为了检测供水管道是否出现泄漏,同时针对人工检测精度低、背景噪音难以去除和现有检测设备的实用性低等问题,研究了一种基于卷积神经网络(convolutional neural network,CNN)的供水管道泄漏声音识别方法.首先设计水下机器人的供水管道内部声音实时采集系统,并利用用户数据报协议(user datagram protocol/internet protocol,UDP/IP)通信技术将该系统采集的声音信息上传至上位机,并对供水管道内的声音进行泄漏和不泄漏的划分且制作成数据集,提取泄漏音频和不泄漏音频样本的梅尔谱特征图,根据实时性选用轻量级卷积神经网络ShuffleNet V2进行训练和识别;其次引入卷积注意力模块(convolution block attention module,CBAM)到网络模型中,并对 ShuffleNet V2 的 Unit1 单元进行改进,提出了 Unit1_y 单元;最后将改进后的网络与MobileNet V3、ResNet18等轻量级网络进行对比.试验结果表明,改进后的网络模型相较于其他的模型对供水管道漏泄声音识别效果最佳且参数量低,占用上位机运算资源少,测试集识别率达到92.14％,验证了算法的有效性.

为了解决光时域反射仪(optical time domain reflectometer,OTDR)中背向散射信号受噪声干扰严重问题,本文提出了一种基于自适应噪声完备集合经验模态分解(complete ensemble empirical mode decomposition with adaptive noise,CEEMDAN)和改进小波阈值的OTDR信号去噪算法,利用CEEMDAN分解算法具有的抗模态混叠现象和降低重构误差等优点,将信号分解为若干IMF分量,根据相关系数的分析方法,找到噪声占主导的本征模态函数(intrinsic mode function,IMF)分量和信号占主导的IMF分量的临界点,去除噪声占主导的IMF分量,并将改进的小波阈值去噪方法对信号占主导的IMF分量进行去噪,最后重构信号.结果表明,本文提出的方法与传统的硬阈值方法、CEEMDAN-硬阈值方法和改进的小波阈值方法相比,能更好地抑制噪声,并达到更好的去噪效果,突显OTDR事件特征,更易于事件的检测.

局部凸连接生长算法(LCCP)存在超体素跨越物体边界,未能利用区域隐含凹凸信息的缺陷,为了改进以上缺陷导致的分割精确度低、物体粘连的问题,提出了结合连通域分割的改进算法.首先采用深度自适应超像素分割法(DASP)根据深度信息和法向量角度将图片划分为超像素;其次根据超像素的法向量夹角判定邻接超像素的凹凸性,合并所有凸连接超像素形成初步结果;最后使用基于超像素的距离变换以及分水岭生长分割方法,把面积较大的凹连通域,快速分割成多个凸区域.在IC-BIN数据集进行分割验证,结果表明平均分割精度(AP)相比于LCCP和约束平面切割法(CPC)分别提升25％和35％,显著改善了欠分割问题.

在进行皮革裁切、板材边角料再利用、PCB板布线等过程时都会遇到求解不规则图型最大内接矩形的问题;当前工厂的裁定手段普遍由人工估测来完成,这种作业方式不仅主观性强,而且效率低下;在研究了基于数字图像处理的遍历法、中心扩散法、遍历中心扩散法后,结合以上方法提出一种高效实用的最大内接矩形快速检测算法:边界排序生长法;实验结果证实:遍历法无法正确适用于凹多边形;中心扩散法虽然可适用于任意不规则图型,但其检测效果不够理想;遍历中心扩散法的检测结果较为理想,但该方法耗时巨大;而边界排序生长法在保证高效率的同时依旧可以拟合出理想的内接矩形.

半导体脉冲激光器的结构、材料、加工工艺和驱动电源特性是其性能的重要影响因素,结合脉冲宽度、输出功率和重复频率等技术指标,综述了半导体脉冲激光器的发展与研究现状,着重介绍了窄脉冲叠加直流偏置法、储能元件的应用、高速开关的级联或阵列、可编程逻辑器件的应用以及器件选型和布局创新这5种驱动电源性能提升方式,并指出目前技术存在的不足.最后,对半导体脉冲激光器的发展趋势进行了展望.

