Quantigraphic Imaging Estimating the Camera Response and Exposures from Differently Exposed
分形几何在计算机图像识别中的应用研究
分形几何在计算机图像识别中的应用研究摘要：计算机图像识别在现代科技发展中具有广泛的应用前景。
分形几何作为一种数学表达方法，具有自相似和无限细节的特性，可以为计算机图像识别提供独特的解决方案。
本文将探讨分形几何在计算机图像识别中的应用研究，包括图像压缩、图像特征提取和图像分类等方面，并分析其优势和挑战。
1. 引言计算机图像识别是一门研究如何让计算机模拟人类视觉系统进行图像分析、识别和理解的科学和技术。
它在计算机视觉、模式识别、人工智能等领域具有极为重要的应用前景。
然而，由于图像数据的复杂性和特异性，如何提高计算机图像识别的准确性和效率一直是研究的难点。
分形几何作为一种新颖的数学工具，被引入到计算机图像识别中，为提高图像识别的准确性和效率提供了新的可能性。
2. 分形几何概述分形几何是1980年代起兴起的一门科学，它以自相似和无限细节为基本特征。
通过简单的几何构造规则可以生成复杂的图案，并能够在各种尺度上保持相似性。
分形几何广泛应用于自然科学、社会科学、艺术等领域。
在计算机图像识别中，分形几何的应用主要体现在图像压缩、图像特征提取和图像分类等方面。
3. 分形几何在图像压缩中的应用图像压缩是计算机图像处理中的重要环节，旨在通过减少图像的数据量，以降低存储空间和传输带宽的需求。
传统的图像压缩方法如JPEG、GIF等，使用基于变换编码和预测编码的算法。
然而，分形压缩是一种基于分形几何理论的新型压缩方法，它通过把图像分割成多个小块，利用自相似特性在小块之间建立映射关系，从而实现高效的压缩效果。
分形压缩具有较好的失真控制性能和高压缩比，适用于图像存储和传输等多个应用场景。
4. 分形几何在图像特征提取中的应用图像特征提取是计算机图像识别中的关键环节，它通过从图像中挖掘出具有判别性的特征，从而实现图像分类或目标检测等任务。
传统的特征提取方法如边缘检测、纹理分析等，往往需要对图像进行前期的预处理和人工选择。
而基于分形几何的特征提取方法可以通过计算图像的分形维度、分形函数等数学特征，从而提取出图像中的自相似和复杂结构等特征。
基于剪切波弹性成像及超微血管成像影像学参数构建乳腺癌新辅助化疗疗效预测模型的预测效能
乳腺癌是女性最常见的恶性肿瘤，约占每年新发女性恶性肿瘤的30%，也是女性癌症死亡的第二大原因［1-2］。
其化疗模式有术前新辅助化疗、术后辅助化疗、转移和复发的化疗，新辅助化疗可降低乳腺癌的分期，如不可手术者降期为可手术，不保乳者降为可保乳。
通过肿瘤变化，可了解其对方案是否敏感，以便判断预后及术后是否行加强治疗［3］。
早期评价新辅助化疗疗效对乳腺癌患者具有重要价值。
超声剪切波弹性成像（shear wave elastography ，SWE ）是DOI ：10.3969/j.issn.1672-0512.2024.03.016［通信作者］ 田妍，Email ：***************。
基于剪切波弹性成像及超微血管成像影像学参数构建乳腺癌新辅助化疗疗效预测模型的预测效能田 妍四川省自贡市第一人民医院超声医学科，四川 自贡 643000［摘要］ 目的：探讨基于剪切波弹性成像（SWE ）及超微血管成像（SMI ）影像学参数构建的乳腺癌新辅助化疗疗效预测模型的预测效能。
方法：收集160例乳腺癌患者，其中训练集120例，验证集40例。
根据化疗效果分为完全缓解（pCR ）组和非pCR 组。
收集患者相关资料，采用单因素和多因素二元logistic 回归分析筛选乳腺癌患者新辅助化疗后非pCR 的独立影响因素，通过RStudio 4.2.1构建列线图风险模型并验证。
结果：单因素分析显示，2组雌激素受体（ER ）、孕激素受体（PR ）、人类表皮生长因子受体-2（HER -2）、内部回声、钙化、边缘毛刺、最大径、动脉收缩期血流峰值速度（PSV ）、RI 、双侧横向剪切波速度比值（SWVr ）及Adler 血流分级差异均有统计学意义（均P <0.05）。
多因素分析：临床特征分析显示，ER 、PR 、HER -2均是非pCR 的独立影响因素（均P <0.05）；影像学特征分析显示，PSV 、SWVr 均是非pCR 的独立影响因素（均P <0.05）；联合分析显示，ER 、PR 、HER -2、PSV 、SWVr 均是非pCR 的独立影响因素（均P <0.05）。
宏观傅里叶叠层技术远距离成像实验研究
航天返回与遥感第44卷第6期38 SPACECRAFT RECOVERY & REMOTE SENSING2023年12月宏观傅里叶叠层技术远距离成像实验研究田芷铭赵明王森李剑（大连海事大学，大连116026）摘要傅里叶叠层是一新型的宽视场高分辨成像技术，但是其在宏观成像领域的应用中，成像模型在米级成像距离下通常仅有2 cm左右的成像视场，难以满足使用要求。
为了提高宏观傅里叶叠层技术的成像距离和视场，文章开展了远距离宏观反射式傅里叶叠层成像模型的理论研究，提出了一种新的宏观傅里叶叠层成像模型，该模型使用发散光束照明，通过球面波移位对目标傅里叶谱进行扫描重建高分辨率目标图像；此外，还分析了宏观相干成像机理和傅里叶成像模型近似条件，由此推导出模型的近似范围，为模型推广提供了理论基础；最后，利用搭建的实验系统对10 m外目标成像，使目标分辨率从1.4 mm提升到0.35 mm，分辨率提升4倍以上，验证了模型具有通过合成孔径技术提升目标成像分辨率的能力。
关键词宏观成像傅里叶叠层成像模型远距离成像超分辨技术傅里叶叠层实验中图分类号: TP391.41文献标志码: A 文章编号: 1009-8518(2023)06-0038-07 DOI: 10.3969/j.issn.1009-8518.2023.06.004Experimental Research on Long-Range Imaging Using MacroscopicFourier Ptychographic TechnologyTIAN Zhiming ZHAO Ming WANG Sen LI Jian（Dalian Maritime University, Dalian 116026, China）Abstract Fourier ptychography is a promising high-resolution imaging technique that has been gradually applied in the field of macroscopic imaging. However, its imaging model typically provides a limited field-of-view of around 2 cm at meter-level imaging distances, which often falls short of practical requirements. To enhance the imaging distance and field-of-view of macroscopic Fourier ptychography, this article conducted theoretical research on the long-distance macro reflection Fourier stack imaging model. The proposed model utilizes diverging light beams for illumination, scans the target Fourier spectrum using spherical wavefront shifting, and reconstructs high-resolution target images. The article analyzes the mechanism of macroscopic coherent imaging and the approximation conditions of the Fourier imaging model, deriving the approximate range of the model and establishing a theoretical foundation for its extension. Finally, the built experimental system was used to image a target 10 meters away, increasing the target resolution from 1.4 mm to 0.35 mm, a resolution increase of more than 4 times, verifying the model’s capability to improve target imaging resolution through the synthetic aperture technology.收稿日期：2023-06-20引用格式：田芷铭, 赵明, 王森, 等. 宏观傅里叶叠层技术远距离成像实验研究[J]. 航天返回与遥感, 2023, 44(6): 38-44.TIAN Zhiming, ZHAO Ming, WANG Sen, et al. Experimental Research on Long-Range Imaging Using Macroscopic Fourier Ptychographic Technology[J]. Spacecraft Recovery & Remote Sensing, 2023, 44(6): 38-44. (in Chinese)第6期 田芷铭 等: 宏观傅里叶叠层技术远距离成像实验研究 39Keywords macroscopic imaging; Fourier ptychographic model; long-range imaging; super-resolution technology; Fourier ptychographic experiment0 引言目前，在监视、遥感等领域，高分辨率成像问题面临着重要挑战。
