Book Reviews

Abstract

Ballard, 1) .H. and C .M . Brown, Computer Vision, Englewood Cliffs, NJ : Prentice-Hall, 1982, 523 pp. , $40 .00, ISBN 0-13-165316-4 . What is computer vision? Computer vision, also known as image understanding, is defined as the construction of explicit, meaningful descriptions of physical object s from images . This definition differs from image processing , a field which studies image-to-image transformations suc h as image enhancement . The book's objective is to foster an understanding of a rapidly changing field . The computer must not only trans form objects, but automatically analyze and understand images so that descriptions — prerequisites for recognizin g and thinking about objects — can be obtained . To reac h this objective, a selection of ideas from disparate fields such as computer graphics, artificial intelligence an d psychology are discussed . These ideas are presented in a clear and meaningful fashion with the extensive aid of diagrams, algorithms, pictures and exercises . Thus, the book is useful both as a reference for someone constructing a computer vision system, and as a text for graduate studies . Although it has a strong artificial intelligence flavour, i t propounds its concepts well as a provocative tool in computer graphics . The book is organized into four major sections whic h follow a progression of increasing abstractness . The authors progress through generalized images, segmente d images, geometrical structures and relational structures . Part I, generalized images, is at the lowest level, bein g chiefly concerned with the set of related images or image like entities that can be derived from a scene . Such entitie s can be the result of edge analysis or may be derived fro m intrinsic physical qualities such as color . The first half of Part I discusses mathematical models of images and imag e formation and specific image formation technologies . The latter half is concerned with preprocessing in order to und o degeneracies in the image being processed (basically a col lection of techniques that exploit pixel redundancy) . Fundamental constraints between the physical parameters an d gray levels are derived without the need for high-level inter nal model information . The reader must be familiar with the material presented here, however, for most topics are merely summarized so as to present a flavour of wha t techniques are available . If there is any need for furthe r clarification on the subject matter, the extensive bibli ography provides excellent references (as do the bibli ographies at the end of each of the chapters in the book) . Part II, segmented images, deals with assembling groups of generalized images with one or more homogene ous features or characteristics . This part is four chapters i n length . The first two describe methods for obtaining seg mentation with respect to boundaries and regions . The latte r two discuss methods for using texture and motion for seg mentation . Shape grammars are also used here in order to present knowledge about the real world of objects . More information on such grammars can be obtained from biologyjournals or from texts concerned with patterns in nature . Part III, geometrical structures, attempts to conve y present-day knowledge of the representation of shape i n order that shapes may be learned, matched against, recollected and used . In so doing, the authors are hampered b y several factors such as the complexity of shapes, the unresolved question of why shape recognition is so easy fo r humans, a lack of classical guidance, and the fact that shap e recognition is in fact still a young discipline . The author s spend only one chapter each on two-dimensional and threedimensional shapes . It is stressed in the second chapter tha t three-dimensional representations are fundamental to th e performance of any image understanding task . As in part I , they present in summary fashion techniques that are full y described in other references . Part IV, relational structures, is the highest level of abstraction . Here, techniques for making the motivation an d world view of a vision system explicit and available are explained . The authors divided this part into five majo r topics : knowledge representation using semantic nets for structuring complex knowledge ; matching to put a derive d representation of an image into correspondence with a n existing representation ; inference techniques applied to bot h problem solving and belief-maintenance activity ; plannin g techniques which are used for resource allocation and atten tional mechanisms ; and control strategies and mechanisms as required in vision processing . A topic that is missin g from the list is learning ; very little is said about that for the domain of vision . One of the nice features of the book is the encapsulation of the ideas that are to be presented at the beginnin g of each major topic . With each summary, a tree diagram shows the relationship of the topics covered . Although the book is well organized . I feel that the algorithms and figures could have been attended to with a little more thought and coherence . While the algorithms are set off in the text, their level of detail varies . Some can be directly translated into programs while others are simpl y disjoint text . The figures, albeit numerous, are sometimes left dangling . More explanations in the body of the tex t would have eliminated this problem . On the positive side , this is a good book on methods for image understanding . It contains fairly complete appendices on mathematica l tools and control mechanisms . Since the publication of thi s book in 1982, advances have been made regarding the prac tical problems of computer vision . The authors' insights into the present flux of ideas of image understanding have made this volume an excellent introduction to fu rther readings in the field .