贝叶斯优化卷积神经网络公共场所异常声识别*

针对公共场所异常声的感知和识别问题,提出一种基于贝叶斯优化卷积神经网络的识别方法。提取声信号的Gammatone倒谱系数、倍频程功率谱、短时能量和谱质心,组合成声信号的特征图。构建卷积神经网络作为分类器,利用递增的卷积核设置和池化操作处理不同尺度的特征。基于贝叶斯优化算法优化卷积神经网络的模型参数,对包括火苗噼啪声、婴儿啼哭声、烟花燃放声、玻璃破碎声和警报声的5种公共场所异常声进行识别。该方法的识别结果与基于不同的特征提取和分类器方案得到的识别结果进行比较,结果表明该方法的识别效果优于其他特征提取和分类器方案的识别效果。最后分析了该方法在不同信噪比噪声干扰下的识别结果,验证了该方法的有效性。     

水下目标多模态深度学习分类识别研究

水下目标的分类识别对于水声探测具有重要意义。提出一种水下目标多模态深度学习分类识别方法。针对水声信号的一维时域模态和二维频域模态特征建立一种多模态特征融合的深度学习结构,结合长短时记忆网络和卷积神经网络的优点,对一维时域信号和二维频谱信号分别进行并行处理,对输出进行典型相关分析,形成特征融合表示,并利用相邻帧的相关性进行参数优化。利用实测水声信号对算法进行了验证。结果表明：提出的算法对于水下目标识别的精度有显著的提高。     

基于深度学习的透过散射介质偏振识别

当偏振光在散射介质中传播时,由于散射粒子的多重散射作用而造成偏振信息的扰乱和丢失.为了保证偏振信息在散射介质中高效和高保真的传输,提出一种基于深度学习的透过散射介质偏振识别方法.通过构建卷积神经网络来提取散斑光强信息中入射光波偏振信息的特征,实现对入射光波偏振态的高分辨率识别,并使用初始相位不同的偏振光来验证卷积神经网络对偏振态识别的鲁棒性.实验结果表明,所提方法具有识别速度快和准确率高的优点,理论上可以通过无限大的数据来训练神经网络,因此该方法在偏振光学成像和激光通信等领域具有巨大的应用潜力.     

水中分层弹性球壳高频时域回波的声学编码研究

针对水下集群目标及敌我目标识别的难题,该文提出了一种基于水中分层弹性球壳高频时域回波的声学编码原理及方法。推导了水中4层弹性球壳目标散射声压的简正级数解,并与有限元结果进行了对比验证。通过构造高频主动声呐的探测脉冲信号,与4层弹性球壳声传递函数的简正级数解做卷积运算,获得了目标的时域回波脉冲序列。研究了分层弹性球壳的厚度、各层材料属性、排布顺序等对时域回波特征的影响规律,提出了基于时域回波特征的声学编码方法。研究表明：利用水中分层弹性球壳目标高频时域回波特征能够实现声学编码,回波结构稳定,且不受限于探测方向。通过携带或安装这种分层弹性球壳结构,有望识别水下航行体/悬浮体等目标。该文的研究对水下目标的主动探测身份识别及导航等具有一定的参考价值。     

基于鲁棒损失函数的标签有噪信号调制方式识别

针对现实信号调制方式标注易发生错误, 即训练数据集中信号调制方式标签存在噪声情形, 我们选取l₁模损失函数及其推广形式作为对标签噪声具有鲁棒性的损失函数, 结合深度卷积神经网络优良的自动特征提取能力, 提出一种针对信号调制方式存在误判噪声的深度学习算法。该算法在训练数据集合标签噪声率达50%情形下, 对信号调制方式的识别准确率依然保持较高水平。相反, 对于采用通常的交叉熵作为损失函数的深度卷积神经网络, 其已无法对信号调制方式进行分类识别。在公开的数据集上的数值实验表明, 所提算法对于标签有噪信号调制方式识别具有较强的鲁棒性。     

语音情感识别中的特征选择方法

为了解决传统卷积神经网络在识别中文语音时预测错误率较高、泛化性能弱的问题,首先以深度卷积神经网络(DCNN)-连接时序分类(CTC)为研究对象,深入分析了不同卷积层、池化层以及全连接层的组合对其性能的影响;其次,在上述模型的基础上,提出了多路卷积神经网络(MCNN)-连接时序分类(CTC),并联合SENet提出了深度SE-MCNN-CTC声学模型,该模型融合了MCNN与SENet的优势,既能加强卷积神经网络的深层信息的传递、避免梯度问题,又可以对提取的特征图进行自适应重标定。最终实验结果表明:SE-MCNN-CTC相较于DCNN-CTC错误率相对降低13.51%,模型最终的错误率达22.21%;算法改进后的声学模型可以有效地提升泛化性能。     