电容层析成像算法综述
电容层析成像算法综述
电容层析成像是用于可视化检测地质体中的电容特征的重要方法。
该技术可以通过将
探测器、地质学数据和成像处理算法结合起来，将地质电容值的分布可视化呈现出来，从
而让地质学家们能够更深入地掌握地质现象，发现更多重要信息。
在电容层析成像之前，地质学家只能通过层析测绘来发现地表现象，而电容层析成像
使他们能够从更深处应用电容测量方法，获得更多丰富的地质学数据。
大多数电容层析成
像算法都是基于处理大量地质学数据，可以分别计算每个数据点的垂直分布及水平分布的
电容值，随后将它们合成成图像，模拟真实的地质结构。
电容层析成像算法通常包括块动态变衰减、几何技术和水平投射计算法等。
其中，块
动态变衰减是地球物理学领域应用最为广泛的电容层析成像算法，它采用从多个深度层中
提取数据的方式，以计算每层的电容值的变化，最终绘制出电容变化的三维图像。
几何技
术主要是通过建立电容变化的几何模型，对各层的平均电容值进行评估，以建立起定量的
相关关系。
水平投射算法是一种比较复杂的算法，它是通过结合层析值、估算值和垂直投
射值来得出最终的相关结果。
电容层析成像算法是地球物理领域重要的数据处理改进算法，可以有效地从大量的矩
阵数据中提取实用型地球物理学知识，进而指导深入考察地质体背后的机理及特征。
但是，不同算法在数据处理过程中都会存在一定误差，因此应当加以考虑，并经过严格的试验和
研究，最终才能用于实际应用。
医学切片图像的配准
这里首先介绍在医学诊断方面广为应用的 CT 的基本思想。 通过一个理想的 X 射线源发出的笔束 X 射线，在对面放置一个 检测器。分别测出 X 射线经过物体衰减前后的强度 I0 和 I，再把 X 射 线源与检测器构成的系统在观察平面内平移一定得步数 N。每平移一 次均做同样的测量，得到一组数据。之后再旋转一个小角度 ΔΦ，接 着再平移 N 步，得到新的角度下的另一组数据。像这样重复旋转 NΦ 次，其中 NΦΔΦ=180º，得到 NΦ 组数据后停止。然后利用这些得到的 数据利用数学方法重新建立物体的模型或者内部图像。 特别要提到的是 CT 值的这个概念。由于 CT 是根据各种组织对 X 射线的吸收系数μ来决定其采用的标准，因此 Hounsfield 将线性衰 减系统分成了 2000 个单位，在医学上成为 CT 值。在 CT 图像的研究 中，提供诊断信息的是组织之间对 X 射线衰减系数的差异来体现的。 因此 CT 值定义如下：
第三章 数值实验模拟………………………………………………………17
3.1 医学图像切片预处理……………………………………………………17 3.2 利用全局匹配的方法配准图像…………………………………………19 3.3 利用局部匹配的方法配准图像…………………………………………21 3.4 结果………………………………………………………………………22
由于制片和图像采集等原因，序列数字切片图像的每两层间都会存在错位现 象，即平移和旋转等变换。基于切片数据的建模分析，其第一步就要对上下相邻 层切片图像进行配准，即通过图像的几何变换来完成校准。本文将从图像边缘曲 线匹配的角度来处理医学切片数据（人体躯干部位）相邻层之间的配准校正问题。
用于生命科学的人工智能定量相位成像方法
用于生命科学的人工智能定量相位成像方法English:Artificial intelligence (AI) has revolutionized the field of quantitative phase imaging (QPI) in life sciences by enabling high-throughput and accurate analysis of complex biological samples. AI-based QPI methods integrate advanced machine learning algorithms with computational imaging techniques to extract quantitative information from phase images with unprecedented precision and speed. These methods leverage deep learning models to perform tasks such as cell segmentation, classification, and tracking, enabling researchers to study dynamic cellular processes and disease progression with minimal human intervention. One example of AI-enabled QPI is label-free cell classification, where AI algorithms can accurately differentiate between different cell types based on their phase images, providing valuable insights into cellular morphology and function. Furthermore, AI-based QPI approaches can also improve the sensitivity and specificity of quantitative phase measurements, allowing for more accurate quantification of cellular and subcellular features. Overall, the integration of AI with QPI holds great potential for advancing our understanding of complexbiological systems and accelerating the development of novel diagnostic and therapeutic strategies in the life sciences.中文翻译:人工智能（AI）已经彻底改革了生命科学中的定量相位成像（QPI）领域，通过实现对复杂生物样品的高通量和准确分析。
quantative imagine analysis
quantative imagine analysisQuantitative image analysis (QIA) is a rapidly growing field that uses mathematical and statistical techniques to measure, analyze, and interpret the content of digital images. QIA has numerous applications in various fields, including medicine, biology, engineering, and social science.In the medical field, QIA is used to analyze medical images, such as X-rays, CT scans, and MRI scans, to detect diseases and evaluate their severity. By analyzing the texture, shape, and density of tissues in these images, QIA can help doctors identify subtle changes that may be difficult to see with the naked eye. This can improve the accuracy of diagnoses and allow for earlier detection of diseases, which can lead to more effective treatment options.In the biological field, QIA is used to analyze microscopic images of cells and tissues. By quantifying the size, shape, and distribution of various cell types and structures, QIA can provide valuable insights into the processes of diseases like cancer. This information can help researchers understand the mechanisms of diseases and develop new treatment strategies.In engineering, QIA is used to measure and analyze physical phenomena that cannot be accessed by traditional感官的方法。
计算机视觉与图像识别
计算机视觉与图像识别引言随着人工智能和机器学习技术的飞速发展，计算机视觉（Computer Vision）和图像识别（Image Recognition）在众多领域中得到了广泛的应用。
从医疗诊断到自动驾驶汽车，从安防监控到人脸识别，这些技术正逐步改变着我们的日常生活和工作方式。
