A region analysis system for natural color scenes has been intensively studies. In the system described here, the input image is converted once into a structured data network, and the knowledge to be used in the higher-level analysis is coded as a set of rules which work on this network.
A wide range of issues have been dealt with in this paper from signal domains to semantic ones.
, . . In chapter 2, we discuss the use of color information in region segmentation . . . . A segmentation scheme, called the “dynamic K-L transformation,” was developed for this purpose. . . . An experiment showing that the color in natural scenes is physically almost two-dimensional was also presented. . . .
, . . A powerful segmentation program was developed, based on an algorithm which uses multihistograms to find the cutoff values for partitioning . . . including a scheme for avoiding the fragmentation of textural parts and a scheme for extracting the detailed structures veiled by dominant ones.
, . . The result of segmentation is organized into a well-structured symbolic data network with powerful retrieval facilities. . . . The “secondary” features can be derived easily from the primary ones when they become to be necessary. . . .
, . . Bottom-up control and top-down control were effectively combined in the framework of region growing. The system generates a plan via bottom-up control as a representation of the rough structures in the input scene. A scheme of approximate reasoning was developed to handle the uncertainty contained in the knowledge and the pictorial data.
A symbolic description of the input scene is made via top-down analysis. The top-down process was constructed using a production system architecture. . . . In order to reduce the computation needed to manage the production system, two phases (scene phase and object phase) have been successfully established in the control structure utilizing the two opposite properties (globality and locality) of scenes.
--From the Author’s Conclusion
This work was performed in the Department of Information Science at Kyoto University. The Author’s Conclusion quoted above nicely summarizes the contents. To this, it should be added that extensive use is made of fuzzy feature definition.
The real-world domain investigated here is rather restricted, since only scenes found on the university campus are used. The real-world knowledge used here seems to be tailored to the environment, although the author claims a “non-specific” system. For example, windows are defined as “vertically-lon- g” rectangles. That is how they look in the photograph; as a result, backtracking is avoided. The author goes on to state:
(2) Whenever a patch is interpreted by executing [sic] an action, every action-patch which is going to give a different interpretation to the patch is declared inconsistent and deleted from the agenda.
(3) Once a patch is given a label, it is not examined any more.
Truly, such a system can only work in a restricted universe--restricted enough for total knowledge to be built-in. This is hardly AI. In spite of this objection, I found the work extremely suggestive and stimulating. However, the publisher does not seem to have fulfilled the promise printed on the back cover of this book: “The series provides a vehicle for rapid publication in softback form. . . .” The author does not give a submission date for the manuscript, but circumstantial evidence points to a long period of waiting in the Pitman editorial offices. The latest reference given in the bibliography is from 1979] It would be strange that Ohta would ignore the major work done in the same field at the Electrical Engineering Department of Kyoto University [1]. In particular, techniques for detecting regular arrangements--with gaps--is described there. This would be most useful in, for example, a more robust and general window-in-office-building-fac- ade detector than that described in the reviewed work. The index is rather sketchy, and a certain number of key concepts used in the work do not appear there at all.
