语音信号回声隐写后其倒谱系数会在回声延迟出产生峰值,传统回声隐写分析主要采用倒谱系数的统计特征作为隐写检测特征,然而在低回声幅度时隐写信号倒谱系数的峰值并不明显,基于统计特征的方法检测性能并不理想.本文将倒谱分析与图像识别技术结合,提出了一种基于倒谱图像的语音回声隐写分析方法,对语音信号分帧加窗后进行倒谱计算,然后以时间为横轴,倒谱序列点为纵轴,倒谱系数幅值为灰度级生成倒谱图像,将生成的倒谱图像作为隐写检测的输入,采用残差神经网络作为分类器进行回声隐写分析.实验结果表明,在3种经典回声隐写算法上低回声幅度时检测准确率分别达到98.2%、98.6%和96.1%,本文方法在低回声幅度时检测准确率相较传统回声隐写分析方法有较大提升,解决了传统回声隐写分析方法在低回声幅度检测效果不佳的问题.

针对语音在高压缩比及低信噪比下传输与重构质量不佳的问题,提出一种基于语谱图的语音压缩传输重构方法.在发送端将语音信号转为语谱图进行传输,再在接收端对语谱图作图像去噪处理,根据去噪后的图像恢复出语音信号的幅度谱;建立发声重构模型,用幅度谱对语音信号进行重构,实现语音恢复.实验结果表明:无噪声环境下,压缩比为10和40的条件下,重构语音质量客观平均得分达到3分以上;低信噪比条件下,压缩比为10时,重构语音质量客观平均得分也能达到2分以上.相比于传统的压缩感知语音重构算法,在高压缩比下,新方法对重构语音质量有明显改善.

为了提高低信噪比下语种识别的准确率,引入一种新的特征提取融合方法.?在前端加入有声段检测,并基于人耳听觉感知模型提取伽玛通频率倒谱系数(Gammatone?Frequency?Cepstrum?Coefficient,GFCC)特征,通过主成分分析对特征进行压缩、降噪,融合每个有声段的Teager能量算子倒谱参数,通过高斯混合通用背景模型进行语种识别验证.?实验结果表明,在信噪比为?5～0?dB时,相对于基于对数梅尔尺度滤波器组能量特征方法,融合特征集方法对5种语言的识别率,分别提升了23.7％～34.0％,其他信噪比等级下识别率也有明显的提升.

语种识别受真实噪声环境的影响较大,识别效果不佳.为了解决真实噪声环境下语种识别的问题,提出一种基于对数灰度语谱图的图像处理语种识别方法.根据噪声能量和语音能量在语谱图上的分布规律对真实噪声中的语音信号进行带通滤波;再结合人耳听觉特性提取对数灰度语谱图;然后提取图像主成分特征作为语种特征,采用残差神经网络模型进行训练测试.实验结果表明,在掠夺者战斗机驾驶舱的环境下,所提方法的平均识别正确率相对于线性灰度语谱图方法提升了27.5％,在其他噪声环境下的平均识别正确率也有提升.

针对低信噪比下语种识别正确率低的问题,提出了一种声道冲激响应频谱参数和Teager能量算子倒谱参数融合的识别方法.根据语音中不同特征信息量分布特性,首先在特征提取前端引入低通滤波器滤除信号高频部分,并采用重采样方法降低采样率,再基于信号频谱提取声道冲激响应频谱参数,然后融合Teager能量算子倒谱参数,最后通过高斯混合通用背景模型进行语种识别验证.不同信噪比条件下性能测试表明,所提方法相对于基于单一的梅尔频率倒谱系数特征、单一的伽玛通频率倒谱系数特征和基于对数梅尔尺度滤波器组能量特征,在低信噪比下提升约15 dB,显著提高了识别正确率.

针对现有的正交频分复用(Orthogonal Frequency Division Multiplexing,OFDM)频偏估计算法普遍存在估计范围小,估计精度不高的问题,提出了一种新的训练序列频率同步算法.首先采用缩短训练序列和周期性发送序列的方式,增大了频偏估计的范围,但频偏范围的增大会导致性能的损失;然后又提出一种通过对短周期重复样式的估计值取平均的方法,在保持估计范围不变的情况下,进一步的提高了频偏估计的性能.最后仿真结果表明,改进的算法频偏估计范围大,并且估计精度较高,均方误差(Mean Square Error, MSE)可以达到10-6.

目前对球员综合能力的评价方法有TOPSIS法、灰色关联分析法等,但当评价指标量多时,存在计算效率低等缺陷,同时有些评价指标并不是值越大越好,且会对最优排序和最劣排序造成影响.为求取最适合的算法、作出准确的NBA球员综合能力评价,采用主成分分析法,以现役NBA联盟中538名运动员的得分、助攻、三分命中率等13项指标为实例,进行评价方法研究.分析球员多方面能力,得出球员在各项成分中得分排名并计算出各项能力最强的前十名球员,与体育界分析结果进行比较,分析球员的强项和弱项.实验结果表明,主成分分析能够高效地将数据降维,表现各项数据之间关联性,并且分析结果正确,适用于NBA球员综合能力评价.

现在的嵌入式系统在信号采集与网络上传方面,常采用RS485等通信方式,通信速率低、稳定性差,控制与数据传输过程复杂.而类似ENC28J60等转换模块编程复杂控制麻烦,而且极易被攻击篡改.基于广播监播机对上位机数据传输的实际要求,设计了一种STC15单片机+W5500网络芯片集成模块的串口-网口转换通信的设计方案,并阐述了设计方法.与以往的网络方案不同的是,W5500是一款硬件TCP/IP协议芯片,提供了SPI作为外设主机接口,设置W5500内部的相关寄存器来实现具体的网络功能.实验证明,在广播监播机上采用该方案具有通信稳定,支持多客户端访问,设计使用简单清晰等特点.

