重庆方言语音识别系统的设计与实现

语音识别赋予了计算机能够识别出语音内容的功能,是人机交互技术领域的重要研究内容。随着计算机技术的发展,语音识别已经得到了成熟的发展。但是关于方言的语音识别还有很大的发展空间。中国是一个幅员辽阔、人口众多的国家,因此方言种类繁多,其中有3000多万人交流使用的重庆方言就是其中之一。采集了重庆方言的部分词语的文本文件和对应的语音文件建立语料库,根据重庆方言的发音特点,选取重庆方言的声韵母作为声学建模基元,选取隐马尔可夫模型(Hidden Markov Model, HMM)为声学模型设计了一个基于HMM的重庆方言语音识别系统。在训练过程利用语料库中训练集语料对声学模型进行训练,形成HMM模型库;在识别过程利用语料库中的测试集语料进行识别测试。实验结果表明,该系统能够实现重庆方言的语音识别,并且识别的正确率为100%。     

基于红外/毫米波双模融合的目标识别方法

在数据融合的基础上,以红外/毫米波双模传感器的智能融合结构为模型,将模糊神经网络与D-S证据理论相结合,提出了一种新的目标识别方法.该算法根据红外/毫米波传感器的性能及工作范围,构造模糊变量作为神经网络的输入,根据神经网络的不同输出判别目标的真伪,并利用D-S证据理论进行目标身份识别.仿真结果证明了该算法的可行性.     

基于小波分解系数的贝叶斯人脸识别方法

本文给出了贝叶斯人脸识别方法中匹配准则的多个近似表达式及一种实用的快速计算方法,在此基础上,利用反对称双正交小波变换的微分算子功能,提出了一种利用两幅人脸图像的小波变换系数差作为模式矢量的贝叶斯人脸识别方法,并利用AR人脸图象库进行了实验,实验结果表明本文方法与基于图像灰度的类似方法相比,识别率提高8％左右,此外本文方法也提供了一条在图像压缩数据域中实现人脸识别的可能途径。     

基于DIGNET网络的数据融合方法

针对数据融合和目标识别的特点,提出了基于DIGNET自组织聚类人工神经网络的数据融合方法。考虑到多传感器系统测量多个参量的特点,用并行的子网络结构代替中间隐层,实现了基于决策层的信息融合目标识别。利用仿真数据对基于DIGNET的数据融合方法进行了实验研究。实验结果表明,该方法具有数据正确分类率高和抗噪能力强等优点,有效地实现了融合识别。将该方法应用于前视红外和可见光双传感器目标跟踪系统的数据融合识别是可行的。     

中高分辨力遥感图像中飞机目标自动识别算法研究

提出了一种中高分辨力的航空航天遥感图像中飞机目标快速自动识别的新算法。在分割和分类过程中充分利用飞机目标的先验知识,提出了一种改进区域分割方法,并应用树分类器对飞机目标进行自动识别。所提出的改进区域分割方法较好地实现了区域分割中阈值的准确自动选取,克服了复杂背景图像中小目标的全局阈值自动分割的失效问题。采用二叉树分类器,通过提取简单的目标几何特征,分层进行种类识别,提高了识别速度,降低了漏检率和虚警率。运用该方法进行了实验。结果表明,识别率达到了100%。     

Wavelet and harmonic filtering for distortion-invariant volume holographic correlation

The wavelet and harmonic filtering method suggested by Zalevsky and Ouzieli is introduced in this paper and adopted in our volume holographic image recognition system. This composite filter combines several scaled versions of the cascaded wavelet and harmonic filter, obtaining high discrimination ability and wide dynamic range of rotation and scale deformations. Optical experiments are conducted to demonstrate the validity and practicability of the algorithm. To the best of our knowledge, this is the first report of using this algorithm in a volume holographic system. Moreover, the separate correlation approach proposed in this paper greatly simplifies the manufacturing process and reduces the cost of the system.     

BP神经网络在光学相关器相关峰识别中的应用

光学相关识别是图像识别的重要方法，有效识别相关器输出平面的相关峰信号是保证光学相关器图像识别准确性的关键。由于激光器输出功率的波动、光学系统本身的误差以及SLM器件本身带来的噪声，采用一般的阈值方法很难达到理想的效果。该文提出对相关器的输出平面进行预处理，充分考虑相关信号的形状信息，提取感兴趣区域(ROI)，采用BP神经网络对输入矢量进行计算，可达到对相关峰信号和噪声的有效分类识别，从而提高了光学相关器识别的可靠性，降低了误判的概率。     

Real time estimation of CO2 laser weld quality for automotive industry

Laser welding is one of the most precise welding processes in joining sheet metals. In laser welding, performing real time evaluation of the welding quality is very important to enhance the efficiency of the welding process. In this study, the plasma and spatter, which are generated during laser welding, are measured using UV and IR photodiodes. The factors that influence weld quality are classified into five categories; optimal heat input, slightly low heat input, low heat input, partial joining due to gap mismatch, and nozzle deviation. The data number deviated from reference signals and their standard deviations were also considered to evaluate the qualities. A system was also formulated to perform real time evaluations of the weld quality using a fuzzy multi-feature pattern recognition with the measured signals.     

改进形态学相关算法以识别高相似度灰度图像

从形态学角度定义了灰度联合图像相似度，提出两种基于位表示法形态学相关算法的改进算法，通过提取位表示法的图像片边缘特征或二元化位表示法的图像片功率谱，以提高位表示法形态学相关算法对高相似度灰度图像的识别能力。     

An eye behavior measuring device for VR system

This work presents an eye-tracking and pupil size-measuring device that interfaces with a computer for applications useful in psychometry, ophthalmology, physiology and virtual reality (VR) systems. This system utilizes a change-coupled device (CCD) camera, appropriate lenses, PC with frame grabber and a DSP unit with various types of VR equipment, i.e., HMD, simulator or LCD projection device. The digital signal processing unit is used to calculate the average brightness and contrast of the VR video image. A CCD camera with various attachments can be mounted on various VR systems to capture the human eye image for testing. An image capture card and a personal computer are used to analyze the test image. From the eye digital image, the computer obtains data on the pupil size and a trace of the tested eye. A pattern recognition computer program and five measurement parameters are used to distinguish the position of the pupil, calculate the pupil location coordinate and analyze the physical conditions of the user. These data can be plotted against the average brightness and contrast of the VR video image in real time. This information is shown on the screen of a personal computer and used for cross-link analysis. This eye-tracking interface can determine the position of a subject's pupil and map that position into a display point on a computer screen. The pupil size and location data versus the average brightness and contrast of a VR video image are computed in real time.