结合关键点与引导向量的旋转目标检测网络

目的 目标检测是遥感智能解译中重要的研究方向之一,大多数目标检测算法难以实现密集排列的旋转目标的高精度检测。提出了一种基于关键点与引导向量预测的目标检测算法,实现高精度旋转目标检测的同时,还可对目标的朝向进行表征。方法 首先提出了一种新的旋转目标建模方式,将目标检测分解成中心点、头部顶点、引导向量以及目标宽度的参数回归以更贴合检测目标;其次设计旋转椭圆高斯核,能够更好地拟合遥感目标的形状,从而提升关键点的预测精度;最后通过预测中心点指向头部顶点的引导向量,完成同一个目标内中心点与头部顶点的匹配,从而生成一个精准的带方向的旋转矩形检测框。结果 在大长宽比舰船目标的HRSC(high-resolution ship collections)数据集上的实验结果表明,相比于其他主流的目标检测算法,本文算法获得了更好的检测结果,在VOC 2007(visual object classes)和VOC 2012的平均精度分别达到了90.78%和97.85%。在小长宽比飞机目标UCAS-AOD(UCAS-high resolution aerial object detection dataset)数据集上达到了98.81%的平均精度。实验结果表明了本文算法的可行性与有效性。结论 本文算法利用椭圆高斯核计算中心点与头部顶点,并设计引导向量对点匹配关系进行约束,实现了旋转目标的方向检测。     

Error analysis in determining the centroids of circular objects in images

Detecting the size and/or location of circular object(s) in an image(s) has application in many areas, like, flow diagnostics, biomedical engineering, computer vision, etc. The detection accuracy of circular object(s) largely depends on the accuracy of centroiding algorithm and image preprocessing technique. In the present work, an error analysis is performed in determining the centroids of circular objects using synthetic images with eight different signal-to-noise ratios ranging from 2.7 to 17.8. In the first stage, four different centroiding algorithms, namely, Center of Mass, Weighted Center of Mass, Späth algorithm, and Hough transform, are studied and compared. The error analysis shows that Späth algorithm performs better than other algorithms. In the second stage, various image preprocessing techniques, consisting of two filters, namely, Median and Wiener, and five image segmentation methods, namely, Sobel, Prewitt, Laplacian of Gaussian (LoG) edge detector, basic global thresholding, and Otsu’s global thresholding, are compared to determine the centroids of circular objects using Späth algorithm. It is found that Wiener filter plus LoG edge detector performs better than other preprocessing techniques. Real images of a calibration target (typical in flow diagnostics) and the secondary atomization of water droplets are then considered for centroids detection. These two images are preprocessed using Wiener filter plus LoG edge detector and then processed using Späth algorithm to detect the centroids of circular objects. It is observed that the results of real image of the calibration target and synthetic images are comparable. Also, based on visual inspection, the centroids detected in the real image of water droplets are found to be reasonably accurate.

     

图注意力网络的场景图到图像生成模型

目的 目前文本到图像的生成模型仅在具有单个对象的图像数据集上表现良好，当一幅图像涉及多个对象和关系时，生成的图像就会变得混乱。已有的解决方案是将文本描述转换为更能表示图像中场景关系的场景图结构，然后利用场景图生成图像，但是现有的场景图到图像的生成模型最终生成的图像不够清晰，对象细节不足。为此，提出一种基于图注意力网络的场景图到图像的生成模型，生成更高质量的图像。方法 模型由提取场景图特征的图注意力网络、合成场景布局的对象布局网络、将场景布局转换为生成图像的级联细化网络以及提高生成图像质量的鉴别器网络组成。图注意力网络将得到的具有更强表达能力的输出对象特征向量传递给改进的对象布局网络，合成更接近真实标签的场景布局。同时，提出使用特征匹配的方式计算图像损失，使得最终生成图像与真实图像在语义上更加相似。结果 通过在包含多个对象的COCO-Stuff图像数据集中训练模型生成64×64像素的图像，本文模型可以生成包含多个对象和关系的复杂场景图像，且生成图像的Inception Score为7.8左右，与原有的场景图到图像生成模型相比提高了0.5。结论 本文提出的基于图注意力网络的场景图到图像生成模型不仅可以生成包含多个对象和关系的复杂场景图像，而且生成图像质量更高，细节更清晰。     

