利用深度卷积神经网络提高未知噪声下的语音增强性能

为了进一步提高基于深度学习的语音增强方法在未知噪声下的性能,本文从神经网络的结构出发展开研究.基于在时间与频率两个维度上,语音和噪声信号的局部特征都具有强相关性的特点,采用深度卷积神经网络（Deep convolutional neural network,DCNN）建模来表示含噪语音和纯净语音之间的复杂非线性关系.通过设计有效的训练特征和训练目标,并建立合理的网络结构,提出了基于深度卷积神经网络的语音增强方法.实验结果表明,在未知噪声条件下,本文方法相比基于深度神经网络（Deep neural network,DNN）的方法在语音质量和可懂度两种指标上都有明显提高.     

基于无监督深度图像生成的盲降噪模型

鉴于有监督神经网络降噪模型的数据依赖缺陷,提出了一种基于无监督深度生成（UDIG）的盲降噪模型。首先,利用噪声水平评估（NLE）算法测定给定噪声图像中的噪声水平值并输入到主流FFDNet降噪模型中,所得到降噪后的图像（称为初步降噪图像）作为UDIG降噪模型的输入。其次,选用编码器—解码器架构作为UDIG模型的骨干网络并用UDIG模型的输出图像（即生成图像）分别与初步降噪图像、噪声图像之间的均方误差之和构建混合loss函数;再次,以loss最小化为优化目标,通过随机梯度下降（SGD）网络训练算法调整网络模型的参数值从而获得一系列生成图像;最后,当残差图像（噪声图像与生成图像之间）的标准差逼近之前NLE算法所测定的噪声水平估计值时及时终止网络迭代训练过程,从而确保生成图像（作为降噪后图像）的图像质量最佳。实验结果表明：与现有的主流降噪模型（算法）相比,UDIG降噪模型在降噪效果上具有显著优势。     

Image enhancement using deep-learning fully connected neural network mean filter

Improving the quality of a noisy image is important for image applications. Many novel schemes pay great efforts in the removal of impulse noise. Most of them restore noisy pixels only by using the neighboring noise-free pixels, but the relationship between a noisy image and its noise-free one, which denotes the clean image not corrupted by noise, is ignored. So the reconstruction quality cannot be further improved. In this study, we employ a deep-learning fully connected neural network (FCNN) to select top N candidates of neighboring un-corrupted pixels for the restoration of a center noisy pixel in an analysis window. Hence, the mean value of the gray levels of these top N pixels is computed and employed to replace the noisy pixel, yielding the noisy pixel being restored. The experimental results reveal that the proposed deep-learning FCNN mean filter can remove impulse noise effectively in corrupted images with different noise densities.

     

A separation–aggregation network for image denoising

Image denoising is the problem that aims at recovering a clean image from a noisy counterpart. A promising solution for image denoising is to employ an appropriate deep neural network to learn a hierarchical mapping function from the noisy image to its clean counterpart. This mapping function, however, is generally difficult to learn since the potential feature space of the noisy patterns can be huge. To overcome this difficulty, we propose a separation–aggregation strategy to decompose the noisy image into multiple bands, each of which exhibits one kind of pattern. Then a deep mapping function is learned for each band and the mapping results are ultimately assembled to the clean image. By doing so, the network only needs to deal with the compositing components of the noisy image, thus makes it easier to learn an effective mapping function. Moreover, as any image can be viewed as a composition of some basic patterns, our strategy is expected to better generalize to unseen images. Inspired by this idea, we develop a separation–aggregation network. The proposed network consists of three blocks, namely a convolutional separation block that decomposes the input into multiple bands, a deep mapping block that learns the mapping function for each band, and a band aggregation block that assembles the mapping results. Experimental results demonstrate the superiority of our strategy over counterparts without image decomposition.     

