基于Focal Loss改进LightGBM的供水管网毛刺数据检测

针对数据不平衡导致的管网毛刺数据检测召回率偏低问题,提出一种Focal Loss改进LightGBM的管网毛刺数据检测方法。首先,结合管网毛刺数据的特点,针对性构造邻域相关特征。其次,将Focal Loss函数引入LightGBM,提高模型对难以检测的毛刺样本的权重,并对Focal Loss不同的参数取值进行实验,以平衡精确率与召回率。最后,选择不同参数的Focal Loss进行模型融合,进一步提升模型对不平衡毛刺数据的检测性能。在某市供水管网的真实数据上进行实验,结果表明,对比基于交叉熵损失函数的单一模型,本文提出的Focal Loss改进后的融合模型在毛刺数据上召回率和F1值的提升幅度达33.3和18个百分点,但毛刺数据的精确率还有待进一步提升。本文所提方法从损失函数入手,动态调整难易样本的权重,有效地提升了不平衡数据下的毛刺数据的检测性能。

基于自适应空间特征增强的多视图深度估计

为了提高多视图深度估计结果精度,提出一种基于自适应空间特征增强的多视图深度估计算法。设计了由改进后的特征金字塔网络(feature pyramid network,FPN)和自适应空间特征增强(adaptive space feature enhancement,ASFE)组成的多尺度特征提取模块,获取到具有全局上下文信息和位置信息的多尺度特征图像。通过残差学习网络对深度图进行优化,防止多次卷积操作出现重建边缘模糊的问题。通过分类的思想构建focal loss函数增强网络模型的判断能力。由实验结果可知,该算法在DTU(technical university of denmark)数据集上和CasMVSNet(Cascade MVSNet)算法相比,在整体精度误差、运行时间、显存资源占用上分别降低了14.08%、72.15%、4.62%。在Tanks and Temples数据集整体评价指标Mean上该模型优于其他算法,证明提出的基于自适应空间特征增强的多视图深度估计算法的有效性。

基于YOLOv8-NFMC的带钢表面缺陷检测算法

针对YOLOv8算法在应用于带钢表面缺陷检测时存在漏检和错检等问题,提出了一种改进YOLOv8算法。针对数据集中的小目标的标签,在原损失CIOU的基础上面加入标准化高斯瓦瑟斯坦距离(normalized Gaussian Wasserstein distance,NWD),提升模型对小目标缺陷的检测能力;采用聚焦调制(focal modulation)替换YOLOv8模型的空间池化金字塔(spatial pyramid pooling-fast,SPPF),在轻量化的同时,提高多尺度特征的表达能力;采用移动翻转瓶颈卷积(mobile inverted bottleneck conv,MBConv)替换C2f中的Conv构建新模块C2f-MB,同时使用C2f-MB替换原有的C2f模块,增强特征表达能力和多尺度特征融合能力;在主干部分加入卷积块注意力机制(convolutional block attention module,CBAM)来抑制背景干扰,能更好捕获全局信息,提升了主干部分的特征提取能力。实验结果表明,改进后的YOLOv8算法在计算量下降的同时,mAP@0.5提高了3%,对漏检和错检等问题有明显改善。

基于聚焦滤波的林业专题地图制图综合方法

森林资源规划设计调查成果总是以连续的面状图斑呈现。当大比例尺二类调查成果图缩编为更小比例尺地图时,常常需要制图综合。如何在制图综合时既要简化多边形边界,又要保持各类型林相面积和位置的相对不变,是林业专题制图的一个难题。经大量实验发现,将实施制图综合的连续面状图斑要素从矢量格式转换为栅格格式,经Focal滤波器以Majority方式滤波处理后,再将滤波后结果从栅格数据转换为矢量数据,可较好地实现森林资源专题连续面状要素信息的制图综合。以云南省某地一个1万hm^(2)的实验区为例,使用该方法将1∶2.5万的森林资源二类调查成果图缩编为1∶25万。结果表明,缩编后,多边形边界简化,目视效果较好,各类型林相面积变化绝对值平均为0.2%,位置精确度平均值为94.68%。此方法已应用于云南森林资源状况图集相关专题图等的生产实践。

