基于CNN的改进行人重识别技术

针对行人重识别研究中训练样本的不足,为提高识别精度及泛化能力,提出一种基于卷积神经网络的改进行人重识别方法。首先对训练数据集进行扩充,使用生成对抗网络无监督学习方法生成无标签图像;然后与原数据集联合作半监督卷积神经网络训练,通过构建一个Siamese网络,结合分类模型和验证模型的特点进行训练;最后加入无标签图像类别分布方法,计算交叉熵损失来进行相似度量。实验结果表明,在Market-1501、CUHK03和DukeMTMC-reID数据集上,该方法相比原有的Siamese方法在Rank-1和mAP等性能指标上有近3～5个百分点的提升。当样本较少时,该方法具有一定应用价值。     

基于驾驶员面部时序数据的疲劳驾驶检测算法

针对传统疲劳驾驶检测方法识别准确率低、泛化能力差的问题,提出了一种基于CNNs和LSTM的端到端可训练网络,检测驾驶员的疲劳状态。根据驾驶员面部特征点提取ROI,将在其他计算机视觉任务上表现较好的深度网络迁移到疲劳检测任务中,并结合LSTM处理时序数据的能力,提出一种新的疲劳检测网络,该网络能够读入视频流中的时序数据并检测出驾驶员的疲劳状态。实验证明所提方法和模型在公开数据集中具有较高的识别准确率,并且在不同的数据集间具有很好的泛化能力,对于减少路面车祸、保障人身安全具有很重要的意义。     

基于域泛化的非均衡电力设备分/合闸X射线图像识别

针对电力开关设备分/合闸X射线图像呈现设备种类多、数据量少、类别不平衡问题,提出一种基于域泛化的非均衡电力设备分/合闸X射线图像识别方法.首先以AlexNet和改进的ResNet作为骨干网络设计识别算法;然后利用最大熵对抗数据增强(MEADA)训练算法对训练集进行样本扩充,以模拟未知类型分/合闸图像的域分布;最后通过将Focal Loss损失函数引入到识别算法中解决分/合闸数据类别不平衡问题.实验结果表明,使用所提方法AlexNet和简化ResNet(ResNet-F)模型的平均准确率相比其基线方法分别提升5.31个和6.52个百分点,且ResNet-F的识别精度比AlexNet高出3.54个百分点.类激活图、受试者工作特征(ROC)曲线和t-随机嵌入(t-SNE)等可视化分析结果进一步验证了所提方法的有效性,为多域非均衡电力设备分/合闸X射线图像识别提供了新思路.     

PNET:像素级台标识别网络

台标识别是典型的细微目标识别问题,针对台标区域小、信息量低,且镂空、半透明台标极易受到画面背景影响的难题,提出一个基于端到端全卷积网络的像素级台标识别网络——PNET.首先构建一个像素级标注的台标数据集,通过视频抽帧和图像预处理获得台标图像集,并提出一种逐图像的像素级半自动标注方法获得二值标签图像集;然后提出一个像素级台标识别网络,在典型分类网络AlexNet,VGG的基础上,通过微调,将分类网络在分类任务中学习到的网络参数转换为像素级台标识别网络在台标分割任务中的所需的网络参数;最后引入跨层架构,融合来自网络深层的全局信息和浅层的局部信息.实验结果表明PNET实现了准确的像素级分割,准确率高达98.3%,在NVIDIA Tesla K80上单幅图像识别时间不超过1.5 s.     

反向目标干扰的图像数据增强

混合样本数据增强方法只注重模型对于图像所属类别的正向表达, 而忽略图像是否属于某一类别的反向判定. 为了解决描述图像类别方式单一而影响模型性能的问题, 提出一种反向目标干扰的图像数据增强方法. 该方法增加图像背景及目标的多样性, 防止网络模型过拟合. 其次采用反向学习机制, 让网络模型在正确辨别原图像所属类别的同时, 对填充图像不属于该类别的属性进行充分学习, 从而增强网络模型对原图像所属类别辨识的置信度. 最后, 为验证该方法的有效性, 使用不同的网络模型在CIFAR-10、CIFAR-100等5个数据集上进行大量实验. 实验结果表明, 本文方法与其他先进的数据增强方法相比较, 可以显著提高模型在复杂背景下的学习效果和泛化能力.     

