基于深度学习的主轴承盖分类识别算法

机械零件自动分类识别算法,在智能工业、自动化加工等领域具有广阔地应用前景.针对汽车发动机主轴承盖零件自动分类时,存在特征多表面分布和光照敏感等难点问题,提出多分支特征融合卷积神经网络(MFF-CNN).MFF-CNN具有2个子网络分支,分别提取主轴承盖2个表面的特征,经过特征融合,形成最终的零件分类特征.在网络结构设计...     

基于FABEMD的变化光照下人脸识别方法

在光照变化的环境下,人脸识别因受到光照强度和方向的非线性干扰而变得困难重重。在人脸局部区域,光照的变化比较缓慢,而皮肤对光照的反射率特征变化比较快,可以认为光照变化是低频信号,而人脸本质特征是高频信号。FABEMD是一种快速自适应的BEMD（Bidimensional Empirical Mode Decomposition,二维经验模式分解）方法,它能够将图像分解为不同尺度的高频图像和低频图像,高频图像代表了人脸皮肤细节纹理特征,而低频图像则代表了轮廓特征。但是并不能定量判别什么样的高频信号以及多少高频信号能够用来消除光照影响,所以提出了两种衡量高频细节信息量的方法,将这些信息量的相对值来推算融合不同尺度的高频信号权重系数。基于Yale B人脸数据库的实验数据证明了所提方法能够取得很好的识别效果。     

自学习规则下的多聚焦图像融合

目的 基于深度学习的多聚焦图像融合方法主要是利用卷积神经网络（convolutional neural network,CNN）将像素分类为聚焦与散焦。监督学习过程常使用人造数据集,标签数据的精确度直接影响了分类精确度,从而影响后续手工设计融合规则的准确度与全聚焦图像的融合效果。为了使融合网络可以自适应地调整融合规则,提出了一种基于自学习融合规则的多聚焦图像融合算法。方法 采用自编码网络架构,提取特征,同时学习融合规则和重构规则,以实现无监督的端到端融合网络;将多聚焦图像的初始决策图作为先验输入,学习图像丰富的细节信息;在损失函数中加入局部策略,包含结构相似度（structural similarity index measure,SSIM）和均方误差（mean squared error,MSE）,以确保更加准确地还原图像。结果 在Lytro等公开数据集上从主观和客观角度对本文模型进行评价,以验证融合算法设计的合理性。从主观评价来看,模型不仅可以较好地融合聚焦区域,有效避免融合图像中出现伪影,而且能够保留足够的细节信息,视觉效果自然清晰;从客观评价来看,通过将模型融合的图像与其他主流多聚焦图像融合算法的融合图像进行量化比较,在熵、Q_w、相关系数和视觉信息保真度上的平均精度均为最优,分别为7.457 4,0.917 7,0.978 8和0.890 8。结论 提出了一种用于多聚焦图像的融合算法,不仅能够对融合规则进行自学习、调整,并且融合图像效果可与现有方法媲美,有助于进一步理解基于深度学习的多聚焦图像融合机制。     

基于深度网络的多形态人脸识别

在实际的自动人脸识别系统中,输入的识别图像往往在表情、分辨率大小以及姿态方面呈现出多种变化。现在很多方法尝试通过线性或局部线性的映射来寻找由这些变化共享的统一的特征空间。利用由受限玻尔兹曼机(RBM)堆叠成的深度神经网络来发掘这些变化内在的非线性表达。深度网络能够学习高维数据到低维数据的映射关系,并且有助于提高图像分类和识别的性能。同时,为了实现在一个统一的深度框架下同时进行特征提取和识别,在网络的顶层增加了一个监督的回归层。在预训练阶段,通过训练集中不同姿态、不同表情以及不同分辨率的图像对网络进行初始化。在微调阶段,通过网络的输出与标签之间的差 并利用标准反向传播的方法 对模型的参数空间进行调整。在测试阶段,从测试库中随机选择一幅图像,获得统一空间下的特征向量。通过与参考图像库中的所有特征向量进行对比,利用最近邻域的方法识别人脸身份。在具有丰富表情以及大姿态变化的CMU-PIE人脸数据库上进行了全面的实验,结果表明,提出的方法取得了比最新的局域线性映射(或局部线性)的人脸识别方法更高的识别率。     

Segmentation of MR and CT Images by Using a Quantiser Neural Network

A Quantiser Neural Network (QNN) is proposed for the segmentation of MR and CT images. Elements of a feature vector are formed by image intensities at one neighbourhood of the pixel of interest. QNN is a novel neural network structure, which is trained by genetic algorithms. Each node in the first layer of the QNN forms a hyperplane (HP) in the input space. There is a constraint on the HPs in a QNN. The HP is represented by only one parameter in d-dimensional input space. Genetic algorithms are used to find the optimum values of the parameters which represent these nodes. The novel neural network is comparatively examined with a multilayer perceptron and a Kohonen network for the segmentation of MR and CT head images. It is observed that the QNN gives the best classification performance with fewer nodes after a short training time.     