本文将介绍计算机视觉和图像识别的基本概念、关键技术及其应用。
什么是计算机视觉？计算机视觉是一门研究如何使机器“看”的科学，即通过使用算法和模型让计算机能够理解和解释视觉信息。
它结合了图像处理、模式识别、人工智能等多个领域的研究成果，旨在实现对图像或视频中物体的检测、分类、跟踪和识别等任务。
什么是图像识别？图像识别是计算机视觉的一个重要分支，主要关注如何从图像中提取有用的信息，并对其进行分类或识别。
它包括面部识别、物体检测、场景理解等多种子领域，是实现智能监控、自动驾驶、医疗影像分析等应用的核心技术之一。
关键技术图像预处理图像预处理是图像识别过程中的第一步，主要包括去噪、增强对比度、灰度化、二值化等操作，以提高后续处理的准确性和效率。
特征提取特征提取是指从图像中提取出能够代表图像内容的特征向量，如边缘、角点、纹理等。
常用的特征提取方法有SIFT（尺度不变特征变换）、SURF（加速鲁棒特征）、HOG（方向梯度直方图）等。
机器学习与深度学习机器学习和深度学习是实现图像识别的两大主流方法。
传统机器学习方法依赖于手工设计的特征和分类器，而深度学习则通过构建深层神经网络自动学习图像特征，取得了显著的性能提升。
应用领域医疗影像分析计算机视觉和图像识别技术在医疗影像分析中发挥着重要作用，可以帮助医生快速准确地诊断疾病，如癌症检测、骨折判断等。
自动驾驶自动驾驶汽车利用计算机视觉技术感知周围环境，实现车辆定位、障碍物检测、交通标志识别等功能，为安全驾驶提供支持。
安防监控在安防监控领域，图像识别技术可以用于人脸识别、行为分析等，提高监控系统的智能化水平，有效预防犯罪行为的发生。
211116239_医学图像融合算法综述
科学研究创医学图像融合算法综述张伯妍1,2*陈翯1,2赵海霞1,2（1.北方民族大学数学与信息科学学院宁夏银川750021；2.北方民族大学宁夏智能信息与大数据处理重点实验室宁夏银川750021）摘 要：医学图像分析是辅助医疗诊断的重要方式，单一模式成像提供的信息有限，无法满足医生全面诊查病情的需求。
医学图像融合作为图像融合的一大重要分支，通过融合，可以提取不同医学影像的显著特征，丰富图像信息，提高融合质量。
本文在现有研究基础上，从图像融合概念展开，阐述医学图像融合原理和框架，梳理医学图像融合的分解方法和融合规则。
关键词：医学图像融合融合层次多尺度变换方法深度学习中图分类号：T P751文献标识码： A文章编号：1674-098X(2022)10(b)-0005-04A Review of Medical Image Fusion AlgorithmsZHANG Boyan1,2*CHEN He1,2ZHAO Haixia1,2( 1.School of Math and Information Science, North Minzu University, Yinchuan, Ningxia Hui Autonomous Region, 750021 China;2.The Key Laboratory of Intelligent Information and Big Data Processing, North Minzu University, Yinchuan, Ningxia Hui Autonomous Region, 750021 China ) Abstract: Medical image analysis is an important way to assist medical diagnosis. The information provided by single-mode imaging is limited and cannot meet the needs of doctors for comprehensive diagnosis. As an important branch of image fusion, medical image fusion can extract salient features of different medical images, enrich image information and improve the quality of fusion. Based on the existing research, this paper starts from the concept of image fusion, expounds the principle and framework of medical image fusion, and combs the decomposition methods and fusion rules of medical image fusion.Key Words: Medical image fusion; Fusion level; Multi-scale transformation methods; Deep learning图像融合是将不同图像的信息综合起来构成一幅全新图像的过程［1］。
基于用户操作的图像分割算法说明书
Phase-Based User-Steered Image Segmentation Lauren O’Donnell1,Carl-Fredrik Westin1,2,W.Eric L.Grimson1, Juan Ruiz-Alzola3,Martha E.Shenton2,and Ron Kikinis21MIT AI Laboratory,Cambridge MA02139,USAodonnell,***********.edu2Brigham and Women’s Hospital,Harvard Medical School,Boston MA,USAwestin,*******************.edu3Dept.of Signals and Communications,University of Las Palmas de Gran Canaria,Spain***************.esAbstract.This paper presents a user-steered segmentation algorithmbased on the livewire paradigm.Livewire is an image-feature drivenmethod thatfinds the optimal path between user-selected image loca-tions,thus reducing the need to manually define the complete boundary.We introduce an image feature based on local phase,which describes lo-cal edge symmetry independent of absolute gray value.Because phase isamplitude invariant,the measurements are robust with respect to smoothvariations,such as biasfield inhomogeneities present in all MR images.Inorder to enable validation of our segmentation method,we have createda system that continuously records user interaction and automaticallygenerates a database containing the number of user interactions,such asmouse events,and time stamps from various editing modules.We haveconducted validation trials of the system and obtained expert opinionsregarding its functionality.1IntroductionMedical image segmentation is the process of assigning labels to voxels in order to indicate tissue type or anatomical structure.This labeled data has a variety of applications,for example the quantification of anatomical volume and shape,or the construction of detailed three-dimensional models of the anatomy of interest.Existing segmentation methods range from manual,to semi-automatic,to fully automatic.Automatic algorithms include intensity-based methods,such as EM segmentation[14]and level-sets[4],as well as knowledge-based methods such as segmentation by warping to an anatomical atlas.Semi-automatic algorithms, which require user interaction during the process,include snakes[11]and livewire [1,7].Current practice in many cases,however,is computer-enhanced manual