1992 Fiscal Year Final Research Report Summary
Studies on unifying contour and texture informations of image recognition
| Project/Area Number |
03650278
|
|---|
| Research Category |
Grant-in-Aid for General Scientific Research (C)
|
| Allocation Type | Single-year Grants |
|---|
| Research Field |
電子通信系統工学
|
|---|
| Research Institution | Kyushu Institute of Design |
|---|
Principal Investigator |
TAKIYAMA Ryuzo Kyushu inst. of Design, Faculty of Design, Professor, 芸術工学部, 教授 (20037815)
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
ONO Naoki Kyushu inst. of Design, Faculty of Design, Research Assistant, 芸術工学部, 助手 (60185642)
SAKAMOTO Hiroyasu Kyushu inst. of Design, Faculty of Design. Lecturer, 芸術工学部, 講師 (70112357)
|
|---|
| Project Period (FY) |
1991 – 1992
|
| Keywords | Image Recognition / Contour Information / Texture Information / Unifying Information / Neural Network / Learning / Fourier Descriptor / Co-occurrence Matrix |
|---|
| Research Abstract |
We described contour and texture information which are very useful for image recognition, implemented a system which unity above informations, and performed various experiments on image recognition. Outlines of obtained results follow: 1. We developed an equilateral-polygonal approximation algorithm applied to contours extracted from images. 2. We developed a smoothing method for contour images, and checked the usefulness of the algorithm in 1 on the smoothed images. 3. We extracted the P-type descriptor of an equilateral polygonal approximated figure in 2, and investigate the relation between the range of frequencies of the descriptor and the degree of approximation of the obtained figure. 4. In various descriptions of texture information we showed that the co-occurrence matrix is most desirable, and verified its effectiveness by use in the Brodaz images. 5. We obtained some properties which clarify relationships between architectures of neural networks and information processing capabilities. 6. We decided to use three layer neural networks to unify contour and texture informations, and proposed parallel and serial methods for the unification. 7. Taking all results above into consideration, we implemented a unifying neural network, experimented on fisher images by the network, and showed its usefulness.
|
