Complex temporal sequence learning based on short-term memory

An approach to storing of temporal sequences that deals with complex temporal sequences directly is presented. Short-term memory (STM) is modeled by units comprised of recurrent excitatory connections between two neurons. A dual-neuron model is proposed. By applying the Hebbian learning rule at each synapse and a normalization rule among all synaptic weights of a neuron, it is shown that a quantity called the input potential increases monotonically with sequence presentation, and that the neuron can only be fired when its input signals are arranged in a specific sequence. These sequence-detecting neurons form the basis for a model of complex sequence recognition that can tolerate distortions of the learned sequences. A recurrent network of two layers is provided for reproducing complex sequences     

同伦方法在图像稀疏去噪中的应用

本文在深入研究稀疏表示和字典学习理论的基础上,建立了图像去噪模型并提出一种新的图像去噪算法。该算法采用同伦方法学习字典,充分利用了同伦方法收敛速度快以及对信号的恢复准确度高的特点。之后利用 OMP 算法求出带噪图像在该字典下的稀疏表示系数,并结合稀疏去噪模型实现对图像的去噪。实验结果显示本文算法在不同的噪声环境下具有较好的去噪效果,同时在与 K-SVD 算法关于收敛速度比较的实验中,实验结果充分显示了使用同伦算法学习字典在收敛速度上的优势。      

一种非开关型快速随机脉冲噪声降噪算法

为提高现有开关型随机脉冲噪声（Random-Valued Impulse Noise,RVIN）降噪算法的降噪性能,提出了一种基于卷积神经网络的非开关型RVIN快速降噪算法（Fast Non-switching RVIN Denoising Algorithm,FNRDA）.首先,利用噪声检测器随机地检测给定噪声图像中少量不同位置处的像素点;然后,将检测为RVIN噪声点的个数除以被检像素点总数转化为噪声比例值;最后,根据噪声比例值调用相应预先训练好的非开关型卷积神经网络降噪模型,快速且高质量地完成图像降噪任务.实验结果表明：所提出的非开关型FNRDA算法在各噪声比例下的综合性能（降噪效果和执行效率）优于经典的开关型RVIN降噪算法,适用于图像恢复、信号检测、无线通讯等实时系统中.     

A directional margin paradigm for noise suppression in face recognition

Convolutional neural networks (CNN) have achieved outstanding face recognition (FR) performance with increasing large-scale face datasets. With face dataset size grown, noisy data will inevitably increase, undoubtedly bringing difficulties to data cleaning. In this paper, the probability that the sample belongs to noise can be determined based on the cosine distance (cos$\theta$) of normalized angle center and face feature vector in the margin-based loss functions. According to this finding, we propose a two-step learning method integrated into the loss function. The new proposed directional margin loss function combines the noise probability with the label as the supervision information. Experiments show that our method can tolerate noisy data and get high FR accuracy when the training datasets mix with more than 30% noise. Our approach can also achieve a great result of 79.33% in MegaFace challenge one using a noisy training dataset.     

决策模板法在决策层融合目标识别中的应用和改进

决策模板法是一种简单直观的决策层融合识别算法,但是经典的决策模板法没有充分利用各传感器对于不同类目标鉴别能力的先验信息。本文提出利用传感器平均度量熵对决策模板法进行修正,合理度量各个传感器对不同类目标的分类鉴别能力,仿真结果表明改进的决策模板法能提高目标正确识别率。     

A distributed robotic control system based on a temporal self-organizing neural network

A distributed robot control system is proposed based on a temporal self-organizing neural network, called competitive and temporal Hebbian (CTH) network. The CTH network can learn and recall complex trajectories by means of two sets of synaptic weights, namely, competitive feedforward weights that encode the individual states of the trajectory and Hebbian lateral weights that encode the temporal order of trajectory states. Complex trajectories contain repeated or shared states which are responsible for ambiguities that occur during trajectory reproduction. Temporal context information are used to resolve such uncertainties. Furthermore, the CTH network saves memory space by maintaining only a single copy of each repeated/shared state of a trajectory and a redundancy mechanism improves the robustness of the network against noise and faults. The distributed control scheme is evaluated in point-to-point trajectory control tasks using a PUMA 560 robot. The performance of the control system is discussed and compared with other unsupervised and supervised neural network approaches. We also discuss the issues of stability and convergence of feedforward and lateral learning schemes.     

分离通道联合卷积神经网络的自动调制识别

针对通信信号的自动调制识别需要大量特征提取的问题,提出了一种分离通道卷积神经网络自动调制识别算法。该算法通过结合深度学习中卷积神经网络(CNN),分别提取时域信号的多通道和分离通道调制特征,再利用融合特征实现不同信号的分类。仿真结果表明,相比基于CNN的算法,所提算法在高信噪比下针对两个数据集的识别率分别提升7%和18%;此外,相比于基于特征提取的传统识别算法,其高阶调制识别性能平均提升3 dB。     

A sparsity-ranking edge-preservation filter for removal of high-density impulse noises

In this study, a novel sparsity-ranking edge-preservation filter (SREPF) is proposed for removal of high-density impulse noise in images. Using the sparse matrix representation, the first stage of SREPF is not only to identify the noisy candidates but also to decide the processing order of them via a rank of noise-pixel sparsity in the working window. Then the second stage of SREPF utilizes a modified double Laplacian convolution to confirm the truly noisy pixels and yield a directional mean to recover them. This new approach has achieved more remarkable success rate of the edge detection than other edge-preservation methods especially in high noise ratio over 0.5. As a result, SREPF has significant improvements in terms of edge preservation and noise suppression exhibited by the peak signal-to-noise ratio (PSNR) and the structural similarity index metric (SSIM). Simulation results show that this method is capable of producing better performance compared to several representative filters.     

