面向目标检测的数据集标签遗漏的协同修正技术

针对图像标注中疲劳、粗心等因素引起的标签遗漏现象,使得模型训练时难以正确区分正负样本,进而影响模型性能。设计了一种协同修正技术,通过多次迭代更新训练集,将潜在无标签的目标进行对象擦除,降低训练集的错误监督信息,避免人工的重复检查和重复标注。该方法无需进行算法参数调整、不依赖具体网络结构,低成本地减少数据集错误从而提升模型训练精度。在YOLOv5算法的实验基础上表明协同修正操作仅迭代1次即有明显的改善效果,并在多个公共数据集上能够提升0.4%~1.4%的检测精度,当数据集中的标签遗漏率达到40%时依然能够生效。该方法对数据集中样本的数据量和类别数没有限制,可应用于电商、遥感、通用等多种目标检测场景,保持着较好的鲁棒性和泛化性。

基于神经辐射场算法的混合现实三维重建技术

针对目前在混合现实(MR)环境中高效率建立高质量三维(3D)模型的需求,基于神经辐射场算法(NeRF)的三维重建技术,提出了一种基于Laplacian算子的数据集优化算法。首先,围绕某线切割设备录制了一段1 min 51 s的视频,并采取等距提取视频帧的方式,获取了训练数据集;然后,使用Laplacian算子对数据集进行了优化,同时保留了原始数据集作为对比,使用了基于NeRF算法的重建方式与传统的基于COLMAP的稠密点云重建方式,分别对两组数据集进行了三维重建;最后,在重建精度与重建速度方面,对不同重建方式、不同重建数据集的重建结果进行了比较。研究结果表明:COLMAP稠密点云重建耗时是基于NeRF重建耗时的9.98倍,而相较于COLMAP稠密点云重建,使用NeRF重建方式的模型表面缺陷较少;此外,使用Laplacian算子优化的数据集的NeRF重建在峰值信噪比(PSNR)和结构相似性(SSIM)指标上分别提升了2.43%、0.72%,有利于提升重建模型的质量。研究结果支持混合现实技术在制造业数字化转型中的应用,可为其提供有益的参考。

一种改进的KNN网页分类算法

针对KNN算法懒惰分类和效率不高的特点,将训练数据集进行优化,提取有代表性的训练样本作为代表样本,用其代替整个训练集进行相似度比较。实验结果表明,使用代表样本集的分类性能与传统KNN算法的性能相当,缩短了分类时间,提高了分类效率,并且不需要估计K值,减少了人工估计值的偏差。

基于YOLO v5的变电站作业人员着装规范性识别

为了对变电站现场作业人员工作服着装规范性进行有效识别,防止作业人员穿短袖或将长袖的袖子卷起,本文采用了YOLO v5实时目标检测算法,针对长袖/短袖两类检测问题,修改分类器,将输出层修改为21维张量。对变电站实际工作现场采集到的7135张样本进行标注,对数据集进行图像增广和色彩调整,并对初始化锚框进行聚类分析。基于YOLO v5在COCO上的预训练模型,利用难例挖掘等手段调整训练目标,最终得到最优的监控场景下工作人员着装规范性的检测模型。实验结果表明,在1000张图片测试集上查准率为97.01%,查全率为96.60%,mAP值为0.8767。在实验环境下平均响应速度达到了38ms。最后,将模型封装为微服务,部署于变电站辅助系统。该着装规范性识别模型速度快、误报少,满足了安全监察需求。

基于OGSA网格技术的连锁超市进销存系统研究

本研究采用网格计算技术使进销存系统从简单的计算资源集中上升到数据共享层面,并且实现了协作处理和网络节点的高质量的服务,能极大的提高连锁超市间计算资源的效率和利用率,有效破解因计算能力差、数据来源单一或存储资源欠缺而无法解决的问题。

Development of a Data Acquisition System to Optimizing the Agricultural Tractor Performance

A novel data acquisition system was successfully integrated on-board the Massey Ferguson 3,000 series agricultural tractors for measuring tractor-implement performance. A commercial load cell was incorporated into the existing system for the needed tractor-implement performance, measurements. This system is capable of measuring, displaying and recording, in real-time, the tractor＇s theoretical travel speed, the actual travel speed, the engine speed, the fuel consumption rate, and the drawbar pull. Static calibration tests on various associated sensors for the required measurements show excellent linearity with correlation coefficients that are close to 1. The developed system was extensively and successfully field demonstrated for tractor-implement performance with offset disc harrows on dry, undisturbed loamy soils. Under these conditions, a ratio of tractor power to implement width is suggested. The data also show the existence of a linear relationship between fuel consumption per hectare and specific draught, for the 4 to 9 kN m^-1 range, which suggests the possibility of extending the American Society of Agricultural and Biological Engineers model of draught prediction to forecast fuel consumption. The configuration of the tractor-harrow combination, based on the measurement of the draught required under operational conditions, provides the manufacturers with a reliable indication of the recommended power required for each harrow model. With this type of information farmers can make decisions regarding selection of a suitable tractor-implement combination for their farms. As a consequence, there is improved tractor-harrow productivity and field efficiency.

The Application of Computer Simulation Technology in Building Design

From the simulation software selection, meteorological data collection, model, boundary conditions, calculation setting and evaluation indicators, the paper summarize simulation method of wind, light and sound, summarize three kinds of environmental. The paper study on optimization analysis method of green building from four aspects of building orientation, architectural form, architectural layout and indoor environment, and puts forward the optimization scheme of the typical simulation analysis of simple questions; provide guidance for optimizing the design of green building. Finally, ECOTECT can adapt to the current domestic engineering design and consulting on green building simulation and design requirements, and has a broad application prospect.

Model-free feature screening for high-dimensional survival data

With the rapid-growth-in-size scientific data in various disciplines, feature screening plays an important role to reduce the high-dimensionality to a moderate scale in many scientific fields. In this paper, we introduce a unified and robust model-free feature screening approach for high-dimensional survival data with censoring, which has several advantages: it is a model-free approach under a general model framework, and hence avoids the complication to specify an actual model form with huge number of candidate variables; under mild conditions without requiring the existence of any moment of the response, it enjoys the ranking consistency and sure screening properties in ultra-high dimension. In particular, we impose a conditional independence assumption of the response and the censoring variable given each covariate, instead of assuming the censoring variable is independent of the response and the covariates. Moreover, we also propose a more robust variant to the new procedure, which possesses desirable theoretical properties without any finite moment condition of the predictors and the response. The computation of the newly proposed methods does not require any complicated numerical optimization and it is fast and easy to implement. Extensive numerical studies demonstrate that the proposed methods perform competitively for various configurations. Application is illustrated with an analysis of a genetic data set.