segmentation.This involves hand-tracing around the structures of interest and perhaps employing morphological or thresholding operations to the data.Although manual segmentation methods allow the highest degree of user control and enable decisions to be made on the basis of extensive anatomical knowledge,they require excessive amounts of user interaction time and mayW.Niessen and M.Viergever(Eds.):MICCAI2001,LNCS2208,pp.1022–1030,2001.c Springer-Verlag Berlin Heidelberg2001Phase-Based User-Steered Image Segmentation1023 introduce high levels of variability.On the other hand,fully automatic algorithms may not produce the correct outcome every time,hence there exists a need for hybrid methods.2Existing Software PlatformThe phase-based livewire segmentation method has been integrated into the 3D Slicer,the Surgical Planning Lab’s platform for image visualization,image segmentation,and surgical guidance[9].The Slicer is the laboratory’s program of choice for manual segmentation,as well as being modular and extensible,and consequently it is the logical place to incorporate new algorithms for our user base.At any time during an average day,one would expect tofind ten people doing segmentations using the3D Slicer in the laboratory,which shows that any improvement on manual methods will be of great utility.Our system has been added into the Slicer as a new module,Phasewire,in the Image Editor.3The Livewire AlgorithmThe motivation behind the livewire algorithm is to provide the user with full control over the segmentation while having the computer do most of the detail work[1].In this way,user interaction complements the ability of the computer tofind boundaries of structures in an image.Initially when using this method,the user clicks to indicate a starting point on the desired contour,and then as the mouse is moved it pulls a“live wire”behind it along the contour.When the user clicks again,the old wire freezes on the contour,and a new live one starts from the clicked point.This intuitive method is driven only by image information,with little geometric constraint,so extracting knowledge from the image is of primary importance.The livewire method poses this problem as a search for shortest paths on a weighted graph.Thus the livewire algorithm has two main parts,first the conversion of image information into a weighted graph,and then the calculation of shortest paths in the graph[6,12].In thefirst part of the livewire algorithm,weights are calculated for each edge in the weighted graph,creating the image forces that attract the livewire. To produce the weights,various image features are computed in a neighborhood around eachgraphedge,transformed to give low edge costs to more desirable feature values,and then combined in some user-adjustable fashion.The general purpose of the combination is edge localization,but individual features are gen-erally chosen for their contributions in three main areas,which we will call“edge detection,”“directionality,”and“training.”Features suchas th e gradient and the Laplacian zero-crossing[13]have been used for edge detection;these features are the main attractors for the livewire.Second,directionality,or the direction the path should take locally,can be influenced by gradients computed with ori-ented kernels[7],or the“gradient direction”feature[13].Third,training is the1024L.O’Donnell et al.process by which the user indicates a preference for a certain type of bound-ary,and the image features are transformed accordingly.Image gradients and intensity values[7],as well as gradient magnitudes[13],are useful in training.In the second part of the livewire algorithm,Dijkstra’s algorithm[5]is used to find all shortest paths extending outward from the starting point in the weighted graph.The livewire is defined as the shortest path that connects two user-selected points(the last clicked point and the current mouse location).This second step is done interactively so the user may view and judge potential paths,and con-trol the segmentation asfinely as desired.The implementation for this paper is similar to that in[6],where the shortest paths are computed for only as much of the image as necessary,making for very quick user interaction.4Phase-Based Livewire4.1Background on Local PhaseThe local phase as discussed in