深度学习识别光网络单元故障的设计与应用

为解决依赖装维上门鉴别光网络单元故障带来的不便,可以从机器视觉入手实现自动化故障识别。近年,ImageNet挑战赛的成功推动了物体识别技术的跨越式发展,特别是基于卷积的深度学习技术在视觉识别方面已经达到人类水平,为光网络单元故障的自动识别提供了技术基础。文章对识别光网络单元的工作状态进行了研究,将设备工作状态分为7个场景,提出了利用手机APP采集图片识别故障的解决方案并投入了实际生产;重点阐述了深度学习模块的设计与实现,提出一种通过算法整合的方式综合运用物体检测和图像分类算法,分3阶段逐步求精,解决了图片过滤,光网络单元型号和状态识别等问题,实现了基于计算机视觉自动识别光网络单元故障。从数据上看产品的端到端准确率超过84%,识别速度达到10 FPS,月均提供服务超过1万人次,在减少用户等待的同时节约了人力资源。     

Kernel-based fourth-order diffusion for image noise removal

The fourth-order partial differential equations have good performance on noise smoothing and edge preservation without creating blocky effects on smooth regions. However, for low signal-to-noise ratio images, the discrimination between edges and noise is a challenging problem. A novel kernel-based fourth-order diffusion is proposed in this paper. It introduces a kernelized gradient operator in the fourth-order diffusion process, which leads to more effective noise removal capability. Experiment results show that this method outperforms several previous anisotropic diffusion methods for noise removal and edge preservation.     

具有光照鲁棒的图像匹配方法

针对现有的基于局部特征的图像匹配算法对光照变化敏感、匹配正确率低等问题，提出一种具有光照鲁棒性的图像匹配算法。首先使用实时对比保留去色（RTCP）算法灰度化图像，然后利用对比拉伸函数模拟不同光照变换对图像的影响从而提取抗光照变换特征点，最后采用局部强度顺序模式建立特征点描述符，根据待匹配图像局部特征点描述符的欧氏距离判断是否为成对匹配点。在公开数据集上，所提算法与尺度不变特征变换（SIFT）算法、加速鲁棒特征（SURF）算法、"风"（KAZE）算法和ORB算法在匹配速度和匹配正确率上进行了对比实验。实验结果表明：随着图像亮度差异的增加，SIFT算法、SURF算法、"风"（KAZE）算法和ORB算法匹配正确率下降迅速，所提算法下降缓慢并且正确率均高于80%；所提算法特征点检测较慢和描述符维数较高，平均耗时为23.47 s，匹配速度不及另外四种算法，但匹配质量却远超过它们。对实时性要求不高的系统中，所提算法可以克服光照变化对图像匹配造成的影响。     

Non-chaotic and chaotic propagation of stationary and dynamic images through MVKS turbulence

Anisoplanatic electromagnetic (EM) propagation across a turbulent atmosphere has been recently examined for an unmodulated carrier propagating over an image-bearing transparency through optical lensing, and for the embedded information inside a carrier recovered using heterodyning and digital demodulation. Carrier modulation yielded better recovery than simple lens-based imaging. A possible mitigation strategy is proposed whereby the image information is encrypted on an RF chaotic carrier, thereafter secondarily embedded onto an optical carrier. Results based on the modified von Karman (MVKS) and the Hufnagel-Valley (H-V) models showed that the signal/image recovery under turbulence is improved compared with non-chaotic propagation. The case of time-varying/dynamic images is also taken up; it is demonstrated via cross-correlation products that turbulence is mitigated by the use of chaotic carrier encryption. Overall, transmission via chaos offers mitigation against distortions due to turbulence along with the security feature inherent via the chaos keys which prevent signal recovery without key-matching.     

Road Extraction from High Resolution Remote Sensing Image based on Shape Analysis with Gestalt Theory

Simulating the psychological experience of human vision，a road extraction model based on the format tower is proposed to extract the road in the high resolution remote sensing image from the perspective of morphology.Firstly，based on the spectral and texture information，the suspected road targets are extracted by using segmentation technology.Then these targets are classified according to their reliability and extract the road targets for each category.Finally，three types of identified road information are verified and merged，and the continuous smooth road extraction results are obtained.Experiments on real high resolution images show that the results are consistent with the visual perception of the human eye，and the overall classification accuracy is higher，indicating that the algorithm is effective and feasible and has good use value.     