基于图像风格迁移的端到端跨域目标检测

跨域目标检测是最近兴起的研究方向,旨在解决训练集到测试集的泛化问题.在已有的方法中利用图像风格转换并在转换后的数据集上训练模型是一个有效的方法,然而这一方法存在不能端到端训练的问题,效率低,流程繁琐.为此,我们提出一种新的基于图像风格迁移的跨域目标检测算法,可以把图像风格迁移和目标检测结合在一起,进行端到端训练,大大简...     

辅助驾驶的车辆与行人检测模型YOLO-DNF

为了解决YOLO系列目标检测算法存在的精度与计算成本不均衡、模型泛化性不足的问题,提出了可满足不同光照场景下目标检测需求的高精度快速的车辆与行人检测模型YOLO-Day Night and Fast(YOLO-DNF)。文中结合当下主流检测模型所使用的卷积神经网络分析卷积结构与网络深度对于主干网络特征提取能力和计算成本的影响,针对网络不同层次选取卷积结构Arrow-Block与CSP-Block搭建网络并通过量化堆叠单元的计算成本确定网络深度,提出低计算成本、高特征提取能力的ACNet网络。此外分析了白天与夜间图像的亮度差异,引入了HSV域扰动并提出亮度处理的数据增强策略,提升了模型的夜间检测精度,改善了模型泛化性不足的问题。实验结果表明：YOLO-DNF模型在SODA10M数据集仅含白天图像的训练集中训练后以每秒24.36帧的检测速率达到32.8%的全时段mAP检测精度,检测精度与速度超过目前主流检测模型。其中夜间精度达到了27.7%,扩展了模型的检测应用场景。     

深度学习遥感影像近岸舰船识别方法

针对复杂背景近岸舰船检测与细粒度识别难题,提出了一种基于深度学习的新型端到端目标识别框架,可有效检测与识别任意方向的舰船目标。针对舰船目标短边尺度较小问题,提出了角度致密化的预设框设置方法,提高了候选区域生成时的召回率;采用改进方位敏感型区域插值池化,减少了坐标量化误差,实现了舰船局部区域特征的精确建模;利用注意力机制下的全局与局部特征区域级融合方法,提升了区域特征的类别判别能力,解决了细粒度舰船识别难题;针对舰船样本稀缺性问题,使用迁移学习提升了模型性能。构建了一个含有25类近岸舰船目标的细粒度数据集,与传统学习模型相比召回率提高2%,平均识别精度提高3%,对复杂背景下目标识别具有重要实用价值。     

自然场景文本检测与识别的深度学习方法

许多自然场景图像中都包含丰富的文本,它们对于场景理解有着重要的作用。随着移动互联网技术的飞速发展,许多新的应用场景都需要利用这些文本信息,例如招牌识别和自动驾驶等。因此,自然场景文本的分析与处理也越来越成为计算机视觉领域的研究热点之一,该任务主要包括文本检测与识别。传统的文本检测和识别方法依赖于人工设计的特征和规则,且模型设计复杂、效率低、泛化性能差。随着深度学习的发展,自然场景文本检测、自然场景文本识别以及端到端的自然场景文本检测与识别都取得了突破性的进展,其性能和效率都得到了显著提高。本文介绍了该领域相关的研究背景,对基于深度学习的自然场景文本检测、识别以及端到端自然场景文本检测与识别的方法进行整理分类、归纳和总结,阐述了各类方法的基本思想和优缺点。并针对隶属于不同类别下的方法,进一步论述和分析这些主要模型的算法流程、适用场景和技术发展路线。此外,列举说明了部分主流公开数据集,对比了各个模型方法在代表性数据集上的性能情况。最后总结了目前不同场景数据下的自然场景文本检测、识别及端到端自然场景文本检测与识别算法的局限性以及未来的挑战和发展趋势。     