基于深度学习的船舶辐射噪声识别研究

为了改善船舶辐射噪声识别系统的性能,进一步提高船舶辐射噪声识别的正确率,该文提出采用一种基于深度学习的船舶辐射噪声识别方法。该方法首先提取了船舶辐射噪声的频谱、梅尔倒谱系数等特征,将提取特征后的图像样本分别用于训练卷积神经网络和深度置信网络,再对船舶辐射噪声进行识别。通过文中所给实例,将深度学习和支持向量机两种识别方法的性能进行比较,得出深度学习方法可以有效地提高船舶辐射噪声识别正确率的初步结论。     

基于双耳模型的倒谱系数在声目标识别中的应用

在声目标识别中,基于单路信号的特征提取算法难以保证特征在低信噪比条件下的鲁棒性。针对该问题,本文使用由两个传声器组成的小孔径阵列采集信号,采用Stereausis双耳听觉模型对两路信号进行综合处理,在此基础上设计了基于双耳模型的倒谱系数,并将其作为识别特征应用于多类地面目标识别。试验表明该特征可有效提高识别系统的识别性能和抗噪声能力。     

基于兴趣区域掩码卷积神经网络的红外-可见光图像融合与目标识别算法研究

建立权重独立的双通道残差卷积神经网络,对可见光与红外频段下的目标图像进行特征提取,生成多尺度复合频段特征图组。基于像点间的欧式距离计算双频段特征图显著性,根据目标在不同成像频段下的特征贡献值进行自适应融合。通过热源能量池化核与视觉注意力机制,分别生成目标在双频段下的兴趣区域逻辑掩码并叠加在融合图像上,凸显目标特征并抑制非目标区域信息。以端到端识别网络作为基础,利用交叉损失计算策略,对含有注意力掩码的多尺度双频段融合特征图进行目标识别。结果表明,所设计的识别网络能够有效地融合目标红外热源物理特征和可见光图像纹理特征,提高了信息融合深度,保留目标热辐射与纹理特征的同时降低了背景信息干扰,对全天候复杂环境下的多尺度热源目标具有良好的识别精度与鲁棒性。     

基于深度学习的252Cf源驱动核材料浓度识别技术

针对核武器/核材料识别系统中核材料浓度识别的关键技术问题,采用Monte Carlo方法,通过建立252Cf源驱动核材料裂变中子信号样本库,模拟分析了随探测器距离和角度及核材料浓度变化的裂变脉冲中子信号特点,基于深度学习之卷积神经网络,构建了一种252Cf源驱动核材料浓度识别方法,实现了对测试样本浓度的识别,且还与BP神经网络和K最近邻方法进行了对比试验研究。结果表明,卷积神经网络算法进行核材料浓度识别,得到了高达92.05%识别准确率,不仅解决了因探测器距离和角度变化时对核材料浓度识别准确率影响的难题,而且还获得了优于BP神经网络和K最近邻算法对核材料浓度识别的认识,这为252Cf源驱动核材料浓度识别提供了一种新的途径。     

Recognition of Putonghua voiceless stop like initials based on speech main periods

I.Intr0ductionNowadays,thereismuchadvancemcntinthcrcsearchintospeechrecognition.Manyresearchershavebecninterestedintheimplementationofareliab1crealtimerec-ognitionsystemofunlimitedv0cabu1ary.Thercareafewproductsconversingsyl1ablesintoChinesecharactersinthemarket.However,theimp1ementationofarobustrealtAnerecognitionsystemofunlimitcdvocabularyisvcrydifficu1t,anditisthcgreataimofourresearch.WehaveaTMS32O-C25signa1processingboardattachedtoacomputerofthM-PC/AT80386.Wehopcthatourspeechrecognit…     

Integrating the original face images and “symmetrical faces” to perform face recognition

Using the original and ‘symmetrical face’ training samples to perform representation based face recognition was first proposed in [1]. It simultaneously used the original and ‘symmetrical face’ training samples to perform a two-step classification and achieved an outstanding classification result. However, in [1] the “symmetrical face” is devised only for one method. In this paper, we do some improvements on the basis of [1] and combine this “symmetrical faces” transformation with several representation based methods. We exploit all original training samples, left “symmetrical face” training samples and right “symmetrical face” training samples for classification and use the score fusion for ultimate face recognition. The symmetry of the face is first used to generate new samples, which is different from original face image but can really reflect some possible appearance of the face. It effectively overcomes the problem of non-sufficient training samples. The experimental results show that the proposed scheme can be used to improve a number of traditional representation based methods including those that are not presented in the paper.     