可见光图像中的高压线缺陷自动诊断方法

研究直升机巡检系统中,基于可见光图像的电力线缺陷诊断方法。对直升机机载系统采集到的可见光图像进行实时分析,自动诊断电力线上可能存在的断股、异物附着缺陷。研究了缺陷在可见光图像中的特征,并提出了一种基于亮度和空间信息的线对象检测方法,自动识别电力线。对线对象进行分析,获取其位置、方向、宽度信息,并进行对象分类。结合缺陷特征和对象与分类信息,设计缺陷诊断流程,诊断断股及异物缺陷。实验证明,该方法对电力线检测准确,能够发现电力线存在的可疑异物和断股缺陷,成功应用于直升机巡检系统中。     

Context and profile based cascade classifier for efficient people detection and safety care system

This study propose a system of extracting and tracking objects for a multimedia system and addresses how to extract the head feature from an object area. It is observed in images taken from real-time records like a video, there is always a variance in human behavior, such as the position, size, etc. of the person being tracked or recorded. This study discusses how to extract and track multiple objects based on context as opposed to a single object. Via cascade extraction, the proposed system allows tracking of more than one human at a time. For this process, an extraction method based on internal and external contexts, which defines features to distinguish a human, is proposed. The proposed method defines shapes of shoulder and head area to recognize the head-shape of a human, and creates an extractor according to its edge information and geometrical shapes context. In this paper, humans in images are extracted and recognized using contexts and profiles. The proposed method is compared with a single face detector system and it shows better performance in terms of precision and speed. This trace information can be applied in safety care system. Extractions can be improved by validating the image using a context based detector when there are duplicated images.     

融合视觉关系检测的电力场景自动危险预警

目的 借助深度学习强大的识别与检测能力,辅助人工进行电力场景下的危险描述与作业预警是一种较为经济和高效的电力安全监管手段。然而,目前主流的以目标检测技术为基础的预警系统只能给出部分危险目标的信息,忽视了电力设备的单目危险关系和成对对象间潜在的二元危险关系。不同于以往的方法,为了拓展危险预警模块的识别能力与功能范畴,本文提出了一种在电力场景下基于视觉关系检测的自动危险预警描述生成方法。方法 对给定的待检测图像,通过目标检测模块得到图中对象的类别名称和限界框位置;分别对图像进行语义特征、视觉特征和空间位置特征的抽取,将融合后的总特征送入关系检测模块,输出单个对象的一元关系和成对对象间的关系三元组;根据检测出的对象类别和关系信息,进行危险预测并给出警示描述。结果 本文自主搜集了多场景下的电力生产作业图像并进行标注,同时进行大量消融实验。实验显示,结合了语义特征、空间特征和视觉特征的关系检测器在前5召回率Recall@5和前10召回率Recall@10上的精度分别达到86.80%和93.93%,比仅使用视觉特征的关系检测器的性能提高约15%。结论 本文提出的融合多模态特征输入的视觉关系检测网络能够较好地给出谓词关系的最佳匹配,并减少不合理的关系预测,且具有一定零样本学习（zero-shot learning）能力。相关可视化结果表明,整体系统能够较好地完成电力场景下的危险预警描述任务。     

基于Hough变换和连通体分析的混合圆形体检测算法

从复杂图像中准确可靠地检测圆形体是计算机视觉和智能化图像理解的关键技术之一. 现存算法存在检测精度低, 对噪声及复杂背景敏感等缺点. 本文提出一种新的混合算法. 首先采用一种改进 Hough 变换获取参数空间并生成其横切面图像; 然后对该横切面图像进行连通体分析, 检测出圆形体的尺寸和中心位置. 改进 Hough 变换重新定义了掩模及积分算子; 连通体分析则采用一种改进圆形测度. 大量实验表明所提算法具有更高检测率、检测精度和鲁棒性.     