基于Res2Net的人脸表情识别方法

为解决自然条件下人脸表情识别易受角度、光线、遮挡物的影响以及人脸表情数据集各类表情数量不均衡等问题,提出基于Res2Net的人脸表情识别方法。使用Res2Net50作为特征提取的主干网络,在预处理阶段对图像随机翻转、缩放、裁剪进行数据增强,提升模型的泛化性。引入广义平均池化(generalized mean pooling, GeM)方式,关注图像中比较显著的区域,增强模型的鲁棒性;选用Focal Loss损失函数,针对表情类别不平衡和错误分类问题,提高较难识别表情的识别率。该方法在FER2013数据集上准确率达到了70.41%,相较于原Res2Net50网络提高了1.53%。结果表明,在自然条件下对人脸表情识别具有更好的准确性。

基于改进U-Net的遥感图像语义分割

全卷积神经网络在遥感图像语义分割中得到了广泛应用,该方法地物分类精度和效率较高,但对地物分布不均匀遥感图像占比较少地物的分类准确率较低。为了提高遥感图像的分类精度,本文通过添加先验知识方法丰富输入数据特征,采用密集链接方式提高上下采样过程中特征的重复利用率,采用可以优化交并比的损失函数Dice Loss和可以提高难分类类别精度的损失函数Focal Loss相加组合作为网络模型的损失函数,采用LayerScale模块加快模型收敛、抑制无用特征、突出有效特征的方式,对U-Net的输入、网络结构、损失函数进行改进,优化语义分割效果。结果表明,基于高分影像数据集(GID)改进的U-Net相较于原始U-Net像素精度、均类像素精度、平均交并比分别提高了0.0233、0.0409、0.0665,提升了地物分类精度,取得了较好的分类效果。

基于Unet+Attention的胸部CT影像支气管分割算法

目前肺气管分割中,由于CT图像灰度分布复杂,分割目标像素近似,易造成过分割;而且肺气管像素较少,难以得到更多目标特征,造成细小肺气管容易被忽略。针对这些难点,本研究提出结合Unet网络和注意力机制的肺气管分割算法,注意力机制使用的是关注通道域和空间域的卷积块注意力模型(CBAM),该模型提高了气管特征权重。在损失函数方面,针对原始数据中正负样本失衡的问题,引入focal loss损失函数,该函数对标准交叉熵损失函数进行了改进,使难分类样本在训练过程中得到更多关注;最后通过八连通域判断将孤立点去除,保留较大的几个连通域,即最后的肺气管部分。选用由合作医院提供的24组CT影像和43组CTA影像,共计26157张切片图像作为数据集,进行分割实验。结果表明,分割准确率能够达到0.86,过分割率和欠分割率均值为0.28和0.39。经过注意力模块和损失函数的消融实验,在改进前的准确率、过分割率和欠分割率分别为0.81、0.30、0.40,可见其分割效果均不如Unet+Attention方法。与其他常用方法在相同条件下进行比较后,在保证过分割率和欠分割率不变的情况下,所提出的算法得到了最高的准确率,较好地解决了细小气管分割不准确的问题。

