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。     

基于高光谱图像的改进SIFT特征提取与匹配

针对尺度不变特征变换(SIFT)算法所提取图像特征点数量少、误匹率高的问题,提出了一种基于高光谱图像的改进SIFT算法。首先,依据传统SIFT算法中高斯金字塔的构造思想,结合在不同波段下的高光谱图像具有相同宏观特征的特点,首次用高光谱图像作为原始算法中经高斯变换产生的图像,使得检测到的具有实际意义的特征点数量大幅增加;其次,传统SIFT算法以及大量的改进方法都只通过目标象元邻域范围内的像素信息来构造特征描述符,而忽略了像素点的位置信息,文中将目标象元的位置信息纳入了特征描述符,在特征描述符的匹配阶段,在利用邻域范围内的像素信息进行粗匹配之后,利用特征描述符中的位置信息进行精细匹配。仿真实验结果表明在限定最优值与次优值之比的情况下,采用高光谱图像构造高斯金字塔的方式能显著增加特征点的提取数量,更多地挖掘出图像中的极值点;在特征描述符中加入目标象元的位置信息作为特征点匹配第二阶段的判断依据,正确匹配数量达到原方法的59倍以上,极大提升了算法的匹配性能。     

Multiemotional product color design using gray theory and nondominated sorting genetic algorithm-III

When users select products, they consider the emotional experience resulting from the color of the product. However, the emotional demands of users for product color are multidimensional and diverse. It is very important yet difficult to accurately grasp multiemotional image requirements and effectively convert them into design elements. Therefore, multiemotional product color design (MEPCD) has become a very important and challenging research topic. In this article, a novel MEPCD system using gray theory (GT) and nondominated sorting genetic algorithm-III (NSGA-III) is proposed to effectively solve the MEPCD problem. First, the image perception spaces of users, which exist in different emotional dimensions, were collected using factor analysis and the semantic differential technique. Then, GT was used to establish a multidimensional emotional product color image evaluation model. Finally, NSGA-III was used to optimize and design a multiemotional color scheme for a product. Furthermore, according to actual conditions, an MEPCD system was established based on the proposed method. The design case study shows that the method and design system proposed in this article have a certain range of applicability and can effectively improve the practicality of MEPCD.     

Copy-move forgery detection of duplicated objects using accurate PCET moments and morphological operators

This paper presents a new method for copy-move forgery detection of duplicated objects. A bounding rectangle is drawn around the detected object to form a sub-image. Morphological operator is used to remove the unnecessary small objects. Highly accurate polar complex exponential transform moments are used as features for the detected objects. Euclidian distance and correlation coefficient between the feature vectors are calculated and used for searching the similar objects. A set of 20 forged images with duplicated objects is carefully selected from previously published works. Additional 80 non-forged images are edited by the authors and forged by duplicating different kinds of objects. Numerical simulation is performed where the results show that the proposed method successfully detect different kinds of duplicated objects. The proposed method is much faster than the previously existing methods. Also, it exhibits high robustness to various attacks such as additive white Gaussian noise, JPEG compression, rotation, and scaling.     

Development of fan‐shaped tracker for single particle tracking

With the development of super‐resolution fluorescence microscopy, complex dynamic processes in living cells can be observed and recorded with unprecedented temporal and spatial resolution. Single particle tracking (SPT) is the most important step to explore the relationship between the spatio‐temporal dynamics of subcellular molecules and their functions. Although previous studies have developed SPT algorithms to quantitatively analyze particle dynamics in cell, traditional tracking methods have poor performance when dealing with intersecting trajectories. This can be attributed to two main reasons: (a) they do not have point compensation process for overlapping objects; (b) they use inefficient motion prediction models. In this paper, we present a novel fan‐shaped tracker (FsT) algorithm to reconstruct the trajectories of subcellular vesicles in living cells. We proposed a customized point compensation method for overlapping objects based on the fan‐shaped motion trend of the particles. Furthermore, we validated the performance of the FsT in both simulated time‐lapse movies with variable imaging quality and in real vesicle moving images. Meanwhile, we compared the performance of FsT with other five state‐of‐the‐art tracking algorithms by using commonly defined measures. The results showed that our FsT achieves better performance in high signal‐to‐noise ratio conditions and in tracking of overlapping objects. We anticipate that our FsT method will have vast applications in tracking of moving objects in cell.     

内插螺旋立式上行管流场及传热性能的实验研究

通过粒子图像测速流场实验与传热实验相结合，研究了内插螺旋立式上行管的螺旋节距、丝径、中径比等结构参数在不同Re下对流场、阻力及传热性能的影响。结果表明，内插螺旋能够有效扰动和混合管内流体，使管内形成多个纵向旋涡的流体结构、增大管壁附近液体涡量，有利于强化传热。当Re相同时，管内平均流速v、Nu和综合换热性能PEC均随丝径增大而增大，随中径比减小而增大；随节距增大，3种参数均出现增大的趋势，节距大于20 mm后开始减小。管内流体的阻力f随丝径和节距增大而减小，随中径比增大而增大。综合比较，在较低Re时，节距p=20 mm、丝径e=1.6 mm、中径比D/d=0.75时综合传热效果最好。     

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.     

Multi-Purpose Forensics of Image Manipulations Using Residual- Based Feature

The multi-purpose forensics is an important tool for forge image detection. In this paper, we propose a universal feature set for the multi-purpose forensics which is capable of simultaneously identifying several typical image manipulations, including spatial low-pass Gaussian blurring, median filtering, re-sampling, and JPEG compression. To eliminate the influences caused by diverse image contents on the effectiveness and robustness of the feature, a residual group which contains several highpass filtered residuals is introduced. The partial correlation coefficient is exploited from the residual group to purely measure neighborhood correlations in a linear way. Besides that, we also combine autoregressive coefficient and transition probability to form the proposed composite feature which is used to measure how manipulations change the neighborhood relationships in both linear and non-linear way. After a series of dimension reductions, the proposed feature set can accelerate the training and testing for the multipurpose forensics. The proposed feature set is then fed into a multi-classifier to train a multi-purpose detector. Experimental results show that the proposed detector can identify several typical image manipulations, and is superior to the complicated deep CNN-based methods in terms of detection accuracy and time efficiency for JPEG compressed image with low resolution.     

采用轮廓特征匹配的红外-可见光视频自动配准

为了精确地配准近平面场景下的红外-可见光视频序列,本文提出了一种基于轮廓特征匹配的自动配准方法,通过迭代匹配目标轮廓特征来解决异源图像中配准特征的提取和匹配难题。首先,采用运动目标检测技术获取目标轮廓,并由曲率尺度空间(CSS)角点检测算法提取轮廓特征点。此后,建立全局形状上下文描述子和局部边缘方向直方图描述子描述特征,从而实现可靠的特征匹配。来自不同时刻的匹配点对被保存在一个基于高斯距离准则的特征匹配库中。最后,为了克服近平面场景中目标深度变化的影响,本文结合前景样本随机抽样策略计算配准矩阵的损失函数,完成对全局配准矩阵的更新。在LITIV数据库上对方法进行实验验证,结果表明本文方法的配准精度优于当前先进的对比方法,在9个测试视频上的平均重叠率误差仅为0.194,与对比方法相比下降了18.5%。基本满足了近平面场景下红外-可见光视频序列配准的精度要求,且具有较高的鲁棒性。