this paper is a multidimensional generalization of the concept of instantaneous phase,which is formally defined as the argument of the analytic functionf A=f−if Hi(1) where f Hi denotes the Hilbert transform of f[2,3].The instantaneous phase of a simple one-dimensional signal is shown in Fig-ure1.This simple signal can be thought of as the intensity along a row of pixels in an image.Note that in this“image”there are two edges,or regions where the signal shape is locally odd,and the phase curve reacts to each,giving phase values ofπ/2and−π/2.Fig.1.AThe output from a local phasefilter,i.e.a Gaborfilter,is an analytic function that can be represented as an even real part and an odd imaginary part,or anPhase-Based User-Steered Image Segmentation1025 amplitude and an argument.Consequently,this type offilter can be used to estimate both the amplitude(energy)and the local phase of a signal.As local phase is invariant to signal energy,edges can be detected from small or large signal variations.This is advantageous for segmentation along weak as well as strong edges.Another advantage is that local phase is in general a smoothly varying function of the image,and provides subpixel information about edge position.4.2ImplementationTo estimate the phase we are using quadraturefilters which have a radial fre-quency function that is Gaussian on a logarithmic scale:ν(ρ)=e−4log2B−2log2(ρρi)(2) whereρi is the center frequency and B is the6dB sensitivity bandwidth in octaves.In contrast to Gaborfilters,thesefilters have zero response for negative frequencies,ensuring that the odd and even parts constitute a Hilbert transform pair and making thefilters ideal for estimation of local phase[10].We multiply thefilters with a cos2radial function window in the frequency domain to reduce ringing artifacts arising from the discontinuity atπwhen usingfilters with high center frequencies and/or large bandwidth.The local phase is our primary feature,serving to localize edges in the image: the livewire essentially follows along low-cost curves in the phase image.The phase feature is scaled to provide bounded input to the shortest-paths search. Figure2shows a phase image derived from the original image shown in Figure 4.Since the argument of the phase is not sensitive to the amount of energy in the signal,it is useful to also quantify the certainty of the phase estimate.In image regions where the phase magnitude is large,the energy of the signal is high,which implies that the the phase estimate in that region has high reliability and thus high certainty.We define the certainty using the magnitude from the quadraturefilters, mapped through a gating function.By inverting this certainty measure,it can be used as a second feature in the livewire cost function.The weighted combination of the phase and certainty images produces an output image that describes the edge costs in the weighted graph,as in Figure2.5ApplicationsIn this section,we describe the behavior of our phase-based segmentation tech-nique through a series of experiments with medical images.In thefirst set of experiments,we show the stability of phase information across different scales[8].The input to the quadraturefilters was blurred using a Gaussian kernel of variance4pixels.Figure3demonstrates the robustness1026L.O’Donnell et al.Local Phase Image Combined ImageFig.2.Local phase images.The left-hand image shows the local phase corre-sponding to the grayscale image in Figure 4.This is the primary feature employed in our phase-based livewire method.The right-hand image displays the weighted combination of the phase and certainty features,which emphasizes boundaries in the image.Darker pixels correspond to lower edge costs on the weighted graph.of phase information to changes in scale,as the two segmentations are quite similar.The leftmost image is the original grayscale and its segmentation,while the center image shows the result of Gaussian blurring.The rightmost image displays the segmentation curve from the blurred image over the original image for comparison.The second experiments contrast the new livewire feature,phase,with our original intensity-based livewire implementation.Phase is more intuitive to use,produces smoother output,and is generally more user-friendly than the intensity-based method as