社交平台不平衡文本数据处理与应用研究

随着社会信息化程度加深,运用自然语言处理技术从海量网络数据中筛选提取有效信息,具有重要的实用价值。然而,从社交平台收集的文本数据存在有效信息类别数据量少、类别不平衡等问题。因此,提出SimDyFeFL方法解决中文应急关联文本识别任务的数据不均衡问题,EdaDyFeFL方法解决英文网络暴力检测任务的数据不均衡问题。采用SimBERT与EDA方法将类间差异较大的原始数据增强至类间数量相近后,融合加入动态反馈过程的Focal Loss函数对各类别加权,并设计BERT、RoBERTa与BERT_DPCNN作为文本分类模型进行三个阶段的对比实验,证明提出方法的有效性。在中、英文两个真实数据集上的大量实验表明,使用SimDyFeFL与EdaDyFeFL改进后的文本分类模型综合性能提升显著,中文模型准确率最高提升7.70个百分点,英文模型准确率最高提升5.15个百分点。与Kaggle平台已有研究取得的最好成绩相比,英文模型准确率高出了2.92个百分点,Macro F1值与Weighted F1值分别高出2.83个百分点与2.95个百分点。

Diagnostic and therapeutic role of endoscopic ultrasound in liver diseases:A systematic review and meta-analysis

BACKGROUND In hepatology,the clinical use of endoscopic ultrasound(EUS)has experienced a notable increase in recent times.These applications range from the diagnosis to the treatment of various liver diseases.Therefore,this systematic review summarizes the evidence for the diagnostic and therapeutic roles of EUS in liver diseases.AIM To examine and summarize the current available evidence of the possible roles of the EUS in making a suitable diagnosis in liver diseases as well as the therapeutic accuracy and efficacy.METHODS PubMed,Medline,Cochrane Library,Web of Science,and Google Scholar databases were extensively searched until October 2023.The methodological quality of the eligible articles was assessed using the Newcastle-Ottawa scale or Cochrane Risk of Bias tool.In addition,statistical analyses were performed using the Comprehensive Meta-Analysis software.RESULTS Overall,45 articles on EUS were included(28 on diagnostic role and 17 on therapeutic role).Pooled analysis demonstrated that EUS diagnostic tests had an accuracy of 92.4%for focal liver lesions(FLL)and 96.6%for parenchymal liver diseases.EUS-guided liver biopsies with either fine needle aspiration or fine needle biopsy had low complication rates when sampling FLL and parenchymal liver diseases(3.1%and 8.7%,respectively).Analysis of data from four studies showed that EUS-guided liver abscess had high clinical(90.7%)and technical success(90.7%)without significant complications.Similarly,EUS-guided interventions for the treatment of gastric varices(GV)have high technical success(98%)and GV obliteration rate(84%)with few complications(15%)and rebleeding events(17%).CONCLUSION EUS in liver diseases is a promising technique with the potential to be considered a first-line therapeutic and diagnostic option in selected cases.

基于SE-RetinaNet的面向玻璃面板的小尺寸低显著性缺陷检测

玻璃面板中的缺陷具有低显著、尺寸小、形态多样、数量少等特点,现有先进目标检测算法难以胜任玻璃面板的质检任务。基于此,提出了SE-RetinaNet—一种面向玻璃面板的小尺寸低显著性的缺陷检测算法。该算法在特征金字塔的顶层和底层引入了SE注意力机制和自注意力机制,增强网络对底层小尺寸特征的提取能力并强化顶层网络捕捉特征的长距离依赖关系的能力,同时在网络末端引入定位子网络SE-Regression,通过结合残差块和Inception模块的优点加强了定位的准确度同时防止网络退化。实验结果表明,所提算法能有效检测玻璃面板中各种尺寸的低显著性缺陷,其检测性能优于现有经典目标检测的算法,能够在玻璃面板缺陷检测问题上发挥较好的性能。