基于深度残差概率随机森林的时间序列分类方法

时间序列数据广泛存在于工业、医疗等应用领域,由于其时序相关性强、特征空间维度大,使得传统的时间序列分类方法普遍存在精度不足和需要复杂特征工程等问题.充分考虑深度神经网络在处理复杂时序数据上的优越性以及决策树方法拟合数据能力强的优势,提出一种基于残差网络和概率决策树的端到端统一深度学习模型.该模型利用残差网络从原始时间序列中提取高级特征,为了更好地建立时序数据特征与类别标签间的映射关系,将概率决策树融入至残差网络的分类层.同时,设计随机子空间的集成策略,缓解由于残差网络的深层结构产生的过度拟合现象,并给出联合优化模型分裂参数和预测参数的迭代优化方案.在大量的基准数据集和工业案例上进行实验和分析,实验结果表明,所提出模型的分类性能优于传统方法与其他深度学习方法,且可有效提高残差网络的泛化能力.     

Finding logos in real-world images with point-context representation-based region search

Finding logos in the real-world images is a challenging task due to their small size, simple shape, less texture and clutter background. In this paper, through visual logo analysis with different types of features, we propose a novel framework for finding visual logos in the real-world images. First, we exploit the contextual shape and patch information around feature points, merge them into a combined feature representation (point-context). Considering the characteristics of logos, this kind of fusion is an effective enhancement for the discriminability of single point features. Second, to eliminate the interference of the complex and noisy background, we transfer the logo recognition to a region-to-image search problem by segmenting real-world images into region trees. A weak geometric constraint based on regions is encoded into an inverted file structure to accelerate the search process. Third, we apply global features to refine initial results in the re-ranking stage. Finally, we combine each region score both in max-response and accumulate-response mode to obtain the final results. Performances of the proposed approach are evaluated on both our CASIA-LOGO dataset and the standard Flickr logos 27 dataset. Experiments and comparisons show that our approach is superior to the state-of-the-art approaches.     

基于可变形部件模型的台标识别方法

背景变化复杂、部分台标相似度高、拉伸变形等因素增加了台标识别的难度,降低了识别的准确率。为此,本文提出了一种鲁棒的基于可变形部件模型的台标识别方法。首先,依据台标特性,利用合适的颜色特征对可变形部件模型的特征进行了改进和增强。其次,利用隐式支持向量机和隐式线性判别分析技术加速台标识别模型训练。然后,为了弥补可变形部件模型的不足,设计了一种基于加权部件的计算方法。最后,提出一种新的可靠机制进行准确率评价。实验结果表明,与基于方向梯度直方图和支持向量机的识别方法相比,该方法具有更高的识别准确率,性能更加稳定。     

特征增强策略驱动的车标识别

目的 小样本情况下的车标识别在实际智能交通系统中具有十分重要的应用价值。针对从实际监控系统中获取的车标图像低分辨率、低质量的特点,考虑如何从车标结构相似性、局部显著特征方面来对车标的整体特征进行增强,提出一种特征增强策略驱动下的车标识别方法（vehicle logo recognition method based on feature enhancement,FE-VLR）。方法 提取车标图像的自对称相似特征,构建图像金字塔,在每层金字塔下提取车标的整体特征和局部显著特征,其中局部显著区域通过基于邻域块相关度的显著区域检测来获取,最后结合CRC （collaborative representation based classification）分类器对车标进行分类识别。结果 在公开车标数据集HFUT-VL （Vehicle Logo Dataset from Hefei University of Technology）和XMU （Xiamen University Vehicle Logo Dataset）上对算法效果进行评估,实验结果表明,在小样本情况下,本文方法优于其他一些传统的车标识别方法,且与一些基于深度学习模型的方法相比,其识别率也有所提升。在HFUT-VL数据集上,当训练样本数为5时,识别率达到97.78%;当训练样本数为20时,识别率为99.1%。在更为复杂的XMU数据集上,本文方法表现出了更好的有效性和更强的鲁棒性,当训练样本在15幅及以下时,本文方法与具有较好表现的OE-POEM （overlapping enhanced patterns of oriented edge magnitudes）算法相比至少提升了7.2%。结论 本文提出的基于特征增强策略的车标识别方法,通过融合自对称相似特征、局部显著特征和车标整体特征来增强特征的表达,提高了对实际道路中的低质量、低分辨率车标图像的识别能力,更能满足实际应用中对车标识别的需求。     

