Lapas attēli
PDF
ePub

62

12

10

9

8

6

relating to color, range, separation of presented a clear summary of past

No. of cameras for stereo images, and other work in Japan. Shirai pointed out that

Topic

Papers parameters that is used at the input Japan has a Computer Vision Group phase. Next, properties of the image with about 500 Japanese members. They Image understanding

82 are invoked to locate image features meet bimonthly and had their first sym- Stereo such as edges, lines, or regions. This posium this summer (this is in addition (or multidimensional) phase is usually called feature extrac- to any international meetings that have Time sequence images

46 tion. Lastly, at the highest level, there been held). The group's chair is Prof. Image database

45 is some underlying object model, for Yachida, mentioned above. There is a example, the designer knows that the Special Interest Group (SIG) in Pattern Shirai pointed out that in a few areas, scene is supposed to be of an auto- Recognition and Understanding (until such as industrial applications, there is mobile, and then matching is done to recently Pattern Recognition and Learn- far more work than is represented by locate these objects in the scene. This ing) sponsored by the Institute of Elec- the number of published papers. involves solving problems such as tronics, Information, and Communi- The only more recent data are from direction, angle, and occlusion. The cation Engineers (IEICE), which pub- the IPSJ's SIG CV for 1990-1991: result is scene description or scene lishes about 125 papers yearly in 10 knowledge. Research in computer vision issues. This group also includes a small

No. of is often compartmentalized into amount ofspeech recognition. There is

Topic

Papers subtopics that follow this modularization a Special Interest Group in Computer as well. For example, “image processing” Vision (SIG CV) sponsored by the Time sequence images

18 usually refers mostly to the lowest levels, Information Processing Society of Japan Feature extraction whereas pattern matching research (IPSJ), focusing on image processing, 3D input and modeling almost always refers to the highest level. that publishes about 60 papers each Stereo

In computer vision research it is not year in a bimonthly journal. Finally, Medical too difficult to get to the leading edge there is also a SIG in Pattern Measure- Matching of what has been accomplished, and ment sponsored by the Society on Instru- Neural network for matching 6 thus almost any project will quickly mentation and Control Engineers Shape from X need to address advanced problems. (SICE), which publishes about 20 papers Face But simply put, because the Japanese yearly in four issues, but this is heavily have tried so many different approaches, oriented toward very practic hard- It is clear that the most important their breadth of research experience is ware problems.

new area is analysis of sequences of very much greater than the Korean's. A survey of the database of infor- images, and this view was also shared They are also trying deeper and more mation processing literature in Japan by the Korean attendees. While there sophisticated techniques, although the (this covers the period 1986-1988, the are only four papers concerning comdisparity might not be too great in a few latest data that are available) charac- puter vision in the field of human faces, specific promising areas such as scene terizes computer vision related papers this is also seen to be a growing area, identification. as follows (excluding coding of images). incorporating human computer inter

face, remote teleconferencing, human JAPAN AND KOREAN

No. of emotional information processing, and COMPUTER VISION

Topic

Papers image coding. SUMMARIZED

Shirai went on to describe several Applications

477 specific Japanese projects that involve From the Japanese side Drawings (96)

computer vision. The most elaborate Medical (85)

of these is the ¥20B ($140M) 1983Prof. Yoshiaki Shirai Characters (81)

1991 “Robot for Extreme Work" projDept of Mechanical Engineering for Industrial (75)

ect, in which the ultimate application is Computer Controlled Machinery Scientific (64)

the development of an autonomous Osaka University Remote sensing (40)

teleoperated robot for nuclear plant, Suita, Osaka 565, Japan

Intelligent robot (36)

pipe cleaning, underwater, and emerTel: +81-6-877-5111 x4706 3D input and recognition

gency (such as fire) operation. This Fax: +81-6-876-4975

Feature extraction

105

particular project involves much more E-mail: shirai@ccm.osaka-u.ac.jp Hardware systems

than just computer vision, and in fact

5

4

132

89

[ocr errors]

research has been done on fundamen- on a highway with a centerline by dis- facilities imported from other countal problems of locomotion, tele- tinguishing line and road boundaries tries. Several Korean companies do existence, manipulation, and sensing, and also road signs. Phase-2, from 1995