Rapid report of the December 18,2023 M_(S)6.2 Jishishan earthquake,Gansu,China

On December 18,2023,the Jishishan area in Gansu Province was jolted by a M_(S) 6.2 earthquake,which is the most powerful seismic event that occurred throughout the year in China.The earthquake occurred along the NWtrending Lajishan fault(LJSF),a large tectonic transformation zone.After this event,China Earthquake Networks Center(CENC)has timely published several reports about seismic sources for emergency responses.The earthquake early warning system issued the first alert 4.9 s after the earthquake occurrence,providing prompt notification that effectively mitigated panics,injuries,and deaths of residents.The near real-time focal mechanism solution indicates that this earthquake is associated with a thrust fault.The distribution of aftershocks,the rupture process,and the recorded amplitudes from seismic monitoring and GNSS stations,all suggest that the mainshock rupture predominately propagates to the northwest direction.The duration of the rupture process is~12 s,and the largest slip is located at approximately 6.3 km to the NNW from the epicenter,with a peak slip of 0.12 m at~8 km depth.Seismic station N0028 recorded the highest instrumental intensity,which is 9.4 on the Mercalli scale.The estimated intensity map shows a seismic intensity reaching up to IX near the rupture area,consistent with field survey results.The aftershocks(up to December 22,2023)are mostly distributed in the northwest direction within~20 km of the epicenter.This earthquake caused serious casualties and house collapses,which requires further investigations into the impact of this earthquake.

Rapid report of source parameters of 2023 M6.2 Jishishan,Gansu earthquake sequence

The M6.2 earthquake in Jishishan,Gansu Province,on December 18,2023,caused extraordinary earthquake disasters.It was located in the northern part of the north−south seismic zone,which is a key area for earthquake monitoring in China.The newly built dense strong motion stations in this area provide unprecedented conditions for high-precision earthquake relocation,especially the earthquake focal depth.This paper uses the newly built strong motion and traditional broadband seismic networks to relocate the source locations of the M3.0 and above aftershocks and to invert their focal mechanisms.The horizontal error of earthquake location is estimated to be 0.5−1 km,and the vertical error is 1−2 km.The focal depth range of aftershocks is 9.6−14.6 km,distributed in a 12-km-long strip with SSE direction.Aftershocks in the south are more concentrated horizontally and vertically,while aftershocks in the north are more scattered.The focal mechanisms of the main shock and aftershocks are relatively consistent,and the P-axis orientation is consistent with the regional strain direction.There is a seismic blank area of M3.0 and above,about 3−5 km between the main shock and aftershocks.It is suggested that the energy released by the main shock rupture is concentrated in this area.Based on the earthquake location and focal mechanism of the main shock,it is inferred that the Northern Lajishan fault zone is the seismogenic structure of the main shock,and the main shock did not occur on the main fault,but on a secondary fault.The initial rupture depth and centroid depth of the main shock were 12.8 and 14.0 km,respectively.The source rupture depth may not be the main reason for the severe earthquake disaster.

基于改进YOLOX-s的车辆检测方法研究

为缓解车辆小目标漏检及误检问题,提出一种基于YOLOX网络的多尺度特征融合的改进车辆检测模型。设计基于深度可分离卷积的Ghost-CSP(cross stage partial),替换网络的部分跨阶段局部结构,加快检测速度;将模型的最大池化方式改进为Softpool方式,并引入坐标注意力机制,增强待检测目标的特征表达,优化目标漏检问题;选用Focal Loss作为模型置信度损失函数以增加分类不准确样本的权重,提高模型对小目标的预测能力。实验结果表明:改进算法平均准确率提高到74.96%,速度达到73帧/s,在满足实时性要求下可以更好地完成车辆目标检测要求。

基于改进轻量化YOLOX的无人机航拍目标检测算法

针对小型无人机在巡逻航拍中的应用,提出了一种改进的轻量化目标检测算法,有效解决巡逻过程中空地无线传输信道和机载端计算能力双重受限的难题;该算法在YOLOX算法的基础上,首先利用Mobilenetv2代替CSPDarknet骨干网络作为特征提取网络,降低了模型参数量和计算量,提高目标检测实时性;其次为了弥补轻量化带来的检测精度下降,考虑检测目标框的长宽比引入CIOU定位损失函数,提升目标定位的精度;同时为了平衡训练过程中的正负难易样本,引入Focal Loss置信度损失函数提升模型的检测性能;基于VisDrone2019-DET数据集实验表明,改进后算法模型参数量降低了56.2%,计算量降低了52.5%,在检测精度没有明显下降情况下单张图片推理时间减少了41.4%;最后,将改进后的算法部署到Nvidia Jetson Xavier NX机载端,测得模型检测帧率可达22 FPS,改进后算法满足巡逻任务的应用需求。