AKSNet: A novel convolutional neural network with adaptive kernel width and sparse regularization for machinery fault diagnosis

Convolutional kernels have significant affections on feature learning of convolutional neural network (CNN). However, it is still a challenging problem to determine appropriate kernel width. Moreover, some features learned by convolutional layers are still redundant and noisy. Thus, adaptive selection of kernel width and feature selection of feature maps are key techniques to improve feature learning performance of CNNs. In this paper, a new deep neural network (DNN) model, adaptive kernel sparse network (AKSNet) is proposed to extract multi-scale fault features from one-dimensional (1-D) vibration signals. Firstly, an adaptive kernel selection method is developed, where multiple branches with different kernels are used to extract multi-scale features from vibration signals. Channel-wise attention is developed to fuse features generated by these kernels to obtain different informative scales. Secondly, a spatial attention is used for dynamic receptive field to focus on salient region of feature maps. Thirdly, a sparse regularization layer is embedded in the deep network to further filter noise and highlight impaction of the feature maps. Finally, two cases are adopted to verify effectiveness of AKSNet-based feature learning for bearing fault diagnosis. Experimental results show that AKSNet can effectively extract features from multi-channel vibration signals and then improves fault diagnosis performance of the classifier significantly. AKSNet shows better recognition performance in comparison with that of shallow neural networks and other typical DNNs.     

A robust neuro-fuzzy network approach to impulse noise filtering for color images

Based on an integration of a simple impulse detector and a robust neuro-fuzzy (RNF) network, an effective impulse noise filter for color images is presented. It consists of two modes of operation, namely, training and testing (filtering). During training, the impulse detector is used to locate the noisy pixels in the color images for optimizing the RNF network. During testing, if a pixel is detected as a corrupted one according to the impulse detector, the trained RNF network will be triggered to output a new pixel to replace it. The proposed impulse noise filter is distinguished by two properties. The first is the use of a simple impulse detector, which is efficient and yet effective in detecting the noisy pixels in color images. The other is the use of a novel membership function in the design of the adaptive RNF network, making the network robust to impulse noise. As demonstrated by the experimental results, the proposed filter not only has the abilities of noise attenuation and details preservation but also possesses desirable robustness and adaptive capabilities. It outperforms other conventional multichannel filters.     

双通道扩张卷积注意力图像去噪网络*

针对深度学习图像去噪算法存在网络过深导致细节丢失的问题,提出一种双通道扩张卷积注意力网络CEANet。拼接信息保留模块将每一层的输出特征图融合,弥补卷积过程中丢失的图像细节特征进行密集学习;扩张卷积可以在去噪性能和效率之间进行权衡,用更少的参数获取更多的信息,增强模型对噪声图像的表示能力,基于扩张卷积的稀疏模块通过扩大感受野获得重要的结构信息和边缘特征,恢复复杂噪声图像的细节;基于注意力机制的特征增强模块通过全局特征和局部特征进行融合,进一步指导网络去噪。实验结果表明,在高斯白噪声等级为25和50时,CEANet都获得了较高的峰值信噪比均值和结构相似性均值,能够更高效地捕获图像细节信息,在边缘保持和噪声抑制方面,具有较好的性能。相关实验证明了该算法进行图像去噪的有效性。     

联合优化深度神经网络和约束维纳滤波的单通道语音增强方法

深度神经网络（Deep neural networks,DNNs）依靠其良好的特征提取能力,在语音增强任务中得到了广泛应用。为进一步提高深度神经网络的语音增强效果,提出一种将深度神经网络和约束维纳滤波联合训练优化的新型网络结构。该网络首先对带噪语音幅度谱进行训练并分别得到纯净语音和噪声的幅度谱估计,然后利用语音和噪声的幅度谱估计计算得到一个约束维纳增益函数,最后利用约束维纳增益函数从带噪语音幅度谱中估计出增强语音幅度谱作为网络的训练输出。对不同信噪比下的20种噪声进行的仿真实验表明,无论噪声类型是否在网络的训练集中出现,本文方法都能够在有效去除噪声的同时保持较小的语音失真,增强效果明显优于DNN及NMF增强方法。     