SmallWorld Model-Based Polylogarithmic Routing Using Mobile Nodes

The use of mobile nodes to improve network system performance has drawn considerable attention recently. The movement-assisted model considers mobility as a desirable feature,where routing is based on the store-carry-forward paradigm with random or controlled movement of resource rich mobile nodes.The application of such a model has been used in several emerging networks,including mobile ad hoc networks(MANETs),wireless sensor networks(WSNs),and delay tolerant networks(DTNs).It is well known that mobility increases the capacity of MANETs by reducing the number of relays for routing,prolonging the lifespan of WSNs by using mobile nodes in place of bottleneck static sensors,and ensuring network connectivity in DTNs using mobile nodes to connect different parts of a disconnected network.Trajectory planning and the coordination of mobile nodes are two important design issues aiming to optimize or balance several measures, including delay,average number of relays,and moving distance.In this paper,we propose a new controlled mobility model with an expected polylogarithmic number of relays to achieve a good balance among several contradictory goals,including delay,the number of relays,and moving distance.The model is based on the small-world model where each static node has"short"link connections to its nearest neighbors and"long"link connections to other nodes following a certain probability distribution.Short links are regular wireless connections whereas long links are implemented using mobile nodes.Various issues are considered,including trade-offs between delay and average number of relays,selection of the number of mobile nodes,and selection of the number of long links.The effectiveness of the proposed model is evaluated analytically as well as through simulation.     

多尺度特征复用混合注意力网络的图像重建

目的 针对以往基于深度学习的图像超分辨率重建方法单纯加深网络、上采样信息损失和高频信息重建困难等问题,提出一种基于多尺度特征复用混合注意力网络模型用于图像的超分辨率重建。方法 网络主要由预处理模块、多尺度特征复用混合注意力模块、上采样模块、补偿重建模块和重建模块5部分组成。第1部分是预处理模块,该模块使用一个卷积层来提取浅层特征和扩张特征图的通道数。第2部分是多尺度特征复用混合注意力模块,该模块加入了多路网路、混合注意力机制和长短跳连接,以此来进一步扩大特征图的感受野、提高多尺度特征的复用和加强高频信息的重建。第3部分是上采样模块,该模块使用亚像素方法将特征图上采样到目标图像尺寸。第4部分是补偿重建模块,该模块由卷积层和混合注意力机制组成,用来对经过上采样的特征图进行特征补偿和稳定模型训练。第5部分是重建模块,该模块由一个卷积层组成,用来将特征图的通道数恢复至原来数量,以此得到重建后的高分辨率图像。结果 在同等规模模型的比较中,以峰值信噪比（peak signal-to-noise ratio,PSNR）和结构相似度（structural similarity index measure,SSIM）作为评价指标来评价算法性能,在Set5、Set14、BSD100（Berkeley segmentation dataset）和Urban100的基准测试集上进行测试。当缩放尺度因子为3时,各测试集上的PSNR/SSIM依次为34.40 dB/0.927 3,30.35 dB/0.842 7,29.11 dB/0.805 2和28.23 dB/0.854 0,相比其他模型有一定提升。结论 量化和视觉的实验结果表明,本文模型重建得到的高分辨率图像不仅在重建边缘和纹理信息有很好的改善,而且在PSNR和SSIM客观评价指标上也有一定的提高。     