illustrated in Figure 4.In contrast to the intensity-based method,no training step is required to bias the phase-based method towards the desired edge type,as the phase information can be used to follow weak or strong edges.6Discussion and Validation FrameworkWe have presented a user-steered segmentation algorithm which is based on the livewire paradigm.Our preliminary results using local phase as the main driving force are promising and support the concept that phase is a fairly stable feature in scale space [8].The method is intuitive to use,and requires no training despite having fewer input image features than other livewire implementations.In order to enable validation of phase-based livewire,we have created a sys-tem that continuously records user interaction and can generate a database con-taining the number of user interactions and time stamps from various editing modules.A selection of the items which are recorded is listed in Table 1.AnPhase-Based User-Steered Image Segmentation1027Original Image Gaussian Blurring Overlay of SegmentationFig.3.Illustration of insensitivity to changes in scale:initial image,result of Gaussian blurring,and overlay of second segmentation(done on the blurred image)on initial image.Note the similarity between the segmentation contours, despite blurring the image with a Gaussian kernel of four pixel variance. ongoing study using this system will enable evaluation of the performance and utility of different image features for guiding the livewire.Analysis of the logged information has shown that great variability exists in segmentation times,partially due to users’segmentation style,learning curve, and type of image data,and also due to factors we cannot measure,suchas time spent using other programs with the Slicer Editor window open.A complete comparison of the phasewire and manual segmentations systems,consequently, is difficult,and perhaps the most reliable indication of the system performance can be found in the opinions of doctors who have used it.Table2demonstrates the utility of the segmentation method using infor-mation from the logging system,and Table3gives doctors’opinions regarding ease of use and quality of output segmentations.For example,Figure5shows a liver segmentation done with phasewire,in which the doctor remarked that the three-dimensional model was more anatomically accurate than the one pro-duced with the manual method,where both segmentations were performed in approximately the same amount of time.This is likely due to the smoothness of the contour obtained with phasewire,in comparison with the contour obtained with the manual method.Item Logged Detailsmouse clicks recorded per slice,per label value,per volume,and per editor module elapsed time recorded per slice,per label value,per volume,and per editor module description includes volumes viewed and location of scene descriptionfile username will allow investigation of learning curveTable1.Selected items logged by the3D Slicer for validation and comparison purposes.1028L.O’Donnell et al.parison of phase-based and intensity-based implementations of livewire.On the left is our original implementation of livewire,which uses image-intensity based features.It does a poorer job of segmenting gray and white mat-ter.Training was performed on the region in the upper left of the image.The segmentation was more difficult to perform than the phase-based segmentation because more coercing was necessary to force the livewire to take an accept-able path.The right image shows segmentation of the same boundary using the phase-based livewire.Contrary to the intensity-based method,no training was necessary before beginning to segment.The conclusion is that the phase-based implementation is more intuitive to use than the intensity-based version,since the phase image gives continuous contours for the livewire to follow.Rather than being distracted by unwanted pixels locally,the phase livewire is generally only distracted by other paths,and a small movement of the mouse can set it back on the desired path.It can be quitefluid to use,with clicks mainly necessary when the curve bends in order to keep the wire following the correct contour. 