用于畸变目标相关识别的匹配滤波器设计

针对尺度缩放和角度旋转变化目标相关识别率低的问题,采用优化折中综合鉴别函数方法设计了具有较大畸变容限的匹配滤波器。在匹配滤波器设计时,通过调整相关输出噪声、训练样本平均功率谱密度和训练样本相似度对训练样本平均值的贡献,得到具有带通性质的匹配滤波器。利用单个匹配滤波器对角度旋转0°～30°、尺度变化0～20%的目标进行了相关识别仿真实验。结果表明：这种匹配滤波器相关输出峰尖锐、鉴别力强、畸变容限大,可用于畸变目标的光学相关识别。     

Iris segmentation for visible wavelength and near infrared eye images

State-of-the-art iris segmentation algorithms exhibit poor performance for non-ideal data, which is mainly because of the noise such as low contrast, non-uniform illumination, reflections, and among others. To address this issue, a robust iris segmentation scheme is proposed that includes the following: First, a set of the Seed-pixels in a preprocessed eye image is marked adaptively. Next, a two-fold scheme based on a Circu-differential accumulator (CDA) and gray statistics is adopted to localize coarse iris region robustly. Notably, the proposed CDA has close resemblance with the Hough transform; however, it consumes relatively less memory and is free from thresholding as well. Similarly, pupillary boundary is localized, which is verified through an intensity test as well. Next, a refine estimate for the limbic boundary is extracted. After that, iris boundaries are regularized using the Fourier series. Finally, the eyelids are localized using a Para-differential accumulator (PDA), and eyelashes and reflections are also localized adaptively in the polar form of iris. Experimental results on the near infrared (NIR) and visible wavelength (VW) iris databases show that the proposed technique outperforms contemporary approaches.     

基于深度优先遍历的图像边缘检测方法

图像边缘识别是图像处理的重要组成。提出一种基于深度优先遍历的梯度分割算法,这种算法首先构造像素点的数据结构,然后从图像任意点出发,估算该点附近的像素点并获取梯度值,如果该点满足边缘点的特征,则从该点出发深度遍历寻找垂直于梯度方向上的边缘点,并标记访问过的点。如果该方向上没有满足条件的点则回退,从某个具有仅次于最大梯度值的方向继续遍历,并标记开始遍历的点为角点,直至遍历全图。该算法将图像的边缘点和角点明显地分割出来,便于识别,对带有不同类型的噪声图像进行处理也可取得较满意效果。     

Quaternion-based improved LPP method for color face recognition

In recent years, pattern recognition and computer vision have increasingly become the focus of research. Locality preserving projection (LPP) is a very important learning method in these two fields and has been widely used. Using LPP to perform face recognition, we usually can get a high accuracy. However, the face recognition application of LPP suffers from a number of problems and the small sample size is the most famous one. Moreover, though the face image is usually a color image, LPP cannot sufficiently exploit the color and we should first convert the color image into the gray image and then apply LPP to it. Transforming the color image into the gray image will cause a serious loss of image information. In this paper, we first use the quaternion to represent the color pixel. As a result, an original training or test sample can be denoted as a quaternion vector. Then we apply LPP to the quaternion vectors to perform feature extraction for the original training and test samples. The devised quaternion-based improved LPP method is presented in detail. Experimental results show that our method can get a higher classification accuracy than other methods.     

A palmprint recognition method based on multi-step representation

Sparse representation uses all training samples to represent a test sample only once, which can be regarded as a one step representation. However, in palmprint recognition, the appearances of palms are highly correlated which means the information provided by all the training samples are redundant while using the representation-based methods. Hence, how to obtain suitable samples for representation deserves exploring. In this paper, we devise a multi-step representation manner to extract the most representative samples for representation and recognition. In addition, the proposed sample selection strategy is based on contributions of the classes, not merely the effort of a single sample. Compared with some other appearance-based methods, the proposed method obtained a competitive result on PolyU multispectral palmprint database.     

A novel method for recognizing face with partial occlusion via sparse representation

In this paper, we propose a novel method to recognize the partially occluded face images based on sparse representation. Generally, occlusions, such as glasses and scarf, fall on some random patch of image's pixels of test images, but which is known to be connected. In our method, all images are divided into several blocks and then an indicator based on linear regression technique is presented to determine whether a block is occluded. We utilize those uncontaminated blocks as the new feature of an image. Finally, the sparse representation-based classification (SRC) method serves as the classifier to recognize unknown faces. In the original work of SRC, the extended SRC (eSRC) scheme is presented to deal with occlusions, which has very heavy computational cost. The experimental results show that our method can achieve good recognition accuracy and has much lower computational cost than eSRC.     

四元量化的综合鉴别函数滤波器实现比例不变相关识别

设计一种基于综合鉴别函数(SDF)的比例不变相关识别滤波器，将频域中被SDF调制的相关输出函数进行复数形式的四元量化（QFCQ），并采用叠代算法调整SDF的系数，使输出的相关峰尖锐，区分能力强，同时满足不同比例图像等相关峰输出的准则，从而达到了比例不变识别的要求.计算机仿真实验对属于训练图像系列的目标图像进行了识别，并与基本的SDF及二元化纯相位SDF做了比较，结果表明该方法具有较好的识别效果.