类别信息生成式对抗网络的单图超分辨重建

目的 基于生成式对抗网络的超分辨模型（SRGAN）以感知损失函数作为优化目标,有效解决了传统基于均方误差（MSE）的损失函数导致重建图像模糊的问题。但是SRGAN的感知损失函数中并未添加明确指示模型生成对应特征的标志性信息,使得其无法精准地将数据的具体维度与语义特征对应起来,受此局限性影响,模型对于生成图像的特征信息表示不足,导致重建结果特征不明显,给后续识别处理过程带来困难。针对上述问题,在SRGAN方法的基础上,提出一种类别信息生成式对抗网络的超分辨模型（class-info SRGAN）。方法 对SRGAN模型增设类别分类器,并将类别损失项添加至生成网络损失中,再利用反向传播训练更新网络参数权重,以达到为模型提供特征类别信息的目的,最终生成具有可识别特征的重建图像。创新及优势在于将特征类别信息引入损失函数,改进了超分辨模型的优化目标,使得重建结果的特征表示更加突出。结果 经CelebA数据集测试表明：添加性别分类器的class-info SRGAN的生成图像性别特征识别率整体偏高（58%97%）;添加眼镜分类器的class-info SRGAN的生成图像眼镜框架更加清晰。此外,模型在Fashion-mnist与Cifar-10数据集上的结果同样表明其相较于SRGAN的重建质量更佳。结论 实验结果验证了本方法在超分辨重建任务中的优势和有效性,同时结果显示：虽然class-info SRGAN更适用于具有简单、具体属性特征的图像,但总体而言仍是一种效果显著的超分辨模型。     

Development of an image-based model for capillary vasculature of retina

The paper presents a method of development of a detailed network model to represent retinal capillary vasculature. The capillary model is a circular mesh consisting of concentric rings with an increasing diameter. Each of the rings has uniformly distributed bifurcation nodes to represent capillary vessels. The model is customized using the data that has been measured from confocal microscopic images of a mouse retina. The capillary model developed can be connected to networks of larger vessels of the vasculature such as arterial and venous networks to form a complete model of the retinal network. A method to automate such interface connections between capillary and other vascular networks using connecting vessels (i.e., pre-capillary and post-capillary) is also presented in the paper. Such a detailed image-based capillary model together with the arterial and venular networks can be used for various circulation simulations to obtain accurate information on hemodynamic quantities such as the spatial distribution of pressure and flow in the vasculature for both physiological and pathological conditions. The method presented for the development of the capillary model can also be adopted for vasculatures of other organs.     

基于MATLAB的BP神经网络在猪等级评定中的应用研究*

为了客观评估猪肉各项指标和猪肉等级,采用MATLAB神经网络工具箱中的BP人工神经网络,利用猪胴体图像特征参数和活体猪图像特征参数建立BP神经网络模型。分别用猪胴体图像特征参数样本60个和活体猪图像特征参数样本80个进行了网络训练,并采用不同的BP神经网络隐含层的传递函数和隐含层神经元数量,得到 BP神经网络模型。通过仿真,将仿真结果与人工评估结果进行对比,结果表明BP人工神经网络模型可以评估猪肉各项指标和等级识别。在猪肉胴体图像特征指标下评价猪肉等级准确率达到98%,在活体猪图像特征参数评价猪肉等级准确率达到80%。说明猪肉胴体图像特征比活体猪图像特征参数更能代表猪肉质量品质也符合客观现实;同时也表明MATLAB神经网络工具箱中的BP人工神经网络可以应用在猪的等级评定中。     

残差密集空间金字塔网络的城市遥感图像分割

目的 遥感图像语义分割是根据土地覆盖类型对图像中每个像素进行分类,是遥感图像处理领域的一个重要研究方向。由于遥感图像包含的地物尺度差别大、地物边界复杂等原因,准确提取遥感图像特征具有一定难度,使得精确分割遥感图像比较困难。卷积神经网络因其自主分层提取图像特征的特点逐步成为图像处理领域的主流算法,本文将基于残差密集空间金字塔的卷积神经网络应用于城市地区遥感图像分割,以提升高分辨率城市地区遥感影像语义分割的精度。方法 模型将带孔卷积引入残差网络,代替网络中的下采样操作,在扩大特征图感受野的同时能够保持特征图尺寸不变;模型基于密集连接机制级联空间金字塔结构各分支,每个分支的输出都有更加密集的感受野信息;模型利用跳线连接跨层融合网络特征,结合网络中的高层语义特征和低层纹理特征恢复空间信息。结果 基于ISPRS （International Society for Photogrammetry and Remote Sensing） Vaihingen地区遥感数据集展开充分的实验研究,实验结果表明,本文模型在6种不同的地物分类上的平均交并比和平均F₁值分别达到69.88%和81.39%,性能在数学指标和视觉效果上均优于SegNet、pix2pix、Res-shuffling-Net以及SDFCN （symmetrical dense-shortcut fully convolutional network）算法。结论 将密集连接改进空间金字塔池化网络应用于高分辨率遥感图像语义分割,该模型利用了遥感图像不同尺度下的特征、高层语义信息和低层纹理信息,有效提升了城市地区遥感图像分割精度。     

Handling uncertainties of robot manipulators and active vision by constraint propagation