针对传统水稻生产中存在的问题,提出了一种新的水稻种植方法,在工厂将优选的水稻种子按照实际种植间距播种到专用的纸膜上,形成水稻种子膜,水稻种植窗口期利用专用的铺膜机将水稻种子膜铺设到田间.该技术具有防草、有效利用秸秆、设备利用率高、改善田间通风环境、节省种子、用高精度机械设备实现精准播种等特点.围绕这一新方法,对生产种子膜的设备进行了研究,完成了送膜机构、供土机构、点种机构、喷胶机构和覆膜机构的原理性设计.原理样机的实验室研究和种子膜的田间试种,证实了水稻种子膜技术的可行性.

针对子弹的表面缺陷检测使用常规的图像处理效果并不理想,文章提出使用改进的卷积神经网络和Faster RCNN网络相结合的方法,首先采集子弹不同类型的缺陷图像作为数据集,然后使用改进后的VGG16和ResNet50两种卷积神经网络分别处理数据集图像,再经过RPN网络和Fast RCNN网络的训练,对图像进行分类与缺陷特征提取,获取比较准确的缺陷区域的框图.实验结果表明,使用改进后的卷积神经网络与Faster RCNN网络相结合的方法可以提高子弹缺陷检测的准确率、召回率和mAP,准确率、召回率和mAP均可达到90％.

An active laser triangulation method based on the telecentric lens is proposed to accurately measure the thickness of the object. The parallel projection method applies to the telecentric lens, and thus the relationship between the imago point displacement and the object height change is linear. We establish a model of active laser triangulation based on the telecentric lens, and discusses the factors that influence the accuracy, including optical system error, machine error, and image processing error. In the experiment, we get the moan error less than 5 mu m for thickness measurement when combing the active laser triangulation based on the telecentric lens and the compensation interpolation method.

针对由多种降质因素导致图像增强效果不理想的情况,提出一种参数化的伽马校正算法实现图像的增强,即整幅图像采用变化的伽马值进行幂律变换.图像处理时,将图像中的每个像素点与平均亮度比较划分为亮点、暗点和平滑点,将像素分类有关的非线性变换函数作为传统伽马校正中幂律值计算的参数,获得参数化的伽马校正增强函数.实验结果表明,对过曝光、曝光不足以及混合多种性质的复杂图像,该方法均有较好的增强效果,鲁棒性较好,在保持输入图像亮度的情况下,能够自适应处理各种质量较低的图像.

For the extraction of minutiae feature and image feature in fingerprint texture,the commonly used methods of texture extraction such as statistical features of LBP texture,texture feature extraction based on fractal dimension,Gabor filter texture extraction and Tamura texture analysis are researched.According to the characteristics of optimal localization in frequency domain and time domain simultane-ously for Gabor filter,the local structure information about spatial frequency(scale),spatial location and direction selectivity can be well described.For this,we propose introducing two different curvature factors on the basis of traditional Gabor function model,and solve the problem that the general Gabor filter cannot be used to extract the local bending features by the characteristics of general Gabor filter de-composed into one-dimensional Gauss filter in two orthogonal directions.After theoretical analysis and practical research,it is indicated that the improved Gabor filter is better than the traditional Gabor filter for the extraction of the fingerprint texture.

针对手机屏幕到外壳间间隙、听筒宽度、装饰边宽度等尺寸参数的实时在线测量需求,提出了一种基于机器视觉的手机间隙尺寸测量方法,通过对标准图像设定感兴趣区域,提取该区域特征后,在待检测手机中自动匹配感兴趣区域,随后使用Canny算子寻找边缘点,再用与边缘垂直的一组平行线求出参与最小二乘法拟合目标直线的边缘点,在减少计算量的同时,提高了测量直线边缘的定位精度.在VS2012平台上,结合运动控制,开发了一套在线手机间隙尺寸测量系统,该系统具有实时、高效、高精度、重复性好等特点,且其精度达到10μm,具有很好的应用价值和前景.