稀疏差分网络和多监督哈希用于高效图像检索

针对基于深度哈希的图像检索中卷积神经网络（CNN）特征提取效率较低和特征相关性利用不充分的问题,提出一种融合稀疏差分网络和多监督哈希的新方法SDNMSH（sparse difference networks and multi-supervised hashing）,并将其用于高效图像检索。SDNMSH以成对的图像作为训练输入,通过精心设计的稀疏差分卷积神经网络和一个监督哈希函数来指导哈希码学习。稀疏差分卷积神经网络由稀疏差分卷积层和普通卷积层组成。稀疏差分卷积层能够快速提取丰富的特征信息,从而实现整个网络的高效特征提取。同时,为了更加充分地利用语义信息和特征的成对相关性,以促进网络提取的特征信息能够更加有效地转换为具有区分性的哈希码、进而实现SDNMSH的高效图像检索,采用一种多监督哈希（MSH）函数,并为此设计了一个目标函数。在MNIST、CIFAR-10和NUS-WIDE三个广泛使用的数据集上进行了大量的对比实验,实验结果表明,与其他先进的深度哈希方法相比,SDNMSH取得了较好的检索性能。     

面向大规模图像分类的深度卷积神经网络优化

在图像分类任务中,为了获得更高的分类精度,需要对图像提取不同层次的特征信息。深度学习被越来越多的应用于大规模图像分类任务中。本文提出了一种基于深度卷积神经网络的,可应用于大规模图像分类的深度学习框架。该框架在经典的深度卷积神经网络AlexNet基础上分别从网络框架和网络内部结构两个方面对网络做了优化和改进,进一步提升了网络的特征表达能力。同时,通过在全连接层引入隐层使得网络能够同时具备学习图像特征和二值哈希的功能,使得该框架具有处理大规模图像数据的能力。通过在三个标准数据库中的一系列比对实验,分析了不同优化方法在不同情况下的作用,并证明了本文所提优化方法的有效性。     

基于复合卷积神经网络的图像去噪算法*

基于深度学习理论,将图像去噪过程看成神经网络的拟合过程,构造简洁高效的复合卷积神经网络,提出基于复合卷积神经网络的图像去噪算法.算法第1阶段由2个2层的卷积网络构成,分别训练阶段2中的3层卷积网络中的部分初始卷积核,缩短阶段2中网络的训练时间和增强算法的鲁棒性.最后运用阶段2中的卷积网络对新的噪声图像进行有效去噪.实验表明文中算法在峰值信噪比、结构相识度及均方根误差指数上与当前较好的图像去噪算法相当,尤其当噪声加强时效果更佳且训练时间较短.     

基于U-Net融合的保留纹理的图像去噪方法

图像的噪声阻碍了高级视觉任务对图像的理解,且去除图像的噪声是一个具有挑战性的任务.现有的基于卷积神经网络的图像去噪方法在去除噪声的同时,对图像纹理会引入一定程度的破坏,导致去噪后图像无法保留图像的纹理.为了解决这个问题,本文提出一种用二分支U-Net网络来融合特征和保留纹理的图像去噪方法.首先选取一种去噪方法的两个不同去噪参数的预训练模型分别得到同一张噪声图像的不同去噪结果,其中一个结果中去噪效果比纹理保留效果好,另一个结果中纹理保留比去噪效果好.然后将这两个去噪图像作为卷积神经网络的输入,利用两个编码器分别提取图像的特征,并同时放入融合模块融合图像的特征,最后利用解码器重建出无噪声图像.实验结果表明,与现有的方法相比本文的方法更有效,在去除噪声的同时能保留更多的图像纹理信息.     