Joint errors are inevitable in robot manipulation. These uncertainties propagate to give rise to translational and orientational errors in the position and orientation of the robot end‐effector. The displacement of the active vision head mounted on the robot end‐effector produces distortion of the projected object on the image. Upon active visual inspection, the observed dimension of a mechanical part is given dimension by the measurement on the projected edge segment on the image. The difference between the observed dimension and the actual dimension is the displacement error in active vision. For different motion of the active vision head, the resulting displacement errors are different. Given the uncertainties of the robot manipulator's joint errors, constraint propagation can be employed to assign the motion of the active sensor in order to satisfy the tolerance of the displacement errors for inspection. In this article, we define the constraint consistency and network satisfaction in the constraint network for the problem of displacement errors in active vision. A constraint network is a network where the nodes represent variables, or constraints, and the arcs represent the relationships between the output variables and the input variables of the constraints. In the displacement errors problem, the tolerance of the displacement errors and the translational and orientational errors of robot manipulators have interval values while the sensor motion has real values. Constraint propagation is developed to propagate the tolerance of displacement errors in the hierarchical interval constraint network in order to find the feasible robot motion. © 2002 Wiley Periodicals, Inc.     

On the sample complexity of learning for networks of spiking neurons with nonlinear synaptic interactions

We study networks of spiking neurons that use the timing of pulses to encode information. Nonlinear interactions model the spatial groupings of synapses on the neural dendrites and describe the computations performed at local branches. Within a theoretical framework of learning we analyze the question of how many training examples these networks must receive to be able to generalize well. Bounds for this sample complexity of learning can be obtained in terms of a combinatorial parameter known as the pseudodimension. This dimension characterizes the computational richness of a neural network and is given in terms of the number of network parameters. Two types of feedforward architectures are considered: constant-depth networks and networks of unconstrained depth. We derive asymptotically tight bounds for each of these network types. Constant depth networks are shown to have an almost linear pseudodimension, whereas the pseudodimension of general networks is quadratic. Networks of spiking neurons that use temporal coding are becoming increasingly more important in practical tasks such as computer vision, speech recognition, and motor control. The question of how well these networks generalize from a given set of training examples is a central issue for their successful application as adaptive systems. The results show that, although coding and computation in these networks is quite different and in many cases more powerful, their generalization capabilities are at least as good as those of traditional neural network models.     

基于共生矩阵分析的自适应神经网络图像复原算法

为了保护图像中的细节信息,提出了一种基于共生矩阵聚类分析的自适应Hopfield神经网络图像复原算法．通过计算图像局部区域的共生矩阵提取其纹理特征,对共生矩阵非零元素进行聚类分析．根据聚类数量和各聚类之间的距离,提出了图像局部区域细节强度的定义及其计算方法．细节强度在准确地区分图像的平坦区域和细节区域基础上,通过非线性函数自适应地调整Hopfield网络的权系数矩阵,以使权系数适合图像的纹理特征,而且权系数的生成过程符合人的视觉特性．图像复原的迭代求解过程和神经网络权系数矩阵的更新过程交替进行．该算法能够在图像的平坦区域有效地抑制噪声,在包含细节的区域突出细节．对比实验结果显示,该算法获得的复原图像的信噪比明显提高,视觉效果明显改善．     

Anisotropic diffusion based denoising on X-radiography images to detect weld defects

This paper proposes a machine vision scheme for denoising, feature space gradient preserving, and detecting weld defects in noisy weld X-radiography images; particularly, for the images that are in low-contrast and contain noises. The detection of small weld defects present on noisy image is extremely difficult in non-destructive testing through machine vision. The presence of high gradient magnitude and the low intensity in the feature space of a noisy image are the main characteristics of weld defects. These characteristics can be considered to refine and obtain noise-free images for detection of weld defects. This study proposes a modified anisotropic diffusion model, which considers a local probability value of gray-level and an adaptive threshold parameter in diffusion coefficient function to adjust the implication of low edge gradient of the feature space from the noisy image. Furthermore, an entropy based stopping criterion has been introduced to terminate the diffusion process. This proposed model is compared with the existing models, and its performance is evaluated through Mean Square Error (MSE), Signal-to-Noise Ratio (SNR), Peak Signal-to-Noise Ratio (PSNR), Entropy (E) and Mean Structural Similarity (MSSIM) measures. Experimental results confirm the reliability of the proposed model.     