基于水动力学、分数阶微分及Steger算法的线状目标提取

目的 线状目标的检测具有非常广泛的应用领域,如车道线、道路及裂缝的检测等,而裂缝是其中最难检测的线状目标。为避免直接提取线状目标时图像分割难的问题,以裂缝和车道线为例,提出了一种新的跟踪线状目标中线的算法。方法 对图像进行高斯平滑,用一种新的分数阶微分模板增强图像中的模糊及微细线状目标;基于Steger算法提出一种提取线状目标中心线特征点的算法,避免了提取整体目标的困难;根据水动力学思想将裂隙看成溪流,通过最大熵阈值处理后,先进行特征点的连接,再基于线段之间的距离及夹角进行线段之间的连接（溪流之间的融合）。结果 对300幅裂缝图像及4种类别的其他线状目标图像进行试验,并与距离变换、最大熵阈值法+细线化Otsu阈值分割+细线化、谷底边界检测等类似算法进行比较分析,本文算法检测出的线状目标的连续性好、漏检（大间隙少）和误检（毛刺及多余线段少）率均较低。结论 本文算法能够在复杂的线状目标图像中准确快速地提取目标的中心线,一定程度上改善了复杂线状目标图像分割难的问题。     

Real-time velocity measurement to linear motion of a rigid object with monocular image sequence analyses

This paper presents a methodology and all procedures used to validate it, which were executed in a physics laboratory under controlled and known conditions. The validation was based on the analyses of registered data in an image sequence and the measurements acquired by high precision sensors. This methodology intended to measure the velocity of a rigid object in linear motion with the use of an image sequence acquired by commercial digital video camera. The proposed methodology does not need a stereo pair of images to calculate the object position in the 3D space: it needs only images sequence acquired for one, only one, angle view (monocular vision). To do so, these objects need to be detected while in movement, which is conducted by the application of a segmentation technique based on the temporal average values of each pixel registered in N consecutive image frames. After detecting and framing these objects, specific points belonging to the object (pixels), on the plane image (2D coordinates or space image), are automatically chosen, which are then transformed into corresponding points in the space object (3D coordinates) by the application of collinearity equations or rational functions (proposed in this work). After obtaining the coordinates of these points in the space object that are registered in the sequence of images, the distance, in meters, covered by the object in a particular time interval may be measured and, consequently, its velocity can be calculated. The system is low cost, use only a computer (architecture Intel I3), and a webcam used to acquire the images (640 × 480, 30 fps). The complexity of the algorithm is linear, fact that allows the system to operate in real time. The results of the analyses are discussed and the advantages and disadvantages of the method are presented.     

Dynamic learning from multiple examples for semantic object segmentation and search

We present a novel “dynamic learning” approach for an intelligent image database system to automatically improve object segmentation and labeling without user intervention, as new examples become available, for object-based indexing. The proposed approach is an extension of our earlier work on “learning by example,” which addressed labeling of similar objects in a set of database images based on a single example. The proposed dynamic learning procedure utilizes multiple example object templates to improve the accuracy of existing object segmentations and labels. Multiple example templates may be images of the same object from different viewing angles, or images of related objects. This paper also introduces a new shape similarity metric called normalized area of symmetric differences (NASD), which has desired properties for use in the proposed “dynamic learning” scheme, and is more robust against boundary noise that results from automatic image segmentation. Performance of the dynamic learning procedures has been demonstrated by experimental results.     

Image Indexing and Retrieval Using Object-based Point Feature Maps

Multimedia data such as audios, images, and videos are semantically richer than standard alphanumeric data. Because of the nature of images as combinations of objects, content-based image retrieval should allow users to query by image objects with finer granularity than a whole image. In this paper, we address a web-based object-based image retrieval (OBIR) system . Its prototype implementation particularly explores image indexing and retrieval using object-based point feature maps. An important contribution of this work is its ability to allow a user to easily incorporate both low- and high-level semantics into an image query. This is accomplished through the inclusion of the spatial distribution of point-based image object features, the spatial distribution of the image objects themselves, and image object class identifiers. We introduce a generic image model, give our ideas on how to represent the low- and high-level semantics of an image object, discuss our notion of image object similarity, and define four types of image queries supported by the OBIR system. We also propose an application of our approach to neurological surgery training.     

