,主要从三个方面改进:首先,使用融合了ECA注意力机制的Resnet50网络替换原SSD中的VGG-16网络,减少了计算量,提高了检测精度;其次,使用Involution算子替换原SSD中的Conv7卷积层,使模型更加轻量;最后,引入FocalLoss损失函数解决正负样本不平衡的问题。经实验分析,改进的SSD算法相较于原SSD算法的mAP提升了4.41个百分点,同时检测速度提升了6.55frame/s。

一种基于Involution和BERT词向量的情感分析算法

文本情感分析是当前数据分析和数据挖掘的一个重要的研究方向,提出了一种融合Involution和BERT词向量的情感分析算法,首先,通过对预处理之后的数据传入BERT模块,用学习质量更高的文本数据词向量表示;其次,利用Involution核函数方法进行更好的特征分析和提取;最后,在三种标准数据集上进行了广泛的实验,研究结果表明,基于Involution和BERT的情感分析模型能够有效地提取文本数据中的有用信息,并准确地进行情感分析,从而更好地了解用户的情感偏好.

A Procedure for the Squaring of a Circle (of Any Radius)

This paper presents a graphical procedure for the squaring of a circle of any radius. This procedure, which is based on a novel application of the involute profile, when applied to a circle of arbitrary radius (using only an unmarked ruler and a compass), produced a square equal in area to the given circle, which is 50 cm2. This result was a clear demonstration that not only is the construction valid for the squaring of a circle of any radius, but it is also capable of achieving absolute results (independent of the number pi (π), in a finite number of steps), when carried out with precision.

雷峰塔下的张爱玲:《雷峰塔》、《易经》,与“回旋”和“衍生”的美学

本文通过张爱玲英文小说《雷峰塔》和《易经》及其它文本的比较阅读，提出以下观察：一、相对于写实／现实主义作为中国现代文学的主流形式，张爱玲穿越修辞、文类以及语言的界限，以重复书写发展出一种“衍生”（derivation）、“回旋”（involution）的美学。二、透过对自身故事的多重叙述，张爱玲颠覆传统家族历史的大叙事。通过书写，她化记忆为技艺，也重塑过往吉光片羽的存在与形式。三、张爱玲创作中回旋、衍生的倾向也带出了一种独特的（文学）史观。张的史观表明她后四十年的创作其实不只限于她以各种形式重写的自传故事，也同时包括另外两项计划：一是将吴语的《海上花列传》翻译为国语，再翻译成英文；另外则是通过细读文本、文献考证以及传记研究的方式参详《红楼梦》。

Renovation and Involution — Paradox of Chinese Literature in Modernity

With a special matter, Chinese literature has differences with traditional theories from western countries in essence, so that Chinese literature demonstrates its renovation and involution. The main reason is its trends of modernity. The transformation from oversea exchange is very important. This paper mainly explains the specific content of paradox of Chinese literature in mo dernity, and explores the reasons for differences between China and western countries.