7AcknowledgementsThe work was funded in part by CIMIT and NIH P41-RR13218(RK,CFW),by a National Science Foundation G raduate ResearchFellowsh ip(LJO),by NIH R01 RR11747(RK),by Department of Veterans Affairs Merit Awards,NIH grants K02MH01110and R01MH50747(MS),and by European Commission and Spanish Gov.joint research grant1FD97-0881-C02-01(JR).Thanks to the Radi-ation Oncology,Medical Physics,Neurosurgery,Urology,and Gastroenterology Departments of Hospital Doctor Negrin in Las Palmas de Gran Canaria,Spain, for their participation in evaluation of the system.Also thanks to Aleksandra Ciszewski for her valuable feedback regarding the Slicer system. References1.W.A.Barrett and E.N.Mortensen.Interactive live-wire boundary extraction.Medical Image Analysis,1(4):331–341,1997.Phase-Based User-Steered Image Segmentation1029Fig.5.A surface model of the liver,created from a phasewire segmentation, shows well-defined surfaces between the liver and the kidney and gallbladder. study method total clicks clicks/slice time/slice(sec)volume(mL) CT brain tumor manual23426.039.368.5phasewire9710.828.767.3phasewire10912.125.569.2CT bladder manual148828.131.7715.6phasewire359 6.821.5710.8Table2.Example segmentations performed withPh asewire.Clicks and time per slice are averages over the dataset.2.Boualem Boashash.Estimating and interpreting the instantaneous frequency of asignal-part1:Fundamentals.Proceedings of the IEEE,80(4),1992.3.Boualem Boashash.Estimating and interpreting the instantaneous frequency of asignal-part2:Algorithms and applications.Proceedings of the IEEE,80(4),1992.4.V.Caselles,R.Kimmel,G.Sapiro,C.Sbert.Minimal Surfaces:A Three DimensionalSegmentation Approach.Technion EE Pub973,1995.5. E.W.Dijkstra A Note On Two Problems in Connexion With Graphs NumerischeMathematik,269–271,1959.6. A.X.Falcao,J.K.Udupa,and F.K.Miyazawa.An Ultra-Fast User-Steered ImageSegmentation Paradigm:Live Wire on the Fly.IEEE Transactions on Medical Imaging,19(1):55–61,2000.7. A.X.Falcao,J.K.Udupa,S.Samarasekera,and er-Steered ImageSegmentation Paradigms:Live Wire and Live Lane.Graphical Models and Image Processing,60:233–260,1998.8. D.J.Fleet and A.D.Jepson.Stability of Phase Information IEEE Trans.PAMI,15(12):1253–1268,1993.9. D.T.Gering,A.Nabavi,R.Kikinis,W.E.L.Grimson,N.Hata,P.Everett,andF.Jolesz and W.Wells.An Integrated Visualization System for Surgical Plan-ning and Guidance Using Image Fusion and Interventional Imaging.Medical Image Computing and Computer-Assisted Intervention-MICCAI’99,809–819,1999.1030L.O’Donnell et al.study method ease of use segmentation qualityCT brain tumor manual33phasewire44CT bladder manual 2.73phasewire44Table3.Doctors’comparison of Phasewire and Manual segmentation methods on the datasets from Table2.The scale is from1to5,with5being the highest.10.H.Knutsson and C.-F.Westin and G.H.Granlund.Local Multiscale Frequencyand Bandwidth Estimation Proceedings of IEEE International Conference on Image Processing,36–40,1994.11.T.McInerney,D.Terzopoulos.Deformable Models in Medical Image Analysis:ASurvey.Medical Image Analysis,1(2):91–108,1996.12. E.N.Mortensen and W.A.Barrett.Interactive Segmentation with IntelligentScissors.Graphical Models and Image Processing,60(5):349–384,1998.13. E.N.Mortensen and W.A.Barrett.Intelligent Scissors for Image Composition.Computer Graphics(SIGGRAPH‘95),191–198,1995.14.W.M.Wells,R.Kikinis,W.E.L.Grimson,F.Jolesz.Adaptive segmentation ofMRI data.IEEE Transactions on Medical Imaging,15:429–442,1996.。
- 1、下载文档前请自行甄别文档内容的完整性，平台不提供额外的编辑、内容补充、找答案等附加服务。
- 2、"仅部分预览"的文档,不可在线预览部分如存在完整性等问题,可反馈申请退款(可完整预览的文档不适用该条件!)。
- 3、如文档侵犯您的权益，请联系客服反馈,我们会尽快为您处理(人工客服工作时间：9:00-18:30)。
(a)
(b)
(c)
Figure 1: Automatic exposure as the cause of differently exposed pictures of the same (overlapping) subject matter: (a) Looking from inside Hart House Soldier’s Tower, out through an open doorway, when the sky is dominant in the picture, the exposure is automatically reduced, and we can see the texture (clouds, etc.) in the sky. We can also see University College and the CN Tower to the left. (b) As we look up and to the right, to take in subject matter not so well illuminated, the exposure automatically increases somewhat. We can no longer see detail in the sky, but new architectural details inside the doorway start to become visible. (c) As we look further up and to the right, the dimly lit interiour dominates the scene, and the exposure is automatically increased dramatically. We can no longer see any detail in the sky, and even the University College building, outside, is washed out (overexposed). However, the inscriptions on the wall (nhe war) now become visible. (a,b,c) The differently exposed pictures of overlapping subject matter can be combined to extend dynamic range and tonal definition, or to provide a true photographic quantity “lightspace” for intelligent vision systems.
