Scale-Covariant and Scale-Invariant Gaussian Derivative Networks

Journal of Mathematical Imaging and Vision - This paper presents a hybrid approach between scale-space theory and deep learning, where a deep learning architecture is constructed by coupling...     

基于尺度不变特征金字塔的输电线路缺陷检测

针对现有目标检测算法在高压电力复杂巡检场景下电力部件与巡检缺陷检测精度较低的问题,提出一种基于尺度不变特征金字塔的输电线路缺陷检测方法.将主流目标检测方法用于该场景,对比得出RepPoints v2网络模型的检测精度最高.针对RepPoints v2中FPN结构不能有效提取跨层次间语义信息及角点验证过程中忽略尺度归一化...     

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.     

基于深度特征学习的复杂环境下陌生人脸匹配算法研究

复杂环境下的陌生人脸匹配,即在人脸存在光照、姿态干扰时,判断两张在训练集中从未出现过的人脸照片是否代表同一个人。在预处理阶段,采用多尺度视皮层算法,降低光照的影响,提出并采用基于PCA-SIFT特征的图片融合算法无监督地对齐人脸,降低人脸姿态的影响。在识别阶段,提出并采用半随机池化方法优化了局部卷积限制波尔兹曼机网络的稳定性,习得深度特征后采用基于信息熵的度量学习算法计算马氏距离并通过SVM分类识别。实验结果显示,提出的方法在LFW数据集上取得了78%的识别率,相比于采用相同训练模式的经典度量学习方法取得了7%的提高,验证了所提方法的有效性。     

Transfer Learning on Deep Neural Networks to Detect Pornography

While the internet has a lot of positive impact on society, there are negative components. Accessible to everyone through online platforms, pornography is, inducing psychological and health related issues among people of all ages. While a difficult task, detecting pornography can be the important step in determining the porn and adult content in a video. In this paper, an architecture is proposed which yielded high scores for both training and testing. This dataset was produced from 190 videos, yielding more than 19 h of videos. The main sources for the content were from YouTube, movies, torrent, and websites that hosts both pornographic and non-pornographic contents. The videos were from different ethnicities and skin color which ensures the models can detect any kind of video. A VGG16, Inception V3 and Resnet 50 models were initially trained to detect these pornographic images but failed to achieve a high testing accuracy with accuracies of 0.49, 0.49 and 0.78 respectively. Finally, utilizing transfer learning, a convolutional neural network was designed and yielded an accuracy of 0.98.     

Quaternion-based Deep Belief Networks fine-tuning

Deep learning techniques have been paramount in the last years, mainly due to their outstanding results in a number of applications. In this paper, we address the issue of fine-tuning parameters of Deep Belief Networks by means of meta-heuristics in which real-valued decision variables are described by quaternions. Such approaches essentially perform optimization in fitness landscapes that are mapped to a different representation based on hypercomplex numbers that may generate smoother surfaces. We therefore can map the optimization process onto a new space representation that is more suitable to learning parameters. Also, we proposed two approaches based on Harmony Search and quaternions that outperform the state-of-the-art results obtained so far in three public datasets for the reconstruction of binary images.     

Pinning Synchronization of Complex Dynamical Networks on Time Scales

The purpose of this paper is to investigate the synchronization problem of complex dynamical networks on time scales, which includes the synchronization problem of continuous-time and discrete-time complex dynamical networks as special cases. A pinning control strategy is designed to achieve synchronization of complex dynamical networks on time scales. Based on the theory of calculus on time scales and the Lyapunov method, pinning synchronization criteria for complex dynamical networks on time scales are established. Moreover, a numerical example is given to verify the effectiveness of theoretical results.

     

Synchronization of Drive-Response Networks With Delays on Time Scales

<正>Dear Editor,This letter deals with the synchronization in drive-response networks with discrete and distributed delays on time scales. Based on the theory of calculus on time scales, Lyapunov functional method and inequality technique, we obtain new sufficient conditions to ensure synchronization criteria which are dependent on boundedness of grainness but independent of the types of delays. Numerical simulations are given to illustrate the effectiveness of our new results.     

DEM: Deep Entity Matching Across Heterogeneous Information Networks

Heterogeneous information networks, which consist of multi-typed vertices representing objects and multi-typed edges representing relations between objects, are ubiquitous in the real world. In this paper, we study the problem of entity matching for heterogeneous information networks based on distributed network embedding and multi-layer perceptron with a highway network, and we propose a new method named DEM short for Deep Entity Matching. In contrast to the traditional entity matching methods, DEM utilizes the multi-layer perceptron with a highway network to explore the hidden relations to improve the performance of matching. Importantly, we incorporate DEM with the network embedding methodology, enabling highly efficient computing in a vectorized manner. DEM’s generic modeling of both the network structure and the entity attributes enables it to model various heterogeneous information networks flexibly. To illustrate its functionality, we apply the DEM algorithm to two real-world entity matching applications: user linkage under the social network analysis scenario that predicts the same or matched users in different social platforms and record linkage that predicts the same or matched records in different citation networks. Extensive experiments on real-world datasets demonstrate DEM’s effectiveness and rationality.