针对传统天地盖纸盒成型机中光电跟踪系统定位精度低、误差大的问题,提出一种基于机器视觉的纸盒包装定位方法,设计并构建了一个视觉定位系统.分析了纸张定位要求与图像特征,采用了以一种以区域中心为特征点的匹配方法确定感兴趣区域,然后采用中值滤波消除图像噪声,运用Canny算子提取直线边缘、总体最小二乘法检测直线,经过几何运算计算面纸位置及旋转角度,实现了面纸位置的实时、准确、非接触式测量.在VS2012平台下调用OpenCV开源函数库实现定位算法和检测程序,将程序测量结果传至PLC控制系统,实现了纸盒的精准包装.

针对当前国内街机游戏开发难度大、难以移植、开发周期长的问题,提出Unity3D引擎结合Android平台技术的街机开发方案.该项目用Unity3D引擎开发一款游戏,完成Android系统与Unity3D的数据交互、添加NGUI事件源,从而实现外部硬件设备控制游戏及游戏中的交互界面.针对厂家生产需求开发一个安卓系统桌面启动器并监听Android开机广播实现开机启动游戏,预装游戏软件实现系统集成.实践表明该系统完成度高,该开发方案开发效率高,在街机游戏开发和应用方面有一定的推广应用价值.

尽管实时排名机制在竞速游戏中的应用已非常普遍，但迄今为止，国内文献中仍没有相关的文字描述和理论研究。因此，本文结合Unity3D引擎的优势和游戏开发的需求，基于样条曲线原理和Unity3D的碰撞检测机制提出了一种新的竞速游戏中实时排名算法的实现思路。游戏测试结果表明，该算法针对不同类型的赛道和地形都表现出良好的稳定性和准确性，具有在竞速类游戏开发中的普遍推广价值。

Establishing category keywords is the key problem in text classification, which should be solved first.On the basis of the classification of text by using the category keywords and TF-IDF algo-rithm, an improved TF-IDF algorithm has been proposed to overcome the shortcomings of the vector space model, which cannot well adjust the weights.Firstly, category keyword library should be established, and the expansion and duplication be carried out.The weight of keywords in the document is modified by the addition of the length of the document, and the shortage of the original features of the entry class dis-tinction ability is solved effectively.By using Bayesian classification method, combined with the experi-ments, the effectiveness of the algorithm is verified, and the accuracy of text classification improved.

Due to the rust on metal surface,it is easily to make the line trace images affected by noise,and causes the problem of the image feature extraction,comparison and analysis.Traditional denosing method like Gauss filter causes damage and excurtion to the edge while mean filter cannot distinguish between edge and background effectively. Therefore a new denoising method is proposed in this paper.The diffusion model of this partial different equation filter algorithm is PM equation.According to the textural feature of striation mark image,the different weight functions in different diffusion directions are introduced.However,the diffusion threshold is set up via gray level histogram of image in this method.The method can eliminate image noise without destroying image edges simultaneously.Experimental results show that the algorithm is superior to the PM equation and LinShi operator and it has good application value in line trace image processing.

Considering the defects existing in the production process,an on-line detection method based on vision is proposed. Through the image collection system of the bottle cap detecting device to acquire the image,then the adaptive threshold is used to extract region of interest. Using geometric properties of circle to find the center of bottle cap region,then image difference method is used to distinguish burr defects. In addition,H and S channel reflects the color difference in HSV color space,chromatic aberration expression with weighing can be used to check out color difference,color difference will be highlight when k value is set greater than 1. The system has been applied in actual production,which indicates that the online detection technique be competent at the defect detection in high speed production of bottle cap.

本文针对传统工具痕迹检验鉴定依赖于人工比对的现状,提出了线条型工具痕迹图像计算机比对的方法,通过提取划痕线条并直线拟合,沿线条方向按照一定间隔获取线条垂直方向的位置参数作为比较特征,判断两个痕迹特征的位置确定痕迹间的关系,实现线条型工具痕迹的自动比对.实验结果表明,该方法能有效地提取划痕线条,实现线条痕迹的自动匹配,有助于后续工具痕迹的自动检测.

本文研究了一种智能节能插座,通过对负载的待机功耗检测,利用低功耗微处理器控制继电器切断连接负载的电路,实现对负载设备消除待机能耗的目的,以此达到节能和环保功效,同时也起到了保护电器的作用.