视觉认知框架：Marr的《视觉》启示与研究进展

需积分: 11 120 浏览量更新于2024-07-27 收藏 30.84MB PDF 举报

《机器视觉》（Vision: A Computational Investigation into the Human Representation and Processing of Visual Information）是大卫·马尔（David Marr）于1982年发表的一部具有里程碑意义的著作，他在书中探讨了人类视觉感知的普遍框架，并以此为基础深入思考大脑如何处理和理解视觉信息。这本书对脑科学和认知科学领域产生了深远的影响，激励了众多研究者投身于这一研究。大卫·马尔在生前未能亲眼见证这本书的出版，但他的思想遗产在后人的传承中得以延续。书中的核心内容围绕以下几个关键知识点展开： 1. **理论框架**：马尔提出了一个计算机科学视角来分析视觉系统，将视觉处理分为三个层次：输入（perception），解释（representation）和行为（processing）。这三个层次分别是：传感器接收外部世界的信息，通过复杂的算法和神经网络模型形成对场景的理解，最后根据这些理解指导行为决策。 2. **人类视觉系统的模型**：他构建了一个理论模型，探讨了视网膜、初级视觉皮层（如V1）以及更高级别的视觉区域之间的信息流，这些区域负责解析、整合和抽象不同的视觉特征，如边缘、纹理、颜色和形状等。 3. **计算理论与生物学相结合**：马尔强调了将认知过程与生物机制相结合的重要性，试图找到神经元活动与视觉处理任务之间的对应关系，这有助于科学家们理解和模拟大脑的工作原理。 4. **视觉信息处理的可计算性**：他提出，视觉系统的功能可以通过可计算的算法实现，这不仅推动了人工智能的发展，也为后来的研究者提供了实验设计和理论验证的基础。 5. **跨学科影响**：这本书不仅在视觉心理学、神经科学和认知科学领域产生了重大影响，还促进了多学科交叉研究，比如计算机视觉、机器学习和人工智能等领域的发展。 6. **版权和授权**：《机器视觉》一书享有版权保护，未经许可，任何电子或机械复制方式的传播都需获得出版商书面许可。该书的重新出版由麻省理工学院出版社发行，展示了学术著作在学术界和社会上的持久价值。《机器视觉》是一部探索人类视觉系统工作原理的杰作，它开启了对大脑认知功能的计算机建模研究，对于理解人眼如何解读世界，以及如何将其应用于人工智能技术的发展具有深远的影响。随着科技的进步，这本书的思想仍然在不断被现代科学家们重新审视和扩展。

Detailed Contents xv

Reformulating the Problem 272

The Information to be Represented 275

General Form of the 2½-D Sketch 277

Possible Forms for the Representation 279

Possible Coordinate Systems 283

Interpolation, Continuation, and Discontinuities 285

Computational Aspects of the Interpolation Problem 288

Discontinuities 289

Interpolation methods 290

Other Internal Computations 291

Chapter 5

Representing Shapes for Recognition 295

Introduction 295

Issues Raised by the Representation of Shape 296

Criteria for judging the effectiveness of a shape representation 296

Accessibility 297

Scope and uniqueness 297

Stability and sensitivity 298

Choices in the design of a shape representation 298

Coordinate systems 298

Primitives 300

Organization 302

The 3-D Model Representation 302

Natural coordinate systems 303

Axis-based descriptions 304

Modular organization of the 3-D model representation 305

Coordinate system of the 3-D model 307

Natural Extensions 309

Deriving and Using the 3-D Model Representation 313

Deriving a 3-D model description 313

Relating viewer-centered to object-centered coordinates 317

Indexing and the catalogue of 3-D models 318

Interaction between derivation and recognition 321

Finding the correspondence between image and catalogued

model 322

Constraint analysis 322

Psychological Considerations 325

xviii Foreword

I knew David closely,  rst as a student and then as a colleague. I

had many long discussions with him during his years at MIT, and I

miss him greatly both as a friend and as a colleague. In these intro-

ductory remarks, I re ect brie y on the development of some of his

ideas during these years, how they looked then, and how they look

today.