the image such as by cos(α4 ), etc., to make the camera become a truly quantimetric instrument.) In this example, we have three very differently exposed pictures depicting parts of the University College building and surroundings.
Abstract
Multiple differently exposed pictures of the same subject matter arise naturally whenever a video camera having automatic exposure captures multiple frames of video with the same subject matter appearing in regions of overlap between at least some of the successive video frames. Almost all cameras have some kind of automatic exposure feature. Generally automatic exposure is center weighted, so that when a light object falls in the center of the frame the exposure is automatically decreased, whereas the exposure is automatically increased when the camera swings around to point at a darker object. In this paper, it is assumed that the spatial (e.g. projective) coordinate transformation between successive frames of the sequence is known (or equivalently that it is the identity), and the contribution of the paper is an efficient way to estimate the tonal relationship between successive frames of the sequence. In particular methods are proposed to simultaneously estimate the unknown camera response function, as well as the set of unknown relative exposure changes among images, up to a single unknown scalar constant. The method comprises a succession of guesses each of which is a refinement of the previous. The first guess is often sufficient, so that no initial solution needs to be provided by the user. Each subsequent guess is a least squares solution so that no computationally expensive optimization is required. Since the method makes use of all the data, it is extremely immune to noise. The method is tested against state-of-the art laboratory measurement instruments to confirm the accuracy of the results.
Differently exposed images (e.g. individual frames of video) of the same subject matter are denoted as vectors: f0 , f1 , . . ., fi , . . ., fI −1 , ∀i, 0 ≤ i < I . Each video frame is some unknown function, f (), of the actual quantity of light, q (x) falling on the image sensor: fi = f ki q Ai x + bi ci x + di , (1)
1. Introduction: Variable gain image sequence processing
Many papers have been published on the problems of motion estimation and frame alignment; for review see [1]. Most of these assume fixed gain. In practice, however, camera gain varies to compensate for varying quantity of light, by way of Automatic Gain Control (AGC), automatic level control, or some similar form of automatic exposure. In fact almost all modern cameras incorporate some form 1
Quantigraphic Imaging: Estimating the camera response and exposures from differently exposed images
Steve Mann University of Toronto, Dept. E.C.E. Toronto, Ontario, Canada, M5S 3G4 Richard Mann Dept. C.S. University of Waterloo
of automatic exposure control. Moreover next generation cameras, such as EyeTap devices () that cause the eye itself to function, in effect, as if it were both a camera and display, also feature an automatic exposure control system to make possible a hands free gaze activated wearable system operable without conscious thought or effort. Indeed, the human eye itself incorporates many features akin to the automatic exposure or AGC of modern cameras. Figure 1 illustrates how such a camera takes in a typical scene. As we look straight ahead we see mostly sky, and the exposure is quite small. Looking to the right, at darker subject matter, the exposure is automatically increased. Since the differently exposed pictures depict overlapping subject matter, we have (once the images are registered, in regions of overlap) differently exposed pictures of identical subject matter. (Registration typically also includes correction of barrel distortion, correction for darkening at the corners of