利用深度神经网络实现分布式相干瑞利光纤振动事件分类

该文利用分布式相干瑞利光纤传感系统,在西气东输一线无锡至苏州段开展现场测试,采集了光纤沿线车辆行走、机械挖掘、人工锄地、定向钻孔等8种振动作业产生的光纤信号,并提出了一种具有5层结构的全连接深度神经网络用于振动事件分类识别以实现不同振动作业的分级管理。振动作业产生的光纤信号能量集中在低频,该文利用梅尔对数频率的非均匀特性提取了25维单帧信号特征量,并将连续40帧信号特征量组合成高维向量作为网络输入特征向量,实现对不同振动作业时变特性的建模。分类识别结果表明,基于深度神经网络结构的振动信号分类识别器能够有效识别不同振动作业类型,实际线路实验验证了该文算法的有效性。     

基于神经网络的二维不变性目标识别方法

提出了一种基于多层前馈神经网络的二维不变性目标识别方法。利用傅里叶描述器提取具有旋转、平移及尺度不变性的目标形状特征。由于所识别的工业工具具有一个自由度,它们形状有一定的动态变化导致同一目标的形状特征敌意一的不唯一性。文中采用含有两个隐层的多层前馈网络学习及识别这些特征矢量。在实验中,在四类机械工具进行测试,并将所提出方法与最近邻分类器进行比较,结果表明,具有反向传输（ＢＰ）学习算法的多层前馈网络     

基于双向循环神经网络的汉语语音识别*

当前基于深度神经网络模型中,虽然其隐含层可设置多层,对复杂问题适应能力强,但每层之间的节点连接是相互独立的,这种结构特性导致了在语音序列中无法利用上下文相关信息来提高识别效果,而传统的循环神经网络虽然做出了改进,但是只能对上文信息进行利用。针对以上问题,该文采用可以同时利用语音序列中上下文相关信息的双向循环神经网络模型与深度神经网络模型相结合,并应用于语音识别。构建具有5层隐含层的模型,其中第3层为双向循环神经网络结构,其他层采用深度神经网络结构。实验结果表明:加入了双向循环神经网络结构的模型与其他模型相比,较好地提高了识别正确率;噪声对双向循环神经网络汉语识别有重要影响,尤其是训练集和测试集附加噪声类型不同时,单一的含噪声语音的训练模型无法适应不同噪声类型的语音识别;调整神经网络模型中隐含层神经元数量后,识别正确率并不是一直随着隐含层中神经元数量的增加而增加,神经元数量数目增加到一定程度后正确率出现了降低的趋势。     

连接时序分类准则声学建模方法优化

对基于连接时序分类准则(connectionist temporal classification,CTC)的端到端声学建模方法进行研究和优化。研究分析了不同声学特征、建模单元以及神经网络结构对CTC声学模型性能的影响,针对CTC模型中blank符号共享导致的建模缺陷提出了建模单元相关的非共享blank方法进行改进,并引入融合建模单元关联信息的模型初始化方法进一步提高CTC模型的性能。在300小时标准英文数据集Switchboard的实验结果显示,结合非共享blank、时延神经网络以及融合建模单元关联信息的初始化方法,CTC声学模型相对于基线系统在词错误率上取得绝对1.1%的下降,同时在训练速度上取得3.3倍的提高,实验结果证明本文针对端到端声学建模提出的优化方法是有效的。      

基于卷积神经网络的低剂量CT图像去噪方法

为了改善低剂量计算机断层扫描(CT)图像的视觉质量,提出一种基于卷积神经网络的图像去噪方法。网络引入批量归一化,并且学习的是低剂量CT图像到其噪声图像之间的映射;使用空洞卷积在不提高复杂度的情况下增大感受野;此外,还将前后层的特征图进行连接,使后方的卷积层能够利用前方各层的特征图作为输入,鼓励网络中特征图的重用。实验结果表明,与目前较先进的方法相比,所提网络结构在实现了更好去噪效果的同时大幅度降低了网络复杂度,能够快速、显著地改善低剂量CT图像的视觉质量。     

基于禁忌算法优化神经网络的海洋船舶流量预测

为了有效缓解船舶交通拥堵和提高通航效率,对海洋港口和航道管理提供一个更可靠的数据,设计了一种基于禁忌算法优化神经网络的海洋船舶流量预测的方法。首先,建立了基于禁忌算法优化神经网络的海洋船舶流量预测模型。然后,设计了三层的脊波神经网络结构,提出了采用禁忌算法优化脊波神经网络结构参数的具体方法,从而得到一个初始化的脊波神经网络流量预测模型。然后,采用有标签的训练样本数据集对网络进行训练,将满足误差阈值的训练模型作为最终的预测模型。以某港口为例进行仿真实验,结果表明文中得到的预测结果与真实值较为接近,且与其它方法相比,具有更好的预测效果。     