基于软阈值深度学习的自动调制识别算法

自动调制识别是认知无线电、电子侦察、电磁态势生成中重要的环节.由于电磁环境日益复杂,噪声对能否正确调制识别影响显著.本文针对低信噪比（signal-noise ratio,SNR）环境条件设计了一种基于软阈值的深度学习模型,在卷积神经网络（convolutional neural networks, CNN）的基础上加入软阈值函数.将IQ数据转化为幅度相位信息作为模型的输入,CNN用于提取幅度相位数据中的特征,软阈值学习网络可以针对不同特征设置不同阈值,用于滤除样本噪声,提高低SNR条件下的识别率.在开源数据集RML2016.10a上验证了所提算法的有效性,对比其他网络结构,本文提出的模型识别率更高且效率更高.     

基于LS-SVM的复轮廓波变换的图像去噪

针对传统的轮廓波变换图像去噪时引入边缘混叠现象,提出了复轮廓波变换(Complex Contourlet Transform,CCT)和最小二乘支持向量机(LS-SVM)的图像去噪方法.该方法充分利用了复轮廓变换的平移不变性、多方向性以及LS-SVM的小样本学习能力,应用训练好的LS-SVM模型将含噪图像的CCT系数分为含噪点和非含噪点,进行去噪处理.仿真结果表明该算法有效保护图像边缘纹理信息,其峰值信噪比明显高于其他算法,并且具有良好的视觉效果.     

一种噪声环境下的雷达目标高分辨率距离像鲁棒识别方法

随着深度学习技术被应用于雷达目标识别领域,其自动提取目标特征的特性大大提高了识别的准确率和鲁棒性,但噪声环境下的鲁棒性有待进一步研究。该文提出了一种在噪声环境下基于卷积神经网络(CNN)的雷达高分辨率距离像(HRRP)数据识别方法,通过增强训练集和使用残差块、inception结构和降噪自编码层增强网络结构,实现了在较宽信噪比范围下的较高识别率,其中在信噪比为0 dB的瑞利噪声条件下,识别率达到96.14%,并分析了网络结构和噪声类型对结果的影响。      

含噪语音压缩感知自适应快速重构算法

本文针对含噪语音压缩感知在低信噪比时重构性能差的问题,提出了一种自适应快速重构算法。该算法将行阶梯观测矩阵与一种新型的快速重构算法结合,并根据含噪语音信号的信噪比自适应选择最佳重构参数,使得在重构语音的同时提高了重构信噪比。算法实现简单快速,且不需要预先计算信号的稀疏度。实验结果表明,自适应快速重构算法重构性能优于基追踪算法和自适应共轭梯度投影算法以及快速重构算法,重构速度略慢于快速重构算法,但快于基追踪算法和自适应共轭梯度投影算法。      

Learning Sparse Overcomplete Codes for Images

Images can be coded accurately using a sparse set of vectors from a learned overcomplete dictionary, with potential applications in image compression and feature selection for pattern recognition. We present a survey of algorithms that perform dictionary learning and sparse coding and make three contributions. First, we compare our overcomplete dictionary learning algorithm (FOCUSS-CNDL) with overcomplete independent component analysis (ICA). Second, noting that once a dictionary has been learned in a given domain the problem becomes one of choosing the vectors to form an accurate, sparse representation, we compare a recently developed algorithm (sparse Bayesian learning with adjustable variance Gaussians, SBL-AVG) to well known methods of subset selection: matching pursuit and FOCUSS. Third, noting that in some cases it may be necessary to find a non-negative sparse coding, we present a modified version of the FOCUSS algorithm that can find such non-negative codings. Efficient parallel implementations in VLSI could make these algorithms more practical for many applications.     

神经网络中将任务学习与模型学习相结合的E-H和E-H-W学习方法

本文从获取好的神经网络泛化能力出发,首先提出了将Hebbian学习与增加问题复杂性统一起来的思想,并通过在总的误差函数中增加一限制函数来实现Hebbian学习.基于此,提出了一种将误差驱动的任务学习与Hebbian规则的模型学习相结合的E-H方法.然后,根据模型学习应同时考虑减小网络复杂性和增加问题复杂性的思想,又提出了一种将误差驱动的学习与Hebbian规则、简单的权退化法结合起来,共同来提高神经网络的泛化能力的E-H-W方法.最后通过大量实例仿真将它们与纯误差驱动的方法、权退化法、其它文献中的相关方法进行了比较.结果表明我们的方法具有最好的泛化能力,是很有效的神经网络学习方法.     

