| Co-Investigator(Kenkyū-buntansha) |
SAKAMOTO Hiroyasu Faculty of design, Kyushu Institute of Design, Associate Professor, 芸術工学部, 助教授 (70112357)
ONO Naoki Faculty of design, Kyushu Institute of Design, Associate Professor, 芸術工学部, 助教授 (60185642)
|
|---|
| Budget Amount *help |
¥2,700,000 (Direct Cost: ¥2,700,000)
Fiscal Year 1999: ¥800,000 (Direct Cost: ¥800,000)
Fiscal Year 1998: ¥1,900,000 (Direct Cost: ¥1,900,000)
|
|---|
| Research Abstract |
At the earlier stage of this study, we constructed database for face recognition. 52 persons co-operated for the database. Each person was photographed in five expressive faces, I.e., normal, simple, anger, sad and wonder. Furthermore each person was also photographed in 29 directional poses with normal expression, i.e., from -90°to +90°with 10°steps. On the database various experiments to reach the aim of this study were performed.
Results obtained in this research follow:
1. Faces in complicated background were detected exactly with inexpensive computational load.
2. We proposed a method for affine-transformation-invariant recognition of digital figures ---"self triangle function method." This method can be used as an approximative method for orientation-independent recognition of face.
3. We proposed a method for 2D object recognition, which is independent of view points. This method, differing from one in 2, is based on point pattern matching, and is useful to recognize parts of face (eyes, nose etc.).
4. As a fundamental of face figure recognition, we proposed a digital figure recognition method which allows for position, scale and rotational variation, which is called a "self distance function." This method was proved to be useful for digital figure recognition but suffers from its computational load.
This method was revised to reduce the computational load, which is called a "central distance method," which needs O(n) computations for an n point-composed figure, whereas the self distance does OィイD12ィエD1.
5. We verified that the recognition rate of expressive faces laises by unifying features obtained by Gabor-wavelet transformation by neural networks.
6. By using the self triangle function method conferred in 2., orientation-independent identification of individuals was tried. This method, however, can be applicable to only small variation in direction, which should be improved hereafter.
|