核磁共振图像重构的即插即用算法

为解决核磁共振图像重构中由于欠采样导致的重构图像质量较低的问题,提出了一种基于凸-非凸稀疏正则和即插即用近似点梯度下降的核磁共振图像重构算法。首先给出了凸-非凸稀疏正则的近似点算子。然后基于该近似点算子提出近似点梯度下降算法。最后将上述算法中的近似点算子用某种合适的去噪器（如神经网络去噪器）替换,得到即插即用近似点梯度下降算法,并将其应用到核磁共振图像重构上。数值实验中,分别用不同的待重构图像、采样模板和去噪器进行对比实验,实验结果表明,所提算法在使用神经网络去噪器时,峰值信噪比较已有算法提升了6.26 dB。同时视觉效果也得到了显著的提升,在处理边缘和纹路方面效果都更加明显,从而验证了算法的有效性。     

两阶段多层感知的随机脉冲噪声比例预测

目的 基于卷积神经网络（CNN）在图块级上实现的随机脉冲噪声（RVIN）降噪算法在执行效率方面较经典的逐像素点开关型降噪算法有显著优势,但降噪效果如何取决于能否对降噪图像受噪声干扰程度（噪声比例值）进行准确估计。为此,提出一种基于多层感知网络的两阶段噪声比例预测算法,达到自适应调用CNN预训练降噪模型获得最佳去噪效果的目的。方法 首先,对大量无噪声图像添加不同噪声比例的RVIN噪声构成噪声图像集合;其次,基于视觉码本（visual codebook）采用软分配（soft-assignment）编码法提取并筛选若干能反映噪声图像受随机脉冲噪声干扰程度的特征值构成特征矢量;再次,将从噪声图像上提取的特征矢量及对应的噪声比例分别作为多层感知网络的输入和输出训练噪声比例预测模型,实现从特征矢量到噪声比例值的映射（预测）;最后,采用粗精相结合的两阶段实现策略进一步提高RVIN噪声比例的预测准确性。结果 针对不同RVIN噪声比例的失真图像,从预测准确性、实际降噪效果和执行效率3个方面验证提出算法的性能和实用性。实验数据表明,本文算法在大多数噪声比例下的预测误差小于2%,降噪效果（PSNR指标）较其他主流降噪算法高24 dB,处理一幅大小为512×512像素的图像仅需3 s左右。结论 本文提出的RVIN噪声比例预测算法在各个噪声比例下具有鲁棒的预测准确性,在降噪效果和执行效率两个方面较经典的开关型RVIN降噪算法有显著提升,更具实用价值。     

基于全卷积神经网络与低秩稀疏分解的显著性检测

为了准确检测复杂背景下的显著区域,提出一种全卷积神经网络与低秩稀疏分解相结合的显著性检测方法,将图像分解为代表背景的低秩矩阵和对应显著区域的稀疏噪声,结合利用全卷积神经网络学习得到的高层语义先验知识,检测图像中的显著区域.首先,对原图像进行超像素聚类,并提取每个超像素的颜色、纹理和边缘特征,据此构成特征矩阵;然后,在MSRA数据库中,基于梯度下降法学习得到特征变换矩阵,利用全卷积神经网络学习得到高层语义先验知识;接着,利用特征变换矩阵和高层语义先验知识矩阵对特征矩阵进行变换;最后,利用鲁棒主成分分析算法对变换后的矩阵进行低秩稀疏分解,并根据分解得到的稀疏噪声计算显著图.在公开数据集上进行实验验证,并与当前流行的方法进行对比,实验结果表明,本文方法能够准确地检测感兴趣区域,是一种有效的自然图像目标检测与分割的预处理方法.     

基于多特征融合卷积神经网络的显著性检测

随着深度学习技术的发展以及卷积神经网络在众多计算机视觉任务中的突出表现,基于卷积神经网络的深度显著性检测方法成为显著性检测领域的主流方法。但是,卷积神经网络受卷积核尺寸的限制,在网络底层只能在较小范围内提取特征,不能很好地检测区域内不显著但全局显著的对象;其次,卷积神经网络通过堆叠卷积层的方式可获得图像的全局信息,但在信息由浅向深传递时,会导致信息遗失,同时堆叠太深也会导致网络难以优化。基于此,提出一种基于多特征融合卷积神经网络的显著性检测方法。使用多个局部特征增强模块和全局上下文建模模块对卷积神经网络进行增强,利用局部特征增强模块增大特征提取范围的同时,采用全局上下文建模获得特征图的全局信息,有效地抑制了区域内显著而全局不显著的物体对显著性检测的干扰; 能够同时提取多尺度局部特征和全局特征进行显著性检测,有效地提升了检测结果的准确性。最后,通过实验对所提方法的有效性进行验证并和其它11种显著性检测方法进行对比,结果表明所提方法能提升显著性检测结果的准确性且优于参与比较的11种方法。     

