Techniques and Applications of Image Understanding
Techniques and Applications of Image Understanding
复制标题
DOI：
10.1080/713820782
复制
发表时间：
1982-11
影响因子：
1.3
通讯作者：
L. Norton-Wayne
中科院分区：
文献类型：
--
作者：
L. Norton-Wayne
这位评论家一直认为，对场景的理解涉及（从包括计算机可处理形式的原始图像的像素阵列中的符号信息）提取其内容的描述，该描述在语义上是有意义的和一致的，而不仅仅是通过粗略的基于统计的分割形状与存储的原型的比较来识别对象。虽然有一些论文，如RA布鲁克斯的《制造业视觉的几何建模》，确实在这个层面上攻击了机器视觉，但大多数论文都关注图像的分析，而不是对图像的理解。这并不是贬低它们的价值，只是指出这卷书的标题并不完全合适。论文分为四组;前两组（在介绍DARPA要实现的目标和概述之后）分别题为“图像理解的战术应用”和“图像理解的侦察应用”。我发现他们相当乏味。一个例外是NB Nill的论文“将矩应用于图像理解”;尽管该论文的标题用词不当（没有真正提供提取高级信息方面的“理解”），但对基础知识的完整和非常清晰的介绍使这篇论文非常值得阅读。第三组“图像理解的制造和检测应用”肯定更有趣，因为这些论文真的不涉及图像理解（布鲁克斯的贡献已经提到是一个明显的例外）。我特别喜欢GM Fitton的论文《实时模式识别--一个工业实例》;尽管它关注简单和传统技术的直接应用，但它描述了一个经过充分考虑和测试的系统，用于一个有用的应用。JR比尔克、RB凯利和JD Dessimoz的《机器人搬运无定向零件的视觉控制》一文中对系统的性能作了定量描述。太多的论文因为省略了这一重要信息而显得软弱无力。最后一组论文是在专题讨论会之后的DARPA研讨会上发表的，它们是迄今为止最好的，因为它们提出了适用于各种各样问题的普遍感兴趣的结果。某些主题的出现是目前的意义是显而易见的，如提取的三维描述从二维图像，纹理的分析，并出现了专用的VLSI并行硬件设计的图像分析。特别的论文脱颖而出;相对深度和局部表面方向从图像运动的K。例如，Prazdny证明了可以独立于相对运动从动态图像计算相对深度的重要结果。遗憾的是，文件中引用的一些参考文献没有在最后列出。利用L.厨房和A。罗森菲尔德应该要求阅读的图像调查，不仅是因为其结果的广泛适用性，但考虑到工作的方式，最后与一个清晰的，定量的比较结果。本卷的最后一节将是极大的兴趣，所有参与机器图像分析;我希望看到它的论文经常引用作为参考。其他部分将主要是那些专业是相关部分的人感兴趣。
This reviewer has always understood theunderstanding'of scenes to involve the extraction (from the symbolic information in the pixel array which comprises the original image in computer processable form) of a description of its content which is semantically meaningful and consistent, rather than mere identification of objects by crude statistically based comparison of segmented shapes with stored prototypes. Though a few of the papers, such asGeometric modelling for vision in manufacturing'by RA Brooks, really do attack machine vision at this level, most of the papers are concerned with the analysis of images rather than with their understanding. This is not to denigrate their worth, merely to point out that the title of the volume is not altogether appropriate. The papers are divided into four groups; the first two (following an introduction explaining what DARPA is setting out to achieve, and an overview) are entitled, respectively,Tactical applications of image understanding'and'Reconnaisance applications of image understanding'. I found them rather pedestrian. An exception is the paper by NB NillApplications of moments to image understanding'; although the title of the paper is a misnomer ('understanding'in terms of the extraction of high-level information is not really offered), a complete and very lucid presentation of basics makes this paper well-worth reading. The third groupManufacturing and inspection applications of image understanding'is definitely more interesting, given that the papers really do not deal with image understanding (Brooks' contribution already mentioned being a notable exception). I enjoyed particularly GM Fitton's paperReal-time pattern recognition-an industrial example'; though concerned with the straight application of simple and conventional techniques, it describes a fully thought-out and tested system for a useful application. The paperVisual control for robot handling of unoriented parts' by JR Birk, RB Kelley, and JD Dessimoz is noteworthy in including a quantitative description of the performance of their system. Too many papers are emasculated when this essential information is omitted. The final group of papers, those presented at the DARPA Workshop which followed the symposium, are by far the best, in presenting results of general interest applicable to a wide variety of problems. The emergence of certain topics as being of current significance is evident, such as the extraction of three-dimensional descriptions from two-dimensional images, the analysis of texture, and the emergence of special purpose VLSI parallel hardware designed for image analysis. Particular papers stand out;Relative depth and local surface orientation from image motions' by K. Prazdny for example, proves the important result that relative depth can be computed from dynamic imagery independently of relative motion. It is a pity that some of the references quoted within the paper are not listed at the end. The paperEdge evaluation using local edge coherence'by L. Kitchen and A. Rosenfeld should be required reading for image investigators, not merely because of the wide applicability of its results, but on account of the way in which the work is presented, concluding with a lucid, quantitative comparison of results.The final section of this volume will be of great interest to all involved in machine image analysis; I expect to see its papers quoted frequently as references. The other sections will be of interest chiefly to those whose speciality is that of the relevant section.