多尺度膨胀卷积神经网络资源三号卫星影像云识别

为提高影像云识别精度,提出一种多尺度膨胀卷积深层神经网络云识别方法。结合卫星影像特征,设计云识别卷积神经网络结构,该结构包含深层特征编码模块、局部多尺度膨胀感知模块以及云区预测解码模块。首先,编码模块中通过基础卷积层获取深度特征;其次,联合多尺度膨胀卷积和池化层共同感知,每层操作连接非线性函数,以提升网络模型的表达能力;最后,云区预测解码模块中融合对应编码模块的特征,再利用L1正则化上采样算法实现端对端的像素级云识别结果。选用典型云遮挡区域影像进行云识别实验,并与Otsu算法和FCN-8S算法进行对比。结果表明,本文所提算法的检测精度较高,Kappa系数显著提升。     

IP/GMPLS/Optical多层网络生存性机制的协调

通过对多层网络生存性问题、独立的网络生存性配置方案和多层网络生存性配置方案的研究分析,基于快速发展的GMPLS技术提出了IP/GMPLS/Optical多层网络结构,明确了多层网络生存性方案存在的问题,并利用GMPLS的统一控制平面进一步提出了基于层间协调机制的集成多层网络生存性方案,获得了优化的多层网络故障恢复能力.     

基于改进卷积神经网络的视网膜血管图像分割

彩色眼底图像视网膜血管分割对于临床医学诊断有重要价值。提出了一种基于改进卷积神经网络的视网膜血管分割方法。首先,将残差学习和密集连接网络(DenseNet)相结合,更充分地利用每一层的特征;通过增加短连接的方式,缩短了低层特征图到高层特征图之间的路径,强化了特征的传播能力。其次,为了提取更多细小血管,在编码器-解码器结构的网络中加入了空洞卷积,在不增加参数的情况下增加感受野。实验结果表明,与现存其他深度学习方法相比,所提出网络结构的参数数量更少,在DRIVE标准数据集上平均准确率达到0.9556,灵敏度达到0.8036,特异性达到0.9778,受试者工作特性(ROC)曲线下的面积(AUC)达到0.9800,比现存其他深度学习方法的分割效果更优。     

基于深度学习的非视域成像

针对非相干光照明下的非视域成像问题,提出一种基于深度学习的解决方法。结合计算机视觉领域中经典的语义分割及残差模型,构造一种URNet网络结构,并改进了经典瓶颈层结构。实验结果表明,改进的网络可以恢复更多的图像细节,并具有一定泛化性,相比于基于非相干光照明的散斑自相关成像技术,该网络恢复性能有较大提升。     

多反复结构模型的精确音乐分离方法

针对基本反复模型音乐分离方法自适应性差的问题,提出一种基于美标度倒谱系数(MFCC)的多反复结构模型的音乐分离方法。首先,提取出音乐信号的MFCC系数矩阵(39维的数据构成);然后利用余弦特性得到其相似矩阵,进而将相似度一致的片段划分到一起,建立不同的反复结构模型;之后结合理想二元掩蔽(]BM)分离出背景音乐及歌声的频谱,相应的时域信号则由傅里叶逆变换获得;最后,在不同类型、长度的音乐文件上测试了算法性能,将提出的算法与Rafii的反复算法和Ozerov的灵活窗非负矩阵分解方法进行对比。实验结果表明,改进方法在分离性能上最高提高3 dB左右,并且对于曲调变换大的音乐提高效果更为明显,从而证实了改进方法是一种有效的音乐分离方法,并且更具稳定性。      

Music/voice separation based on the multi-repeating structure of Mel cepstrum coefficient

For the poor adaptability of the original repeating pattern,an improved music separation method of multi-repeating structure of Mel cepstrum coefficient(MFCC) is proposed.Firstly,the MFCC coefficient matrix(39-dimensional data) of the music signal was extracted.Then the cosine characteristic was applied to the count of similarity matrix of MFCC,and the fragments with consistent similarity are putted together.Next different repeating patterns are built for different groups.Thereby the spectrums of the background music and vocal were separated combined with ideal binary masking(IBM),and the corresponding time domain signals were obtained by inverse Fourier transform.Fnally,the improved method was tested on the music database of different types and length,and the separation results were compared with repeating method of Rafii and the non-negative matrix factorization based on flexible framework method of Ozerov.The experimental results showed that the separation performance of improved method was improved about 3 dB,and the performance of music with melody changed larger was significantly improved.Experiments verified that the improved method was an effective music separation algorithm and more stability.     

On the genre-fication of music: a percolation approach

We analyze web-downloaded data on people sharing their music library. By attributing to each music group usual music genres (Rock, Pop ...), and analysing correlations between music groups of different genres with percolation-idea based methods, we probe the reality of these subdivisions and construct a music genre cartography, with a tree representation. We also discuss an alternative objective way to classify music, that is based on the complex structure of the groups audience. Finally, a link is drawn with the theory of hidden variables in complex networks.     