Automatic logo transition detection in digital video contents

The amount of user created contents has been increasing rapidly and is associated with a serious copyright problem. Automatic logo detection and recognition in videos is a natural and efficient way of overcoming the copyright problem. However, logos have varying characteristics, which make logo detection and recognition very difficult. Moreover, logo transitions between two different logos exist in one video comprising several video contents. This disrupts the automatic logo detection and recognition. Therefore, in order to improve logo detection, it is necessary to take into account the logo transitions explicitly. This paper proposes an accurate logo transition detection method for recognizing logos in digital video contents. The proposed method accurately segments a video according to logo and efficiently recognizes various types of logos. The experimental results demonstrate the effectiveness of the proposed method for logo detection and video segmentation according to logo.     

Logo localization and recognition in natural images using homographic class graphs

We propose a method for localization and classification of brand logos in natural images. The system has to overcome multiple challenges such as perspective deformations, warping, variations of the shape and colors, occlusions, background variations. To deal with perspective variation, we rely on homography matching between the SIFT keypoints of logo instances of the same class. To address the changes in color, we construct a weighted graph of logo interconnections that is further analyzed to extract potentially multiple instances of the class. The main instance is built by grouping the keypoints of the graph connected logos onto the central image. The secondary instance is needed for color inverted logos and is obtained by inverting the orientation of the main instance. The constructed logo recognition system is tested on two databases (FlickrLogos-32 and BelgaLogos), outperforming state of the art with more than 10 % accuracy.     

Smart learning of logo detection for mobile phone applications

With the advance of mobile phone cameras and broadband networks, gaining access to digital information and services via logo recognition has become of high industrial interest. The fundamental subsystem for logo recognition must be a logo database, whose images link real-world information to specific corporate entities. However, few attempts have been made to create and update such a logo database, i.e., how to automatically collect the latest logos. Moreover, the few existing methods are limited in their application and unattractive in terms of logo detection accuracy and performance overhead. In this article, we describe a practical system for automatic logo extraction. Websites are an optimal source of a huge number of up-to-date logos, and experts can easily find logos from webpages without rendering. For instance, an expert can locate elements with the term “logo” using the websites’ entity names as attribute values, and then download images connected to them. Our system mimics this human behavior to automate logo extraction. Given a website, it learns its entity name and uses that name to locate elements that lead to the logo. Evaluation tests showed that this contextual reasoning significantly contributes to the performance of the system, which achieved high precision with negligible overhead.     

抗模糊特征提取策略下的车标识别

目的 现有的车标识别方法尽管取得了不错的识别效果,但最终的识别率容易遇到瓶颈,很难得到提升。车标识别是智能交通系统中至关重要的一部分,识别率的微小提升也能带来巨大的社会价值。通过挖掘与分析车标识别中潜在的问题和难点,发现未能得到正确分类的图像大部分为模糊车标图像。针对车标图像中存在的成像模糊等情况,本文提出一种基于抗模糊特征提取的车标识别方法。方法 构建车标图像金字塔模型,分别提取图像的抗纹理模糊特征和抗边缘模糊特征。抗纹理模糊特征的提取使用局部量化的LPQ（local phase quantization）模式,可以增强原始特征的鲁棒性,抗边缘模糊特征的提取基于局部块弱梯度消除的HOG（histogram of oriented gradient）特征提取方法,可以在描述车标图像边缘梯度信息的同时,提升特征的抗模糊能力。最后利用CCA（canonical correlation analysis）方法进行两种抗模糊特征的融合并用于后续的降维与分类。结果 本文方法在多个数据集上均取得了很好的识别效果,在20幅训练样本下,本文方法在公开车标数据集HFUT-VL（vehicle logo dataset from Hefei University of Technology）上取得了99.04%的识别率,在本文构建的模糊车标数据集BVL（blurring vehicle logo dataset）上也取得了97.19%的识别率。而在难度较大的XMU（Xiamen University vehicle logo dataset）上,本文方法在100幅训练样本下也达到了96.87%的识别率,识别效果高于一些具有较好表现的车标识别方法,表现出很强的鲁棒性和抗模糊性。结论 本文方法提高了对成像质量欠缺的车标图像的识别能力,从而提升了整体识别效果,更符合实际应用中车标识别的需求。     