基于颜色四通道及空间金字塔的鱼类图像检索

随着计算机视觉技术在海洋水产领域中的应用不断加深，鱼类图像检索在渔业资源调查、鱼类行为学分析等方面发挥了巨大的作用。通过研究发现，鱼类图像的背景信息会对鱼类图像检索造成极大干扰，而且鱼类图像中颜色、纹理、形状等特征由于空间位置信息的缺乏而使检索的准确率不高。为解决以上问题，提出了一种新的基于颜色四通道及空间金字塔的鱼类图像检索算法。首先，提取视觉显著性图将鱼类图像的前景和背景分开，从而减少图像背景对检索的干扰；其次，为了使图像特征包含一定的空间位置信息，利用空间金字塔的理论对图像进行分割，在此基础上，将图像转为HSVG四通道图并提取SURF特征；；最后，得到检索结果。为验证所提算法的有效性，在QUT_fish_data数据集和DLOU_fish_data数据集上对算法的查全率、查准率与经典的HSVG算法和显著性分块算法进行对比：在两个数据集上查准率分别比传统的HSVG算法最多分别提高12%和5%，查全率最多分别提高7%和22%；比传统的显著性分块算法查准率最多分别提高15%和5%，查全率最多分别提高36%和22%；从而证明所提算法是有效的，能有效提升鱼类图像的检索效果。     

Sparse non-convex Lp regularization for cone-beam X-ray luminescence computed tomography

Cone-beam X-ray luminescence computed tomography (CB-XLCT) is an attractive hybrid imaging modality, and it has the potential of monitoring the metabolic processes of nanophosphors-based drugs in vivo. However, the XLCT imaging suffers from a severe ill-posed problem. In this work, a sparse nonconvex L_p (0?p?L₁ regularization. Further, an iteratively reweighted split augmented lagrangian shrinkage algorithm (IRW_SALSA-L_p) was proposed to efficiently solve the non-convex L_p (0?p?p-values (1/16, 1/8, 1/4, 3/8, 1/2, 5/8, 3/4, 7/8) in both 3D digital mouse experiments and in vivo experiments. The results demonstrate that the proposed non-convex methods outperform L₂ and L₁ regularization in accurately recovering sparse targets in CB-XLCT. And among all the non-convex p-values, our L_p(1/4?p?相似文献   

Assessment of local strain field in adhesive layer of an unsymmetrically repaired CFRP panel using digital image correlation

The present study focuses on experimental investigation of through the thickness displacement and strain field in thin adhesive layer in single sided (unsymmetrical) patch repaired CFRP (carbon fiber reinforced polymer) panel under tensile load. Digital image correlation (DIC) technique is employed to acquire the displacement and strain (longitudinal, peel and shear) field. Experimental determination of shear transfer length based on shear strain field obtained from DIC is introduced to estimate the optimum overlap length which is an essential parameter in patch design for the repair of CFRP structures. Further, DIC experiment with magnified optics is performed to get an insight into complex and localized strain field over thin adhesive layer especially at critical zones leading to damage initiation. The failure mechanism, load displacement behavior, damage initiation and propagation are closely monitored using DIC. The influence of patch edge tapering on strain distribution in adhesive layer is also investigated. The DIC successfully captures the global and localized strain field at critical zones over thin adhesive layer and further helps in monitoring the damage based on strain anomalies. Strains are found to have maximum magnitude at the patch overlap edge and the shear strain level in adhesive layer is higher than the peel strain. Normal tapering increases the peel strain and has negligible influence on shear strain level in adhesive layer. The recommended overlap length is found to be consistent with the recommendation in the literature. Whole field strain pattern and the overlap length obtained from experiment are further compared with the finite element analysis results and they appear to be in good coherence.     

Image Super-Resolution Based on Generative Adversarial Networks: A Brief Review

Single image super resolution (SISR) is an important research content in the field of computer vision and image processing. With the rapid development of deep neural networks, different image super-resolution models have emerged. Compared to some traditional SISR methods, deep learning-based methods can complete the superresolution tasks through a single image. In addition, compared with the SISR methods using traditional convolutional neural networks, SISR based on generative adversarial networks (GAN) has achieved the most advanced visual performance. In this review, we first explore the challenges faced by SISR and introduce some common datasets and evaluation metrics. Then, we review the improved network structures and loss functions of GAN-based perceptual SISR. Subsequently, the advantages and disadvantages of different networks are analyzed by multiple comparative experiments. Finally, we summarize the paper and look forward to the future development trends of GAN-based perceptual SISR.     

基于密集神经网络的灰度图像着色算法

针对在灰度图像着色领域中，传统算法信息提取率不高、着色效果不理想的问题，提出了基于密集神经网络的灰度图像着色算法，以实现改善着色效果，让人眼更好地观察图片信息的目的。利用密集神经网络的信息提取高效性，构建并训练了一个端到端的深度学习模型，对图像中的各类信息及特征进行提取。训练网络时与原图像进行对比，以逐渐减小网络输出结果的信息、分类等各类型的损失。训练完成后，只需向网络输入一张灰度图片，即可生成一张颜色饱满、鲜明逼真的彩色图片。实验结果表明，引入密集网络后，可有效改善着色过程中的漏色、细节信息损失、对比度低等问题，所提算法着色效果较基于VGG网络及U-Net、双流网络结构、残差网络（ResNet）等性能优异的先进着色算法而言取得了显著的改进。