基于深度学习的汽车轮胎受损检测方法研究

针对汽车轮胎受损尺度小、特征与背景相似导致检测困难等问题,提出一种基于改进YOLOX(YOLOX-O)的深度学习轮胎受损检测方法.以YOLOX模型为框架,对原YOLOX网络进行下采样层剪枝,使最小特征图尺度增加以实现对小尺度受损目标检测;为解决目标特征和背景相似导致受损难以检测的问题,在主干网络CSP结构中添加最大池化层和1×1卷积层,以保留最突出的目标特征;为平衡受损样本不均衡问题,将原网络中用于计算损失函数的BCE-loss函数替换为Focal-loss.利用Aidlux平台对模型进行压缩和优化,并将其部署到手机上实现移动端目标检测.试验结果表明:本文方法能够检测尺度小、特征和背景相似的受损轮胎,检测平均准确率均值达到90.8%,而且能够在手机端实现快速检测,适于推广应用.

Design and fabrication of compound varifocal lens driven by polydimethylsiloxane film elastic deformation

A compound varifocal lens based on electromagnetic drive technology is designed and fabricated, where the polydimethylsiloxane(PDMS) film acts as a driving component, while the PDMS biconvex lens and the plane-concave lens form a coaxial compound lens system. The plane-concave lens equipped with driving coils is installed directly above the PDMS lens surrounded by the annular magnet. When different currents are applied, the annular magnet moves up and down, driving the PDMS film to undergo elastic deformation, and then resulting in longitudinal movement of the PDMS lens. The position change of the PDMS lens changes the focal length of the compound lens system. To verify the feasibility and practicability of this design, a prototype of our compound lens system is fabricated in experiment. Our proposed compound lens shows that its zoom ability reaches 9.28 mm when the current ranges from -0.20 A to 0.21 A.

Constraint on the focal mechanism of the 2011 Tohoku earthquake from the radial modes

Different from other normal modes of the Earth’s free oscillation that depend on all the six components(M_(rr),M_(tt),M_(pp),M_(rt),M_(rp),and M_(tp))of the centroid moment tensor,the amplitudes of the radial modes depend on the M_(rr)component(e.g.,scalar moment(M_(0)),dip(δ),and slip(λ))and hypocenter depth of the focal mechanism,and hence can be easily used to constrain these parameters of the focal mechanism.In this study,we use the superconducting gravimeter(SG)records after the 2011 Tohoku earthquake to analyze the radial modes_(0)S_(0)and_(1)S_(0).Based on the solutions of the focal mechanism provided by the GCMT and USGS,we can obtain the theoretical amplitudes of these two radial modes.Comparing the theoretical amplitudes with the observation amplitudes,it is found that there are obvious differences between the former and the latter,which means that the GCMT and USGS focal mechanisms cannot well represent the real focal mechanism of the 2011 event.Taking the GCMT solution as a reference and changing the depth and the three parameters of the M_(rr)moment,the scalar moment(M_(0))and the dip(δ)have significant influences,but the effects of the slip(λ)and the depth are minor.After comparisons,we provide a new constraint(M_(0)=5.8±0.09×10^(22)N·m,δ=10.1±0.08°,λ=88°,and depth=20 km)for the focal mechanism of the 2011 event.In addition,we further determine the center frequency(1.631567±2.6e^(-6)mHz)and quality factor(2046.4±50.1)of the_(1)S_(0)mode.