     

Scale-Invariant Minimum-Cost Curves: Fair and Robust Design Implements

Four functionals for the computation of minimum cost curves are compared. Minimization of these functionals result in the widely studied Minimum Energy Curve (MEC), the recently introduced Minimum Variation Curve (MVC), and their scale-invariant counterparts, (SI-MEC, SI-MVC). We compare the stability and fairness of these curves using a variety of simple interpolation problems. Previously, we have shown MVC to possess superior fairness. In this paper we show that while MVC have fairness and stability superior to MEC they are still not stable in all configurations. We introduce the SI-MVC as a stable alternative to the MVC. Like the MVC, circular and helical arcs are optimal shapes for the SI-MVC. Additionally, the application of scale invariance to functional design allows us to investigate locally optimal curves whose shapes are dictated solely by their topology, free of any external interpolation or arc length constraints.     

Research on Point-wise Gated Deep Networks

Stacking Restricted Boltzmann Machines (RBM) to create deep networks, such as Deep Belief Networks (DBN) and Deep Boltzmann Machines (DBM), has become one of the most important research fields in deep learning. DBM and DBN provide state-of-the-art results in many fields such as image recognition, but they don't show better learning abilities than RBM when dealing with data containing irrelevant patterns. Point-wise Gated Restricted Boltzmann Machines (pgRBM) can effectively find the task-relevant patterns from data containing irrelevant patterns and thus achieve satisfied classification results. For the limitations of the DBN and the DBM in the processing of data containing irrelevant patterns, we introduce the pgRBM into the DBN and the DBM and present Point-wise Gated Deep Belief Networks (pgDBN) and Point-wise Gated Deep Boltzmann Machines (pgDBM). The pgDBN and the pgDBM both utilize the pgRBM instead of the RBM to pre-train the weights connecting the networks' the visible layer and the hidden layer, and apply the pgRBM learning task-relevant data subset for traditional networks. Then, this paper discusses the validity that dropout and weight uncertainty methods are developed to prevent overfitting in pgRBMs, pgDBNs, and pgDBMs networks. Experimental results on MNIST variation datasets show that the pgDBN and the pgDBM are effective deep neural networks learning     

基于深度神经网络的社会媒体网络分析

社会媒体网络中不仅包含了用户、文本、图片和视频等多种模态的数据,还包含了反映不同模态数据之间交互的群体特征。为了更好地描述社会媒体网络,从而为上层应用提供更好的服务,提出了一种基于深度神经网络的社会媒体网络模型。该模型采用深度神经网络对单个模态的数据进行学习,从而得到任意一个模态数据的潜在特征表示方法。对于两种不同模态的数据,利用具有高斯分布的先验矩阵与两个模态数据的后验分布建立反映这两个模态数据间群体特征的生成模型。实验结果表明,提出的模型在网络结构的链接分析中具有更好的预测效果,能有效地描述社会媒体网络的整体特征。     

Deep Shading: Convolutional Neural Networks for Screen Space Shading

In computer vision, convolutional neural networks (CNNs) achieve unprecedented performance for inverse problems where RGB pixel appearance is mapped to attributes such as positions, normals or reflectance. In computer graphics, screen space shading has boosted the quality of real‐time rendering, converting the same kind of attributes of a virtual scene back to appearance, enabling effects like ambient occlusion, indirect light, scattering and many more. In this paper we consider the diagonal problem: synthesizing appearance from given per‐pixel attributes using a CNN. The resulting Deep Shading renders screen space effects at competitive quality and speed while not being programmed by human experts but learned from example images.     

Deep-SBFL:基于频谱的深度神经网络缺陷定位方法

深度神经网络已经在自动驾驶和智能医疗等领域取得了广泛的应用.与传统软件一样,深度神经网络也不可避免地包含缺陷,如果做出错误决定,可能会造成严重后果.因此,深度神经网络的质量保障受到了广泛关注.然而,深度神经网络与传统软件存在较大差异,传统软件质量保障方法无法直接应用于深度神经网络,需要设计有针对性的质量保障方法.软件缺陷定位是保障软件质量的重要方法之一,基于频谱的缺陷定位方法在传统软件的缺陷定位中取得了很好的效果,但无法直接应用于深度神经网络.在传统软件缺陷定位方法的基础上提出了一种基于频谱的深度神经网络缺陷定位方法 Deep-SBFL.该方法首先通过收集深度神经网络的神经元输出信息和预测结果作为频谱信息;然后将频谱信息进行处理作为贡献信息,以用于量化神经元对预测结果所做的贡献;最后提出了针对深度神经网络缺陷定位的怀疑度公式,基于贡献信息计算深度神经网络中神经元的怀疑度并进行排序,以找出最有可能存在缺陷的神经元.为验证该方法的有效性,以E_Inspect@n (结果排序列表前n个位置内成功定位的缺陷数)和EXAM (在找到缺陷元素之前必须检查元素的百分比)作为评测指...     