本文提出一种基于STC15W4K48S4单片机的高精度数字频率计的设计方法,内部软件设计采用多周期同步测量法实现,设计中对测量的数据进行相应的调整减少误差.由于采用了32 MHz的晶振,测量范围可在1 Hz~10 MHz,并且在高频下误差相对很小.本次设计给出的频率计的设计方案,不但切实可行,而且设计简单、成本低、可测频带宽,大大降低了设计成本和实现复杂度.

In order to increase the license plate recognition rate and recognition speed in complex en-vironments, a phased license plate recognition algorithm based on BP neural network and convolution neural network( CNN) is proposed. This method BP neural network used to recognize license plate charac-ters, non-similar letters and numbers in the first stage;and improved CNN used to identify similar license plate letters and numbers in the second stage. Finally through the experimental results of vertical and hor-izontal comparison, the advantage of this method is obtained. Experimental results show that compared with other algorithms such as BP neural network, this method has improved the recognition rate while the recognition time is reduced.

A recognition algorithm for similar characters on license plates based on convolution neural networks (CNN) is proposed in this study to improve the recognition rate in a complex environment. The algorithm adopts the improved CNN to recognize similar letters and numbers in license plate images. Experimental results suggest that the improved CNN may improve the recognition rate and speed of similar characters on license plates.

Super resolution image reconstruction is a new technology which means to use multiple video sequences,or single-frame image and the training sample images of complementary information between the images to reconstruct a better quality,higher spatial resolution image data,make up the original image data is the lack of spatial resolution,improved image spatial resolution for force and clarity.Describes the method based on regularization of the super-resolution image reconstruction.On this basis,using the L1 norm of the reconstructed image fidelity constraint,the use of total variation regularization to overcome the ill-conditioned reconstruction problems,effectively maintain the edge of the image.To achieve a text message containing the image of regularized super-resolution reconstruction,experimental verification of the effectiveness of the method.

Remote sensing technology arised in the 1960s and developed into a comprehensive detection technology,while remote sensing technology has been used in many fields widely.the face massive remote sensing image data,the quality and efficiency of interpretation is important.Based on summarizing the main difficulties and deficiencies of remote sensing information extraction technology in the past,the paper focuses on urban features of remote sensing information extraction technology ideas about object-oriented,the finally,summarfized and analyzed this idea,pointing out that the current problems and future research directions.

An averaging filtering algorithm was proposed based on the analysis of classical median and aneraging filter algorithm,and the principles and implementing steps were given.Comparison experiments between it and the classical averaging filter algorithm.The experimental results indicated the excellent filtering performance and accelerating calculating speed from subjective impression and objective parameters.

The Gabor motion energy filter(GME) is applied to dynamic facial expressions.Finally,comparative experiments are carried out in the Cohn-Kanade face database.They show that GME outperforms GE on low intensity expression discrimination.

An efficient algorithm was presented for the statistics of traffic flows including that illumination was not sufficient at night.Firstly,In order to separate the foreground from background,the original video sequences had been thresholded,and by using morphological erosion to eliminate some isolated noises and smaller regions,these regions whose ranges are smaller than the one enacted were discarded by looking for connective regions,and the beams which were projected by headlamps and detached,were connected by the morphological expansion.Secondly,calculating and judging the areas and ranges of connective regions to extract accurately the headlights.Finally,comparing the headlights' areas and position of two adjacent frames to match and track,then calculate traffic flows.Experiment results indicate that this algorithm is low about computation,and its detection ratio reaches above 97% in fine condition.

Camshift is a tracking algorithm based on color histogram.It operates on a back-projection image produced from object histogram model.Initialize a search window size and location.Computer information is used previously to adjust current search window size and location,then location center of the object in the current image.The tracking speed is raised by virtue of predicting the position that a moving object arrives at the next time and reducing the search region.The experimental results are given to show that the proposed algorithm can improve object tracking speed even if complex background and irregular motion.

In allusion to describe detection and tracking of moving targets, it is an import role of the appearance of optical flow, especially for real-time detection of vehicle under dynamic background. There are many advantages for optical flow than other means. The key idea in this paper is to introduce detection both under static and dynamic background. Optical Flow is a more complex algorithm than others because of its large calculation. So in this paper choose to extract features to assist optical flow algorithm in order to lessen work. At the same time, to collect ROI at the basis of some prepro- cessing play a very important part to detect under dynamic background. After these steps we can obtain rough optical flow, so we should make template matching to cut out noise which we don＇t like. So we can get nice result at the end.

The paper introduces how to extract the human star skeleton and by using the particle swarm to optimize the skeleton characteristic vector quantization.Star skeleton,generated by connecting human center-of-mass point to the human limbs and head endpoint,is a kind of fast skeleton extraction techniques.Connecting center-of-mass point with endpoint,human skeleton could be represented by five dimensional vector Si,Si∈Rn,Rn is the star skeleton feature space.Star skeleton sequence can take the place of the timing sequences of human actions.Finally,we generated the codebook G by using particle swarm optimization quantification process.

Frames difference to get moving regions,gradient threshold to get binary images.extracting feature points'optical flow of moving regions,marking optical flow vector by section,setting up ROI,using optical flow to get identification,orientation,tracking of moving target are introduced.This method has real-time and robustness to moving target tracking.It can be used in traffic flow statistics and is also a bedding of driver assistance research.Experimental results show the effectiveness and practical of this method.