and also road signs. Phase-2, from 1995 market low-price machine vision sysas well as the development of a system to 2000, will deal with multilane high- tems developed in Korea but, to date, integration language. The part of the ways, tunnels, rain, windshield wipers, their performance has not been project dealing with these fundamen- and using stereo for obstacle avoid- impressive. Production line utilization tal issues actually received the bulk of ance. Phase-3, from 2000 to 2030, will of computer vision is infrequent and the funding, and more applied aspects, (hopefully) deal with normal roads, limited to simple inspection and very i.e., to really develop such a robot, were crossings, parking, backing up, and using repetitive tasks. Park claimed that not so well funded. In addition to a mirror and will involve tools of scene Korean companies would rather not Japanese universities, the Electrotech- understanding and map matching. This purchase a general purpose vision system nical Laboratory (ETL), Fujitsu, project also has a very unique perspec- such as a Sammi-AB but prefer to Toshiba, and other companies partici- tive on wanting to use active sensing, obtain very task-specific systems. pated--Toshiba working on feature for example, to help the scene under- Industry does see a very strong need for extraction and Fujitsu on projecting standing by using sound and to under

standing by using sound and to under- efficient algorithms for segmentation, images onto a sphere (which Shirai stand the sounds being received by use classification and, of course, for high claimed works well in clean environ- of the input visual data. Thus the proj- reliability. ments). ETL has done a great deal of ect designers are thinking about sensor Before 1989 work was very scatwork on sensing, stereo, and robot vision fusion and multisensor integration. tered and mostly restricted to worklanguage development and actually These parts of the program will begin shops and courses in medical imaging, issued a special issue of the ETL Bulletin soon at Tokyo University. Shirai computer-aided design/manufacturing on Robotics in which this has been admitted that at the moment image (CAD/CAM), image processing, and summarized. Shirai showed several segmentation is one of their most diffi- computer graphics. Modern work really photos of the prototypes that had been cult problems, but he did show us some begins only in 1989 with an Image developed. One of these looked like a film of the motor home on its test road Processing and Image Understanding monster from "Star Wars II,” and Shirai and it seemed to be working, although Workshop (at POSTECH) at which admitted that 8 years was a long time rather slowly. This appeared to be at a time it was decided to have annual for this technology and that a newer much less advanced state than the CMU workshops in order to share research project would have designed a less clumsy project I saw more than a year ago. activities. Subsequently, two workshops looking robot.

From the Korean side

have been held with a total of 42 papers Another interesting Japanese proj

presented. Two related meetings are ect is a vision-based vehicle system.

Prof. Chan Mo Park

worth mentioning, an International This shares some of the same goals as Professor and Chairman

Workshop on Medical Imaging (March similar projects in the United States, Dept of Computer Science

1991 at the Korea Institute of Science such as at Carnegie Mellon University & Engineering

and Technology) and a Chapter Meet(CMU). The Japanese project (which Pohang Institute of Science ing of the Korea Information Society is also supported by the Ministry of and Technology

(May 1991 at Chung-Joo University), International Trade and Industry P.O. Box 125

which had as its theme “Current Status (MITI)) is in two phases. The initial or Pohang 790-600, Korea

of Pattern Recognition Technology" phase-o part was mostly done by Fujitsu Tel: +82-562-79-2251

and generated half a dozen overview and Nissan around 1989 and involved a Fax: +82-562-79-2299

papers. There are now three SIGs intervehicle on a special test course, shadow- E-mail: parkcm@vision.postech. ested in vision: SIG AI (artificial intelless illumination, and only large obsta

ac.kr

ligence) (Korea Information Science cles. The vehicle (a motor home) has

Society), SIG IP-TV (information three cameras for lane detection and

gave a summary of computer vision processing) (Korean Institute of Teletwo more for obstacle avoidance and a activities in Korea. Until very recently matics and Electronics), and SIG Images sonar system. Techniques used are line there was not much to report, and even (Korean Institute of Communication following for lane finding and sonar today he emphasized that industrial Science). for obstacles and for range finding applications are very limited. Most Park also gave a list of research Phase-1, which runs from 1989 to 1995, research is occurring at universities

research is occurring at universities activities at various Korean research involves learning how to run the vehicle and government research institutes using centers (see the Appendix) but did not

W

go into detail about the projects. This list gives a realistic sense of the work going on in Korea. Because the data were collected by asking scientists, the amount of thought and detail provided varies greatly (how many PCs does a Cray-2 equate to). But by scanning this, it is very clear that there are only a very few places with substantial equipment resources with respect to vision. I will try to obtain more details about the actual progress of the research at those institutes. Park did show AVIS (a project at POSTECH), which is an automated inspection system for use in the Pohang steelmaking factory using the PIPE computer (purchased from Aspex). It is also installed at the Korea Advanced Institute of Science and Technology (KAIST).

For the future Park felt that vision work should concentrate on factory automation, that biomedical applications were still a promising field that could have broader applications, and that handwritten character recognition was the key to office automation applications. In the area of more fundamental research, he felt that Korean scientists should work on moving target detection, remote sensing, mobile robots, and other motion-related problems and that the Korean Government needed to take a more active role with additional funding, manpower development, and mechanisms to encourage cooperation between industry and university, as well as international cooperation.