YIN,a fundamental frequency estimator for speech and music

An algorithm is presented for the estimation of the fundamental frequency (F0) of speech or musical sounds. It is based on the well-known autocorrelation method with a number of modifications that combine to prevent errors. The algorithm has several desirable features. Error rates are about three times lower than the best competing methods, as evaluated over a database of speech recorded together with a laryngograph signal. There is no upper limit on the frequency search range, so the algorithm is suited for high-pitched voices and music. The algorithm is relatively simple and may be implemented efficiently and with low latency, and it involves few parameters that must be tuned. It is based on a signal model (periodic signal) that may be extended in several ways to handle various forms of aperiodicity that occur in particular applications. Finally, interesting parallels may be drawn with models of auditory processing.     

Novel biological visual attention mechanism via Gaussian harmony search

Most of the visual attention models are based on the concept of a two-dimensional saliency map, which encodes the conspicuity of the object in the visual scene. The visual attention model proposed by Laurent Itti is used in this work. In Itti's model, the saliency map is calculated via combining the information across several modalities, including color, intensity, and orientation. In this work, we propose a pre-training process to select the weightings used in the combining of feature maps to make the target more conspicuity in the saliency map. Harmony search (HS) algorithm is used in the pre-training process to obtain the weightings. HS is a new heuristic algorithm, which mimics the improvisation of music players. Its performance has been verified by many benchmark problems. We modify the pitch adjustment process of the original HS to improve the optimization performance and accelerate the convergence rate. The modified algorithm is named Gaussian harmony search (GHS).     

一种音乐情绪参数化的方法

针对目前基于情绪的音乐分类研究存在的弊端,为了方便音乐检索,本文提出一种音乐情绪参数化的方法。该方法通过提取反映音乐情绪的特征向量,然后利用fisher算法进行维数压缩,再通过大量的音乐样本训练得到节奏、音调和音色3个描述音乐情绪的参数,参数的大小反映了情绪的强弱。实验结果表明,音乐情绪参数化的结果符合音乐实际的情绪。     

A central spectrum model: a synthesis of auditory-nerve timing and place cues in monaural communication of frequency spectrum

A probabilistic psychophysical model for monaural communication from the auditory nerve to the brain is given in the form of a tonotopic display of stimulus spectrum, termed central spectrum. The model builds upon prior research demonstrating the potential of neural timing cues from the auditory nerve for conveying information on complex spectra, and was designed to meet the quantified demands of the psychophysics of frequency measurement. The central spectrum magnitude at each frequency is determined by the response of the auditory-nerve fibre with characteristic frequency matching that frequency. An interval histogram from each fiber is passed through a filter matched to the characteristic frequency of the fiber. This output versus characteristic frequency defines the central spectrum. Detailed analysis demonstrates that efficient probabilistic processing of the central spectrum described known psychophysical properties of frequency measurement in discrimination and periodicity pitch experiments. Psychophysical models based upon the central spectrum model followed by optimum probabilistic pattern recognition are potentially relevant for predicting human communication limits in response to arbitrary sounds of speech and music.     

图像序列中目标关键帧快速搜索算法

在目标测量时所获得的图像序列中,如何定位目标关键帧(最有利于目标测量的图像)的位置,对目标识别的效率和测量设备的性能有着显著的影响.针对具有复杂特性的目标图像序列,提出了一种基于帧间像素灰度差值来定位目标关键帧的快速搜索算法.该算法仅仅利用像素灰度值这一最基本的特征,将图像序列中相邻两张图像的同一像素的灰度差值与给定阈值相比较,统计高于阈值的像素个数,再与另一给定阈值相比较,进而确定目标关键帧的位置.实验结果表明,该算法对目标大小不同、形状不同,环境不同,信噪比较高的图像序列都具有快速、稳定的搜索效果.     

基于音色单元分布的音乐结构分析

音乐的结构是音乐作品表达作者思想的一种重要形式,也是听众理解音乐作品内涵的有效途径。本文研究了基于音乐特征的音色单元建模方法,研究了在Fisher准则下,根据局部范围音色单元的分布,采用非监督聚类方法分析音乐的结构。实验结果证明了基于离散余弦变换的音色特征,用音色单元分布聚类算法进行音乐结构分析的有效性。      

Acoustical analysis and model-based sound synthesis of the kantele

The five-string Finnish kantele is a traditional folk music instrument that has unique structural features, resulting in a sound of bright and reverberant timbre. This article presents an analysis of the sound generation principles in the kantele, based on measurements and analytical formulation. The most characteristic features of the unique timbre are caused by the bridgeless string termination around a tuning pin at one end and the knotted termination around a supporting bar at the other end. These result in prominent second-order nonlinearity and strong beating of harmonics, respectively. A computational model of the instrument is also formulated and the algorithm is made efficient for real-time synthesis to simulate these features of the instrument timbre.