Vol.4 No.2 2011
12/66

Research paper : ARGUS: Adaptive Recognition for General Use System (N. Otsu et al.)−83−Synthesiology - English edition Vol.4 No.2 (2011) N. Otsu, T. Kurita and I. Sekita: Pattern Recognition – Theory and Application, Behaviormetrics Series 12, Asakura Shoten, Tokyo (1996) (in Japanese).N. Otsu: Mathematical studies on feature extraction in pattern recognition, Research Report of Electrotechnical Laboratory, No.818, 210 pages (1981) (in Japanese).N. Otsu, T. Shimada and S. Mori: Shape feature extraction using Nth order auto-correlation mask, IECE Technical Report, PRL-78 (31) (1978) (in Japanese).N. Otsu and T. Kurita: A new scheme for practical flexible and intelligent vision systems, Proc. IAPR Workshop on [1][2][3][4]ReferencesNotesNote 1) Initially, it was intended to be called ARGUS (Adaptive Recognition for General Use System), after the giant of Greek mythology with a hundred eyes. In recent years, although HLAC/CHLAC has often been used as an abbreviation for this methodology, this actually refers to the first-stage feature extraction, and therefore is not appropriate. As such, the system/methodology as a whole will be referred to as ARGUS.Note 2) HLAC features of a binary image are closely related to the image spectra due to the N +1th vector in perceptrons[10]. Here, combinations of “black” (1) and “white” (0) have been considered, and at first glance, it may appear that our approach that considers only “black” would not be sufficient, but it actually is. For example, ■□ , with f0 = f(r) = 1 and f1 = f(r+a1) = 1, and logically f0 ・ f1 = f0 ・ (1−f1) = f0 −f0 ・ f1, is represented in the range of the linear sum of feature values due to the masks (No.1 and No.3).Note 3) This method was proposed[2][3] prior to the back propagationlearning method[7] in neural networks.Note 4) HLAC features count the number of those topological geometry elements and their coefficients are adaptively determined in the second stage by multiple regression. Note 5) Minimum Distance Decision: the method whereby the distance from the unknown input feature vector to the center of each class is measured, and the class with the shortest is identified.Note 6) Helios Watcher (KK Hitachi Building Systems),http://www.hbs.co.jp/lineup/elevator/hw outline.htmlNote 7) A vector with elements of 1or 0 to express positive or negative response for each corresponding word.statistics,” which are representative of the recent HOG and SIFT features. In addition, not limited to images, it is widely applicable to the multi-channel time series data for audio and various kinds of sensor information and the like, as well as to general three-way data. Future goals include extension from quantitative data to qualitative (categorical) data, and the development of technique is already underway[29].The application of this method is expected in a wide range of computer vision applications, such as automatic (unattended) video surveillance for intelligent security cameras, various appearance inspection systems, image annotation and retrieval, robot vision, motion analysis, and evaluation in sports and rehabilitation. At the moment, we are promoting its medical application through collaborative research with university hospitals and cancer centers, specifically toward an automatic inspection system for cancer using microscopic images. In addition, centering on an AIST-approved venture (United Technologies Institute), applications are being developed for the commercial viability of semiconductor substrate inspection and various kinds of visual inspection for agriculture and livestock fields, including inspection of rice quality and forecast of estrus and delivery in milk cows in local consortium projects.The practical application of this method requires adjustments such as pre-processing and parameter tuning (correlation width). Future topics include automation of those settings, accumulating such knowledge base.AcknowledgementsThis study significantly progressed under the guidance of graduation and master’s degree theses at the Department of Mechano-Informatics, the University of Tokyo, where I (the author) served concurrently. I hereby offer my appreciation to my dear students. I am especially indebted to the great contribution made by Dr. Takumi Kobayashi (currently in the Information Technology Research Institute, AIST) for the subsequent collaborative research and wish to offer my appreciation. I am also indebted to Dr. Takio Kurita (currently in Hiroshima University) for his work in the early HLAC experiments, and to various postdoctoral fellows as well as the Dr. Sakaue group and the Dr. Higuchi group for their contributions in recent years to the development of applications. Lastly, I would like to thank all of those concerned.

元のページ 

page 12

※このページを正しく表示するにはFlashPlayer10.2以上が必要です