基于信噪比分类网络的调制信号分类识别算法

针对传统降噪算法损伤高信噪比(SNR)信号而造成信号识别准确率下降的问题,该文提出基于卷积神经网络的信噪比分类算法,该算法利用卷积神经网络对信号进行特征提取,用固定K均值(FK-means)算法对提取的特征进行聚类处理,准确分类高低信噪比信号。低信噪比信号采用改进的中值滤波算法降噪,改进的中值滤波算法在传统中值滤波的基础上增加了前后采样窗口的关联性机制,来改善传统中值滤波算法处理连续噪声效果不佳的问题。为充分提取信号的空间特征和时间特征,该文提出卷积神经网络和长短时记忆网络并联的卷积长短时(P-CL)网络,利用卷积神经网络和长短时记忆网络分别提取信号的空间特征与时间特征,并进行特征融合与分类。实验表明,该文提出的调制信号分类模型识别准确率为91%,相比于卷积长短时(CNN-LSTM)网络提高了6%。     

CSM neural network for degraded printed character optical recognition

In pattern recognition applications, the classification power of a system can be improved by combining several classifiers. Obviously performance of the system cannot be improved if the individual classifiers make all the same mistakes, thus it is important to use different features and different structures in the individual classifiers. In this context, we propose a two subnets neural network called CSM net. The first subnet, or similarity layer, is operating as a similarity measure neural network; it is based on the complementary similarity measure method (CSM). The second subnet is a competitive neural network (CNN) based on the winner takes all algorithm (WTA) that is used for the classification. In the proposed neural architecture, the statistical CSM method is analyzed, and implemented in the form of a feed forward neural network, it is named “similarity measure neural network” (SMNN). We show that the resulting SMNN synaptic weights are modified versions of the model patterns used in the training set, and that they can be considered as a memory network. We introduce a relative distance data calculated from the SMNN output, and we use it as a quality measurement tool of the degraded characters, what makes the SMNN classifier very powerful, and very well-suited for features rejections. This relative distance is used by the SMNN and compared to a first rejection threshold to accept, or reject, the incoming characters. In order to guarantee a higher recognition and reliability rates for the cascaded method, the SMNN is combined with a second subnet based on the WTA for classification using a second specific rejection threshold. These two submits combination (CSM net) boost the performance of the SMNN classifier. This is resulting in a robust multiple classifiers that can be used for setting the entire rejection threshold. The experimental results that we introduce are related to the proposed method, but the tests are introduced with various impulse noise levels, as well as the tests with broken and manually corrupted characters, and characters with various levels of additive Gaussian noise. The experiments show the effective ability of the model to yield relevant and robust recognition on poor quality printed checks, and show that the CSM net outperforms the previous works, both in efficiency and accuracy.     

一种基于Chirp原子分解的语音增强方法

由于语音信号在时频面上具有局部连续结构，文章提出了一种基于Chirp时频原子分解的语音增强方法．该方法将含噪的语音信号使用匹配追踪算法分解成Chirp原子的组合，根据语音和噪声所对应的Chirp原子在参数上的不同，从中分离出属于语音的Chirp原子来重构语音信号，从而达到去除增强语音的目的。仿真实验结果表明．经该法处理后的语音信号的信噪比有较大的提高，主观试昕效果也较好。     

结合深度迭代缩放卷积神经网络的PRNU提取算法

光响应非均匀性（photo-response non-uniformity，PRNU）是用于数字图像设备溯源的一种重要特征，也被称为成像设备指纹。针对图像真实噪声包含PRNU和大量未知噪声的复杂特性，本文提出一种结合深度迭代缩放卷积神经网络的PRNU数字成像设备指纹提取算法。首先，通过连续重复的缩小与放大特征图的分辨率来提高GPU内存利用效率和生成大的感受野，尽可能的提取包含完整PRNU指纹的真实噪声。然后，利用来自同一数字成像设备多幅图像的噪声残差来估计PRNU指纹。本文算法在相机溯源数据集Dresden和手机溯源数据集Daxing上进行了测试。实验结果显示，与传统方法相比本文算法具有的更好的识别率和普适性。      

基于专家场先验模型的图像超分辨重建算法

为了进一步提高图像超分辨率重建的质量,针对非局部集中稀疏表示算法中重建图像的噪声问题,提出了一种基于专家场先验模型的图像超分辨率重建改进算法。首先,利用专家场模型从图像训练集中学习整幅图像的先验知识建立全局先验模型;然后将学习到的先验信息用于非局部集中稀疏表示模型求解最优稀疏表示系数;最后,得到高分辨率图像估计。该算法在超分辨率重建迭代运算的同时,同步更新专家场模型参数,因此在不显著增加运算复杂度的情况下,通过选取合适的先验约束,有效地增强了图像重建的效果。实验结果表明:相比非局部集中稀疏表示算法,文中算法对无噪和有噪降质图像均能取得较好的峰值信噪比结果,并且能够进一步提高有噪图像的去噪效果。