多尺度密集残差网络的单幅图像超分辨率重建

目的 近几年应用在单幅图像超分辨率重建上的深度学习算法都是使用单种尺度的卷积核提取低分辨率图像的特征信息,这样很容易造成细节信息的遗漏。另外,为了获得更好的图像超分辨率重建效果,网络模型也不断被加深,伴随而来的梯度消失问题会使得训练时间延长,难度加大。针对当前存在的超分辨率重建中的问题,本文结合GoogleNet思想、残差网络思想和密集型卷积网络思想,提出一种多尺度密集残差网络模型。方法 本文使用3种不同尺度卷积核对输入的低分辨率图像进行卷积处理,采集不同卷积核下的底层特征,这样可以较多地提取低分辨率图像中的细节信息,有利于图像恢复。再将采集的特征信息输入残差块中,每个残差块都包含了多个由卷积层和激活层构成的特征提取单元。另外,每个特征提取单元的输出都会通过短路径连接到下一个特征提取单元。短路径连接可以有效地缓解梯度消失现象,加强特征传播,促进特征再利用。接下来,融合3种卷积核提取的特征信息,经过降维处理后与3×3像素的卷积核提取的特征信息相加形成全局残差学习。最后经过重建层,得到清晰的高分辨率图像。整个训练过程中,一幅输入的低分辨率图像对应着一幅高分辨率图像标签,这种端到端的学习方法使得训练更加迅速。结果 本文使用两个客观评价标准PSNR（peak signal-to-noise ratio）和SSIM（structural similarity index）对实验的效果图进行测试,并与其他主流的方法进行对比。最终的结果显示,本文算法在Set5等多个测试数据集中的表现相比于插值法和SRCNN算法,在放大3倍时效果提升约3.4 dB和1.1 dB,在放大4倍时提升约3.5 dB和1.4 dB。结论 实验数据以及效果图证明本文算法能够较好地恢复低分辨率图像的边缘和纹理信息。     

基于神经网络的异构网络向量化表示方法

当网络中存在不同类型的对象时,对象与对象之间的关系会变得多种多样,网络的结构也会变得更为复杂。针对网络的异构化问题,提出了一种基于神经网络的异构网络向量化表示方法。针对具有图片和文本两种类型对象的异构网络,采用多层次的卷积网络将图片映射到一个潜在的特征空间,采用全连接的神经网络将文本对象也映射到相同的特征空间。在该特征空间内,图片与图片、文本与文本以及图片和文本之间的相似性采用相同的距离计算方法。在实验中,应用提出的方法进行异构网络的多种应用测试,结果表明提出的方法是有效的。     

Knowledge-based image retrieval with spatial and temporalconstructs

A knowledge-based approach to retrieve medical images by feature and content with spatial and temporal constructs is developed. Selected objects of interest in an image are segmented and contours are generated. Features and content are extracted and stored in a database. Knowledge about image features can be expressed as a type abstraction hierarchy (TAH), the high-level nodes of which represent the most general concepts. Traversing TAH nodes allows approximate matching by feature and content if an exact match is not available. TAHs can be generated automatically by clustering algorithms based on feature values in the databases and hence are scalable to large collections of image features. Since TAHs are generated based on user classes and applications, they are context- and user-sensitive. A knowledge-based semantic image model is proposed to represent the various aspects of an image object's characteristics. The model provides a mechanism for accessing and processing spatial, evolutionary and temporal queries. A knowledge-based spatial temporal query language (KSTL) has been developed that extends ODMG's OQL and supports approximate matching of features and content, conceptual terms and temporal logic predicates. Further, a visual query language has been developed that accepts point-click-and-drag visual iconic input on the screen that is then translated into KSTL. User models are introduced to provide default parameter values for specifying query conditions. We have implemented the KMeD (Knowledge-based Medical Database) system using these concepts     

ForegroundNet：一种基于语义与动态特征的前景检测算法

针对以往的前景检测方法对场景信息依赖较多的问题,提出了一种实时的无需迭 代更新背景模型的前景检测深度学习模型ForegroundNet。ForegroundNet 首先通过骨干网络从 当前图像和辅助图像中提取语义特征,辅助图像为相邻的图像帧或者是自动生成的视频背景图 像;然后将提取得到的特征输入到包含短连接的反卷积网络中,使得最终特征图在与输入图像 具有相同的大小,并且包含不同尺度的语义及动态特征;最后使用softmax 层进行二值分类, 得到最终检测结果。在CDNet 数据集上进行的实验结果表明,相比于当前F 值为0.82 的次优 方法,ForegroundNet 能够获得0.94 的F 值,具有更高的检测精度;同时ForegroundNet 检测速 度达到123 fps,具有良好的实时性。