基于 ADSP-2181 的 PCB标志定位系统

结合当前电子技术和图像处理技术的最新发展,设计了一种采用ADSP-2181数字信号处理器建构的图像检测定位的硬件平台,以满足图像检测的工业应用需要。运用并行数据流的分析方法系统论证了ADSP-2181作为图像检测系统核心处理器方案,提出了系统的数据流模型和控制模型,并以此设计实现了一个较高性能的硬件系统,具有视频图像获取、实时显示、图像处理和图像检测等多种功能。从PCB自动检测的标志识别及定位的需要出发,提出了适合硬件处理的Hough变换算法进行PCB圆形定位标志的检测,并给出了实验结果。     

Detecting changes in high-resolution satellite coastal imagery using an image object detection approach

This article presents a spatial contrast-enhanced image object-based change detection approach (SICA) to identify changed areas using shape differences between bi-temporal high-resolution satellite images. Each image was segmented and intrinsic image objects were extracted from their hierarchic candidates by the proposed image object detection approach (IODA). Then, the dominant image object (DIO) presentation was labelled from the results of optimal segmentation. Comparing the form and the distribution of bi-temporal DIOs by using the raster overlay function, ground objects were recognized as being spatially changed where the corresponding image objects were detected as merged or split into geometric shapes. The result of typical spectrum-based change detection between two images was enhanced by using changed spatial information of image objects. The result showed that the change detection accuracies of the pixels with both attribute and shape changes were improved from 84% to 94% for the strong attribute pixel, and from 36% to 81% for the weak attribute pixel in study area. The proposed approach worked well on high-resolution satellite coastal images.     

Automated extraction of welds from digitized radiographic images based on mlp neural networks

It is desired to automate inspection of welding flaws. Automated extraction of welds forms the first step in developing an automated weld inspection system. This article presents a multilayered perceptron (MLP) based procedure for extracting welds from digitized radiographic images. The procedure consists of three major components: feature extraction, MLP-based object classification, and postprocessing. For each object in the line image extracted from the whole image, four features are defined: the peak position (x1), the width (x2), the mean square error between the object and its Gaussian intensity plot (x3), and the peak intensity (x4). Fiftyone training samples were used to train MLP neural networks. The training of MLP classifiers is discussed. Trained MLP neural networks are subsequently used to test unlearned feature patterns and to identify whether the patterns are welds or not. Postprocessing is performed to remove noises (misclassified nonweld objects) and restore the continuity of weld line (discontinuity due to missed weld objects). Test results show that the procedure can successfully extract all welds (100%) from 25 radiographic images.     

Preserving Topology by a Digitization Process

The main task of digital image processing is to recognize properties of real objects based on their digital images. These images are obtained by some sampling device, like a CCD camera, and represented as finite sets of points that are assigned some value in a gray-level or color scale. Based on technical properties of sampling devices, these points are usually assumed to form a square grid and are modeled as finite subsets of Z². Therefore, a fundamental question in digital image processing is which features in the digital image correspond, under certain conditions, to properties of the underlying objects. In practical applications this question is mostly answered by visually judging the obtained digital images. In this paper we present a comprehensive answer to this question with respect to topological properties. In particular, we derive conditions relating properties of real objects to the grid size of the sampling device which guarantee that a real object and its digital image are topologically equivalent. These conditions also imply that two digital images of a given object are topologically equivalent. This means, for example, that shifting or rotating an object or the camera cannot lead to topologically different images, i.e., topological properties of obtained digital images are invariant under shifting and rotation.     