Logo detection with extendibility and discrimination

Logos are specially designed marks that identify goods, services, and organizations using distinguished characters, graphs, signals, and colors. Identifying logos can facilitate scene understanding, intelligent navigation, and object recognition. Although numerous logo recognition methods have been proposed for printed logos, a few methods have been specifically designed for logos in photos. Furthermore, most recognition methods use codebook-based approaches for the logos in photos. A codebook-based method is concerned with the generation of visual words for all the logo models. When new logos are added, the codebook reconstruction is required if effectiveness is a crucial factor. Moreover, logo detection in natural scenes is difficult because of perspective tilt and non-rigid deformation. Therefore, this study develops an extendable, but discriminating, model-based logo detection method. The proposed logo detection method is based on a support vector machine (SVM) using edge-based histograms of oriented gradient (HOGE) as features through multi-scale sliding window scanning. Thereafter, anti-distortion affine scale invariant feature transform (ASIFT) is used for logo verification with constraints on the ASIFT matching pairs and neighbors. The experimental results using the public Flickr-Logo database confirm that the proposed method has a higher retrieval and precision accuracy compared to existing model-based methods.

     

增强边缘梯度特征局部量化策略驱动下的车标识别

目的 车标是车辆的显著性特征,通过车标的分类与识别可以极大缩小车辆型号识别的范围,是车辆品牌和型号识别中的重要环节。基于特征描述子的车标识别算法存在如下缺点：一方面,算法提取的特征数量有限,不能全面描述车标的特征;另一方面,提取的特征过于冗杂,维度高,需要大量的计算时间。为了提取更加丰富的车标特征,提高识别效率,提出一种增强边缘梯度特征局部量化策略驱动下的车标识别方法。方法 首先提取车标图像的增强边缘特征,即根据不同的梯度方向提取梯度信息,生成梯度大小矩阵,并采用LTP（local ternary patterns）算子在梯度大小矩阵上进一步进行特征提取,然后采用特征码本对提取的特征进行量化操作,在确保车标特征描述能力的同时,精简了特征数目,缩短了局部向量的长度,最后采用WPCA（whitened principal component analysis）进行特征降维操作,并基于CRC（collaborative representation based classification）分类器进行车标的识别。结果 基于本文算法提取的车标特征向量,能够很好地描述车标图像的特征,在HFUT-VL1车标数据集上取得了97.85%的识别率（平均每类训练样本为10张）,且在识别难度较大的XMU车标数据集上也能取得90%以上的识别率（平均每类训练样本为100张）,与其他识别算法相比,识别率有明显提高,且具有更强的鲁棒性。结论 增强边缘梯度特征局部量化策略驱动下的车标识别算法提取的特征信息能够有效地描述车标,具有很高的识别率和很强的鲁棒性,大大降低了特征向量的维度,提高了识别效率。     

Intelligent trademark recognition and similarity analysis using a two-stage transfer learning approach

The ubiquitous and fast-growing e-marketplaces are causing serious concerns of unauthorized trademark usages, especially their pictorial images. There is a significant need for e-commerce service intermediaries and trading sites to actively check whether product images posted online are not infringing upon others’ intellectual property rights (IPRs). To ensure registered trademark (TM) or copyrighted logo-image (logo) protection, this research develops an intelligent system that can detect, locate, and crop (cut) logos posted online and check whether they are substantially or deceptively similar to registered TM logos. This research integrates two deep learning models to achieve the research goal. The first is the logo detection and localization model for cropping trademark like images from complex online merchandise photos, which could have images of many product views and even worn and displayed on a human model. Only the cropped logo image is needed for comparison to a database of registered trademarks. The second model performs TM similarity analysis using the cropped logos compared to the published TM images. Yolo v4 is adopted as the general logo locator for intelligent logo image cropping. The triplet convolutional neural network model is used to fine-tuned for intelligent trademark similarity analysis. The models are trained with an image dataset, combining image samples from LogoDet-3k and images found through web search. The system performance is quantitatively tested using the FlickrLogos-32 dataset and consumer product images extracted from an e-commerce platform. The test results of the system achieve high precision (0.91) for trademark class matching.