Looking back, it is striking to observe the amazing rate at which

the basic framework of the theory evolved soon after Marr’s arrival

at MIT. In the Arti cial Intelligence laboratory at MIT, ongoing re-

search was often described in internal publications called “AI mem-

os.” During his years at MIT, Marr produced a  urry of these memos

re ecting the rate and intensity of his research. In 1974, his  rst year

at MIT, a series of three AI memos described in detail the theory of

early vision, with an initial implementation of the so-called primal

sketch. As was characteristic of much of his work, the  rst in the se-

ries was a careful consideration of the overall goal of low-level vision:

an autonomous process, which produces a symbolic representation,

useful for higher-level processes. Subsequent memos then described

the details of the process, for example,  nding peaks and derivatives

in intensity pro les and making assertions about edges and bars,

their location, width, and blur.

An important insight gained from the work on the primal sketch

was the realization of the inherent complexity of early visual com-

putations, including edge detection. A number of edge-detection

techniques, such as the so-called Sobel operator, were widely used

at the time. They were fast and simple to use, but performed poorly

when applied to natural images. Marr, together will Ellen Hildreth,

devised a principled and systematic approach to edge detection, later

used in the popular Canny edge detector.

The primal sketch and edge-detection models also had implica-

tions for the study of cortical circuitry. Following the seminal work

of Hubel and Wiesel on the physiology of the primary visual cor-

tex, cells in this cortical region were often described as “edge detec-

tors.” The computational work on edge detection made it clear that

simple cells in the primary visual cortex could not, by themselves,

be edge detectors. They could play a useful role in the process, but

more elaborate circuitry, involving multiple units, will be required

for reliable edge detection. The general implication was that com-

putational studies of speci c visual tasks, such as edge detection and

binocular vision, can play a useful, sometimes crucial, role in the

understanding of neural circuitry.

Foreword xix

The work on the primal sketch and subsequently on binocular

stereo matching fostered the belief that, due to the enormous com-

plexity inherent in low-level vision, understanding the circuitry and

response properties in the visual system would be dif cult to at-

tain and remain incomplete without complementary studies at the

computational and algorithmic levels. At the same time, to be of

relevance to neurophysiology, computational studies of vision would

have to address in detail speci c visual problems, rather than pursue

general mathematical formulations. This conclusion is manifested

in Marr’s sharp criticism of a book titled Physics and Mathematics of

the Nervous System. A review published in Science in 1975 opens with

Marr’s characteristic unabated style: “Many experimental biologists

dismiss with contempt the approach of even very able theoreticians

to developmental of neurophysiological problems. The outsider

needs look no further than this volume to understand why. Some

of the papers describe attempts to elucidate problems of biological

information processing, but in one way or another they all make the

same error of strategy—engaging in the search of a general theory

before and actually instead of tackling any of the particular problems

at hand.”

How is the primal sketch viewed today in neurophysiology and

in computational vision? Marr viewed the primal sketch as a rich

symbolic description of intensity changes in the image, composed of

two main stages: the extraction and classi cation of local intensity

changes, followed by the grouping of the local changes into more

extended entities. Plausible anatomical candidates for these compu-

tations are cortical areas V1 and V2, with V2 playing perhaps a more

important role in the grouping stages. There has been some evidence

relating V2 to grouping process, based in particular on the responses

of V2 units to subjective contours, and their sensitivity to border

ownership and  gure-ground relationships. Area V1 is still often

considered in neurophysiology to be a bank of oriented or Gabor-

like  lters applied to the image. Many in the  eld, however, suspect

that V1 may provide a substantially richer description of the image,

along the line proposed by Marr. Evidence from single units and

from brain imaging suggests that V1 may not be as autonomous in

its function as suggested by Marr: top-down signals from higher-

level visual areas appear to have signi cant effects on the computa-

tions performed by V1. The complexity of early visual processes is

now broadly appreciated from both computational and biological

standpoints. Because of this complexity, it is perhaps not surprising

剩余428页未读，继续阅读

samtyn

粉丝: 0
资源: 1

视觉认知框架：Marr的《视觉》启示与研究进展

MATLAB Machine Vision Toolbox 4.1：机器人视觉与控制的实用工具

LabVIEW机器视觉模块详解：IMAQ Vision与Machine Vision控件

LabVIEW机器视觉模块解析：IMAQ Vision与Machine Vision控件

Machine Vision Handbook

Machine Vision Toolbox

机器视觉 Machine Vision

Machine Vision.zip

Computer and Machine Vision

Machine Vision Toolbox 4.1

GigE Vision Camera Interface Standard for Machine Vision

最新资源