Cloud-based approach for automatic CNC workpiece origin localization based on image analysis

Computer vision applications in the industry have been a constant field of research in academia. Industrial daily challenges such as quality inspection, object detection, and measurement are examples of situations where some automation could be done by using computer vision techniques. In this paper, a cloud-based approach of an automatic system based on stereo vision and image analysis has been developed to automate a daily routine present in machining companies: workpiece referencing. The proposed architecture uses two cameras mounted in the spindle of a machining center. All images are processed in custom software, running on the cloud, to return the position of the Workpiece Coordinate System (WCS) directly to the Computer Numerical Control (CNC) machine controller. Experimental results validate the application of the proposed architecture in a real machining process machine.     

利用线画图象二维串的图象理解方法

本文提出一种基于符号运算和面向规则线画图象的自动图象理解方法，方法用代数符号表达空间物体，用两个符号串描述物体在特定的投影空间上的相互关系，上方法所构造的代数系统规定了包括微分，腐蚀等等代数操作定义了若干条运算规则。利用这些操作和规则，可以实现对图象中物体的空间位置，运动趋势，自身大小的变化等等的自动分析。     

Phase congruency-based detection of circular objects applied to analysis of phytoplankton images

Detection and recognition of objects representing the Prorocentrum minimum (P. minimum) species in phytoplankton images is the main objective of the article. The species is known to cause harmful blooms in many estuarine and coastal environments. A new technique, combining phase congruency-based detection of circular objects in images, stochastic optimization-based object contour determination, and SVM- as well as random forest (RF)-based classification of objects was developed to solve the task. A set of various features including a subset of new features computed from phase congruency preprocessed images was used to characterize extracted objects. The developed algorithms were tested using 114 images of 1280×960 pixels. There were 2088 P. minimum cells in the images in total. The algorithms were able to detect 93.25% of objects representing P. minimum cells and correctly classified 94.9% of all detected objects. The feature set used has shown considerable tolerance to out-of-focus distortions. The obtained results are rather encouraging and will be used to develop an automated system for obtaining abundance estimates of the species.     

多特征融合的文档图像版面分析

目的 在文档图像版面分析上,主流的深度学习方法克服了传统方法的缺点,能够同时实现文档版面的区域定位与分类,但大多需要复杂的预处理过程,模型结构复杂。此外,文档图像数据不足的问题导致文档图像版面分析无法在通用的深度学习模型上取得较好的性能。针对上述问题,提出一种多特征融合卷积神经网络的深度学习方法。方法 首先,采用不同大小的卷积核并行对输入图像进行特征提取,接着将卷积后的特征图进行融合,组成特征融合模块;然后选取DeeplabV3中的串并行空间金字塔策略,并添加图像级特征对提取的特征图进一步优化;最后通过双线性插值法对图像进行恢复,完成文档版面目标,即插图、表格、公式的定位与识别任务。结果 本文采用mIOU（mean intersection over union）以及PA（pixel accuracy）两个指标作为评价标准,在ICDAR 2017 POD文档版面目标检测数据集上的实验表明,提出算法在mIOU与PA上分别达到87.26%和98.10%。对比FCN（fully convolutional networks）,提出算法在mIOU与PA上分别提升约14.66%和2.22%,并且提出的特征融合模块对模型在mIOU与PA上分别有1.45%与0.22%的提升。结论 本文算法在一个网络框架下同时实现了文档版面多种目标的定位与识别,在训练上并不需要对图像做复杂的预处理,模型结构简单。实验数据表明本文算法在训练数据较少的情况下能够取得较好的识别效果,优于FCN和DeeplabV3方法。     

利用Ontology改进的自动化图像标注方法

图像检索系统大多是利用图像的底层特征如颜色、纹理和图像来分析图像,没有考虑图像内容及其对象的内容语义,导致对图像的理解不佳.为使系统能更准确的理解图像中的对象及其深层语义,分析了目前图像标注的优缺点,提出了一种以底层特征为基础,利用本体论建构的知识辅助计算机分析图像中实体对象,判断对象与对象间在现实世界中存在的合理相关性,进而对图像进行标注.实验结果显示加入本体论辅助标注图像大大提高了图像识别的准确性.