3D Filtering by Block Matching and Convolutional Neural Network for Image Denoising

Block matching based 3D filtering methods have achieved great success in image denoising tasks. However, the manually set filtering operation could not well describe a good model to transform noisy images to clean images. In this paper, we introduce convolutional neural network (CNN) for the 3D filtering step to learn a well fitted model for denoising. With a trainable model, prior knowledge is utilized for better mapping from noisy images to clean images. This block matching and CNN joint model (BMCNN) could denoise images with different sizes and different noise intensity well, especially images with high noise levels. The experimental results demonstrate that among all competing methods, this method achieves the highest peak signal to noise ratio (PSNR) when denoising images with high noise levels (σ > 40), and the best visual quality when denoising images with all the tested noise levels.     

形状的全尺度可视化表示与识别

目的视觉目标的形状特征表示和识别是图像领域中的重要问题。在实际应用中,视角、形变、遮挡和噪声等干扰因素造成识别精度较低,且大数据场景需要算法具有较高的学习效率。针对这些问题,本文提出一种全尺度可视化形状表示方法。方法在尺度空间的所有尺度上对形状轮廓提取形状的不变量特征,获得形状的全尺度特征。将获得的全部特征紧凑地表示为单幅彩色图像,得到形状特征的可视化表示。将表示形状特征的彩色图像输入双路卷积网络模型,完成形状分类和检索任务。结果通过对原始形状加入旋转、遮挡和噪声等不同干扰的定性实验,验证了本文方法具有旋转和缩放不变性,以及对铰接变换、遮挡和噪声等干扰的鲁棒性。在通用数据集上进行形状分类和形状检索的定量实验,所得准确率在不同数据集上均超过对比算法。在MPEG-7数据集上精度达到99.57%,对比算法的最好结果为98.84%。在铰接和射影变换数据集上皆达到100%的识别精度,而对比算法的最好结果分别为89.75%和95%。结论本文提出的全尺度可视化形状表示方法,通过一幅彩色图像紧凑地表达了全部形状信息。通过卷积模型既学习了轮廓点间的形状特征关系,又学习了不同尺度间的形状特征关系。本文方法...     

Adaptive threshold selection for impulsive noise detection in images using coefficient of variance

This paper proposes an adaptive threshold selection strategy to detect impulsive noise in images. The proposed method utilizes a simple neural network with statistical characteristics of noisy images. The method is adaptive in the sense that the threshold obtained is adaptable to different type of images and noise conditions. The network tuned for one image works for other images as well at different noise conditions. Comparative analysis with other standard techniques reveals that the proposed scheme outperforms its counterparts in terms of noise suppression.     

基于感知掩蔽深度神经网络的单通道语音增强方法

本文将心理声学掩蔽特性应用于基于深度神经网络（Deep neural network,DNN）的单通道语音增强任务中,提出了一种具有感知掩蔽特性的DNN结构.首先,提出的DNN对带噪语音幅度谱特征进行训练并分别得到纯净语音和噪声的幅度谱估计.其次,利用估计的纯净语音幅度谱计算噪声掩蔽阈值.然后,将噪声掩蔽阈值和估计的噪声幅度谱联合计算得到一个感知增益函数.最后,利用感知增益函数从带噪语音幅度谱中估计出增强语音幅度谱.在TIMIT数据库上,对不同信噪比下的20种噪声进行的仿真实验表明,无论噪声类型是否在语音的训练集中出现,所提出的感知掩蔽DNN都能够在有效去除噪声的同时保持较小的语音失真,增强效果明显优于常见的DNN增强方法以及NMF（Nonnegative matrix factorization）增强方法.