Dr. Sung Kwun Kim

Dr. Yoshinori Kuno Executive Director

Senior Research Scientist Robotics & Automation R&D

Toshiba Corporation Division

Information Systems Laboratory Samsung Electronics

Research and Development Center 259 Gong Dan-Dong

1, Komukai Toshiba-cho Gumi, Kyung Buk, Korea

Saiwai-ku, Kawasaki 210, Japan Tel: +82-546-460-2015

Tel: +81-44-549-2241 Fax: +82-546-461-8038

Fax: +81-44-549-2263

E-mail: kuno@isl.rdc.toshiba.co.jp Prof. Jong Soo Choi Dept of Electronic Engineering The panel was co-chaired by Chung-ang University 221 HukSeok Dong

Dr. Masakazu Ejiri DongJak Gu, Seoul, Korea

Senior Chief Scientist Tel: +82-2-815-9231-7 X2235

Corporate Technology Fax: +82-2-815-9938

Hitachi Central Research

Laboratory Prof. Kwang Ik Kim

Kokubunji, Tokyo 185, Japan Dept of Mathematics

Tel: +81-423-23-1111 Pohang Institute of Science

Fax: +81-423-27-7700 and Technology P.O. Box 125

and Pohang, Kyung Buk 790-330, Korea Tel: +82-562-79-2044

Prof. Sang Uk Lee Fax: +82-562-79-2799

Signal Processing Laboratory E-mail: kimki@vision.postech.ac.kr Department of Control and

Instrumentation Engineering Dr. Takeshi Shakunaga

Seoul National University
NTT Human Interface Laboratories Seoul, Korea
NTT
3-9-11, Midori-cho

Unfortunately, none of the panelists Musashino-shi, Tokyo 180, Japan provided handouts and so my summary Tel: +81-422-59-3336

below is based on notes that may not be Fax: +81-422-59-2245

completely accurate. E-mail: shaku@nttarm.ntt.jp

Ejiri (Hitachi) only made a few

remarks but pointed out that vision Dr. Johji Tajima

systems were realized in Japan 20 years Research Manager

ago. (See my comments earlier about Pattern Recognition Research depth and breath of research vis-a-vis Laboratory

Japan and Korea.) He also pointed out NEC C&C Information Technology that there was very tough competition Research Laboratories

between Japanese companies but very 1-1 Miyazaki 4-chome

friendly discussions between researchers. Miyamae-ku, Kanagawa 216, Japan (Isn't this the Japanese way; maybe this Tel +81-44-856-2145

is the reason that everybody's solderFax: +81-44-856-2236

ing inspection systems look alike.) E-mail: tajima@pat.cl.nec.co.jp Kuno (Toshiba) claimed that more

than 100 computer vision applications were developed at Toshiba. Not all were successful and most were developed for position detection. The ones that

PANEL DISCUSSION: APPUCATION OF COMPUTER VISION FOR AUTOMATION

This was the most fascinating part of the meeting, as it placed six experts together and gave each an opportunity to describe work that they had seen and work that they were hoping would be done in the future. Panelists were

[ocr errors]

work have a common thread that they Presumably Toshiba's research support standardized database of images would begin with a good (high contrast) image will follow these paths.

bevery helpful for studying algorithms. input. He mentioned three specific Tajima (NEC) felt that for image As far as new directions, he mentioned examples of vision systems now in use processing (as opposed to image the importance of sensor fusing to within Toshiba but did not go into any understanding) there were already very

understanding) there were already very enhance the reliability of existing real detail about any of the specific cheap general purpose systems with techniques. hardware or software techniques that many operators built into hardware for

Shakunaga (NTT) claimed that NTT were used.

preprocessing (such as thresholding, was trying to combine visual information

etc.). He then went on to give a rapid processing with other technologies to • Soldering inspection system for the description of a collection of vision develop applications in the area of

mounting of integrated circuits (ICs) applications within NEC, again with advanced visual telecom services and onto printed circuit boards (PCBs). few details.

network support, both of obvious In some sense this is a very simple

importance to NTT. He gave two problem, as there is a clean model • Multilayer substrate inspection examples. of what the image is supposed to station to detect short circuits, pin look like. The hard part of this holes, etc. for use with the boards • Maintenance. A manhole facility problem is to get good input images. NEC uses on their supercomputers inspection system using a truckToshiba's inspection station uses 168 (SX series). This system can inspect mounted, underground-looking light-emitting diodes (LEDs) to illu- a 225-mm? board area in 25 minutes. radar that eliminated the need for minate different parts of the target.

digging to locate pipes. This uses • Soldering inspection station, looking pattern recognition and frequency • Agricultural automation. This a great deal like Toshiba's, with domain analysis. This is said to

involves using a robot to cut young five cameras and lights for three- work to a depth of 1.5 meters, which plants at the proper stem length. dimensional (3D) views.

includes 75% of the company's pipes.