构建带空洞卷积的深度神经网络重建高分辨率图像

为了在滤波器参数保持不变的情形下扩大感受野,在非常深的卷积网络超分辨率模型网络中引入空洞卷积方法.首先,分析不同膨胀系数组合的空洞卷积块的感受野,并选择更好的结构作为空洞卷积块.然后,堆叠卷积块并加入残差连接构成深度卷积网络.最后,使用多种训练技巧对网络进行重新训练.实验表明,对于数据集Set5上较大的扩大因子,文中构建网络可提升重建效果,并在视觉上也有明显优势.     

针对深度神经网络模型指纹检测的逃避算法

随着深度神经网络在不同领域的成功应用,模型的知识产权保护成为了一个备受关注的问题.由于深度神经网络的训练需要大量计算资源、人力成本和时间成本,攻击者通过窃取目标模型参数,可低成本地构建本地替代模型.为保护模型所有者的知识产权,最近提出的模型指纹比对方法,利用模型决策边界附近的指纹样本及其指纹查验模型是否被窃取,具有不影响模型自身性能的优点.针对这类基于模型指纹的保护策略,提出了一种逃避算法,可以成功绕开这类保护策略,揭示了模型指纹保护的脆弱性.该逃避算法的核心是设计了一个指纹样本检测器——Fingerprint-GAN.利用生成对抗网络(generative adversarial network, GAN)原理,学习正常样本在隐空间的特征表示及其分布,根据指纹样本与正常样本在隐空间中特征表示的差异性,检测到指纹样本,并向目标模型所有者返回有别于预测的标签,使模型所有者的指纹比对方法失效.最后通过CIFAR-10,CIFAR-100数据集评估了逃避算法的性能,实验结果表明：算法对指纹样本的检测率分别可达95%和94%,而模型所有者的指纹比对成功率最高仅为19%,证明了模型指纹比对保护方法的不可靠性.     

DeepTAM: Deep Tracking and Mapping with Convolutional Neural Networks

International Journal of Computer Vision - We present a system for dense keyframe-based camera tracking and depth map estimation that is entirely learned. For tracking, we estimate small pose...     

UpdateSearch: A New Dynamic Channel Allocation Scheme for Mobile Networks That Can Adjust to System Loads

Distributed dynamic channel allocation techniques are an integral part of distributed mobile computing systems where nodes communicate among themselves via wireless radio channels. The channel allocation schemes can be broadly categorized as search based or update based. Search based techniques have low messaging complexity and are suited for high system load and low request rates. On the other hand update based schemes have higher messaging complexity but are more suitable for low system load and high request rates. This paper presents a combined scheme, called UpdateSearch, which provides the advantages of both types of schemes. UpdateSearch is parameterized by the number of channel classes k, 1kn, where n is the total number of channels in the system. The parameter k can be adjusted to control the number of concurrent searches and degree of contention between cells competing for channels in the system. For k=1 and k=n the scheme respectively behaves as basic search and basic update scheme [2]. A simple analytical model is used to compare the performance of UpdateSearch for various values of k with the basic update and search techniques in terms of channel allocation time and number of simultaneous channel selections allowed in the system under different system loading conditions.     

Fine-tuning Deep Belief Networks using Harmony Search

In this paper, we deal with the problem of Deep Belief Networks (DBNs) parameters fine-tuning by means of a fast meta-heuristic approach named Harmony Search (HS). Although such deep learning-based technique has been widely used in the last years, more detailed studies about how to set its parameters may not be observed in the literature. We have shown we can obtain more accurate results comparing HS against with several of its variants, a random search and two variants of the well-known Hyperopt library. The experimental results were carried out in two public datasets considering the task of binary image reconstruction, three DBN learning algorithms and three layers.     

深度神经网络测试研究综述

随着深度神经网络技术的快速发展、大数据的涌现和计算能力的显著提升,深度神经网络被越来越多地应用到各个安全攸关领域,例如自动驾驶、人脸识别、飞机碰撞检测等.传统的软件系统通常由开发人员手工编写代码实现其内部的决策逻辑,并依据相应的测试覆盖准则设计测试用例来测试系统代码.与传统的软件系统不同,深度学习定义了一种新的数据驱动的编程范式,开发人员仅编写代码来规定深度学习系统的网络结构,其内部逻辑则由训练过程获得的神经元连接权值所决定.因此,针对传统软件的测试方法及度量指标无法直接被移植到深度神经网络系统上.近年来,越来越多的研究致力于解决深度神经网络的测试问题,例如提出新的测试评估标准、测试用例生成方法等.调研了92篇相关领域的学术论文,从深度神经网络测试度量指标、测试输入生成、测试预言这3个角度对目前已有的研究成果进行了系统梳理.同时,分析了深度神经网络测试在图像处理、语音处理以及自然语言处理上的已有成果,并介绍了深度神经网络测试中应用到的数据集及工具.最后,对深度神经网络测试的未来工作进行了展望,以期为该领域的研究人员提供参考.