ProNet Adaptive Retinal Vessel Segmentation Algorithm Based on Improved UperNet Network

This paper proposes a new network structure,namely the ProNet network.Retinal medical image segmentation can help clinical diagnosis of related eye diseases and is essential for subsequent rational treatment.The baseline model of the ProNet network is UperNet(Unified perceptual parsing Network),and the backbone network is ConvNext(Convolutional Network).A network structure based on depth-separable convolution and 1×1 convolution is used,which has good performance and robustness.We further optimise ProNet mainly in two aspects.One is data enhancement using increased noise and slight angle rotation,which can significantly increase the diversity of data and help the model better learn the patterns and features of the data and improve the model’s performance.Meanwhile,it can effectively expand the training data set,reduce the influence of noise and abnormal data in the data set on the model,and improve the accuracy and reliability of the model.Another is the loss function aspect,and we finally use the focal loss function.The focal loss function is well suited for complex tasks such as object detection.The function will penalise the loss carried by samples that the model misclassifies,thus enabling better training of the model to avoid these errors while solving the category imbalance problem as a way to improve image segmentation density and segmentation accuracy.From the experimental results,the evaluation metrics mIoU(mean Intersection over Union)enhanced by 4.47%,and mDice enhanced by 2.92% compared to the baseline network.Better generalization effects and more accurate image segmentation are achieved.

A monolithic integrated medium wave Mercury Cadmium Telluride polarimetric focal plane array

A medium wave(MW)640×512(25μm)Mercury Cadmium Telluride(HgCdTe)polarimetric focal plane array(FPA)was demonstrated.The micro-polarizer array(MPA)has been carefully designed in terms of line grating structure optimization and crosstalk suppression.A monolithic fabrication process with low damage was explored,which was verified to be compatible well with HgCdTe devices.After monolithic integration of MPA,NETD<9.5 mK was still maintained.Furthermore,to figure out the underlying mechanism that dominat⁃ed the extinction ratio(ER),specialized MPA layouts were designed,and the crosstalk was experimentally vali⁃dated as the major source that impacted ER.By expanding opaque regions at pixel edges to 4μm,crosstalk rates from adjacent pixels could be effectively reduced to approximately 2%,and promising ERs ranging from 17.32 to 27.41 were implemented.

IR-YOLO: Real-Time Infrared Vehicle and Pedestrian Detection

Road traffic safety can decrease when drivers drive in a low-visibility environment.The application of visual perception technology to detect vehicles and pedestrians in infrared images proves to be an effective means of reducing the risk of accidents.To tackle the challenges posed by the low recognition accuracy and the substan-tial computational burden associated with current infrared pedestrian-vehicle detection methods,an infrared pedestrian-vehicle detection method A proposal is presented,based on an enhanced version of You Only Look Once version 5(YOLOv5).First,A head specifically designed for detecting small targets has been integrated into the model to make full use of shallow feature information to enhance the accuracy in detecting small targets.Second,the Focal Generalized Intersection over Union(GIoU)is employed as an alternative to the original loss function to address issues related to target overlap and category imbalance.Third,the distribution shift convolution optimization feature extraction operator is used to alleviate the computational burden of the model without significantly compromising detection accuracy.The test results of the improved algorithm show that its average accuracy(mAP)reaches 90.1%.Specifically,the Giga Floating Point Operations Per second(GFLOPs)of the improved algorithm is only 9.1.In contrast,the improved algorithms outperformed the other algorithms on similar GFLOPs,such as YOLOv6n(11.9),YOLOv8n(8.7),YOLOv7t(13.2)and YOLOv5s(16.0).The mAPs that are 4.4%,3%,3.5%,and 1.7%greater than those of these algorithms show that the improved algorithm achieves higher accuracy in target detection tasks under similar computational resource overhead.On the other hand,compared with other algorithms such as YOLOv8l(91.1%),YOLOv6l(89.5%),YOLOv7(90.8%),and YOLOv3(90.1%),the improved algorithm needs only 5.5%,2.3%,8.6%,and 2.3%,respectively,of the GFLOPs.The improved algorithm has shown significant advancements in balancing accuracy and computational efficiency,making it promising for practical use in resource-limited scenarios.