(If you have ever lived in “dig we • Digital audio tape, and VCR, • Deformation measurement by laser must” New York you will know how

magnetic head gap-width adjusting range finding for circuit boards. welcome such a system would be.) system using computer processing

A second system uses a TV camera of images of Moire patterns. • Inspection system for determining on a stalk that looks inside manholes

if foreign objects are inside empty and generates a stereo view of the Kuno commented succinctly about (Coke) bottles, and another system facility's layout inside (using vertical the state of the art, that “we are using for determining the amount of liquid camera movement for the second '70s algorithms on '80s hardware." As in a bottle.

image) and then generates a drawing for the future he felt that there would

of the manhole contents. This uses be no general purpose vision system in • A 3D human body measurement edge detection, which is said to be the near future because of cost issues. system. This was the most intriguing accurate to 0.05 pixel. In his view there are three basic ways of the lot. The application here is to to use computer vision systems.

determine the tailoring of apparel • Human computer interface. The idea

by measuring cross sections of is to transmit less feature data for • Use simple (e.g., low cost) vision humans. The subject is in a box and teleconferencing. NTT has been

system cleverly for factory is illuminated by six lasers. The experimenting with human head, lip, automation, human computer software uses a body model that and body image readers. The idea is interface, etc.

runs on a workstation and a database to interpret head motion and that runs on a minicomputer.

generate understanding based on • Apply heuristic tuning to fields with

head movement. This uses edge strong needs, e.g., character scanning As far as industry was concerned, detection of head silhouette and recognition is a perfect example. Tajima felt that the important work analysis of facial area.

needs to be done in 3D recognition as • Do basic research on sophisticated well as motion detection, and that Shakunaga divided future research

vision systems for future applications, recognition of features needs to be above themes in three directions.
such as robots, nuclear power plants, 99% to be industrially useful. He felt a
space work, etc.

• Early vision. Because human sys- Choi (Chung-ang University) not in a factory situation. There are no

tems are very adaptable, we should described work in 3D vision. Of course, camera standardizations. Missing scene study adaptive tuning of input data the major problem is to extract infor- parts and shadowing are a problem, as and attention getting mechanisms. mation about a 3D world from two- obviously it isn't possible to deduce 3D We should also study human imple- dimensional (2D) images. This can begin data for missing parts of a scene. mentation of early vision algorithms with range finding, for knowing the for edge, region, texture, shape, distance to objects will then allow one COMMENTS ABOUT distance, motion, etc.

to determine which one is in front, etc.; SPECIFIC CONTRIBUTED

or it can begin with finding segments, PAPERS • Middle level vision. Requires features, objects ... to which stereo, etc.

research into model based match- can be applied. (An occlusion bound- A complete list of titles/authors of ing, from specific (recognition) to ary, for instance, allows triangulation the presented papers is being prepared generic (cognition).

on the occluding edge--it is not a fea- and will be distributed (electronically)

ture of the occluded object.) The two as soon as it is ready. However, topics • High level vision. Study 3D world approaches are:

discussed included description and manipulation. Consider integration of vision and Passive

• Character recognition & document semantic databases.

understanding • Monocular vision requires a priori Kim (Samsung) felt their problems information (in some problems this • Image processing & coding were similar to NTT's and to Toshiba's. is available) He also felt that the cost of vision sys

• Hough transform tems will be coming down quickly, • Photometric stereo, e.g., using difalthough this is now still a bottleneck. ferent light sources.

• Scene understanding & object He gave a short list of computer vision

recognition applications but with even fewer details • Shape from shading, although than the other industrially based recovering surface orientation from • Neural nets speakers.

a single image is obviously ill posed

as are many of the monocular • Stereo & shape recognition • System for mounting a screw in a techniques. microwave oven cavity.

• Motion & sequential image • Range data from two different processing • Simple assembly.

images, or a sequence.

• Sensing & recognition • Soldering system for inspection and Active modification, again very similar to

• Mobile robots NEC's and Toshiba's.

• Structured lighting (mentioned work

by Prof. Sato and also work at CMU). • Vision paradigm • Color monitor. • Time of flight (sonar, etc.).

• Computer vision hardware & • Mobile navigation.

algorithms • Conventional triangulation as with The motivation for reducing the cost

• Motion & shape reconstruction of inspection was made clear to the audience--Kim pointed out that at None of the techniques is best for all • Intermediate & high level vision Samsung a very large fraction of the situations and ultimately the system electronic manufacture employees are designer must choose the most appro- • Thinning, quadtree, & component doing inspection and adjustment related priate. Systems with low cost and high labeling jobs. performance are not available, certainly

• 3D modeling & recognition

range finders.

« iepriekšējāTurpināt »